2026-09-06 19:08:45 +00:00
23 changed files with 2572 additions and 368 deletions
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -117,9 +117,10 @@ sweep, and refreshes whenever the poller runs.
 is **never** refused for exceeding it — it **gates nothing**. Overage is billed
 against the account's credit balance (`allowance_remaining_usd` from the
 provider). `/metrics` now reports balance, estimated burn rate, and projected
-runway (hours remaining) derived from the provider-reported
-`allowance_remaining_usd` over a configurable window, using a configurable
-minimum segment length and sample count to avoid wild extrapolations.
+runway **per provider** inside `quota.by_provider`, plus a `total_balance_usd`
+field that sums whatever each provider reports. The flat balance/burn/runway
+keys at the top level of `quota` were removed; every consumer was updated in the
+same change, so there are no deprecated aliases.

 Three signals said `deepseek-v4-flash` — catalog token price (7.8x cheaper),
 NeuralWatt's own published per-request energy (~10x lower), and a live 70k
@@ -222,7 +223,7 @@ rather than from months of history.
 - `poller.py` — also seeds/updates `provider='ollama-local'` rows from `config.yaml` each poll so local rows stay current even when NeuralWatt is unreachable.
 - `logs.py` — per-request trace id (ContextVar), logfmt, journald priority prefixes; `logs.bind()` survives StreamingResponse generators. [operations](docs/operations.md).
 - `metrics.py` / `GET /metrics` — read-only observability; takes `(conn, cfg)`, never imports `dispatcher`. Also carries the three detectors added after the incidents below: capability sub-ceilings, the reactive rejection detector, and the classifier-degradation share. [api](docs/api.md).
- `tui.py` — Textual dashboard over `/metrics` + `/events/decisions`; live feed, category→model panel, detail popup; data layer split into `tui_model.py`. The decision table leads with a `time` column and carries `profile` plus an `E` flag for exploratory picks; the quota panel leads with balance and runway. [architecture](docs/architecture.md).
+- `tui.py` — Textual dashboard over `/metrics` + `/events/decisions`; live feed, category→model panel, detail popup; data layer split into `tui_model.py`. The decision table leads with a `time` column and carries `profile` plus an `E` flag for exploratory picks; the quota panel now shows one row group per provider plus a `total_balance_usd` line, with a lead calculated only from providers whose balance is polled from a prepaid account. [architecture](docs/architecture.md).
 - `tests/test_tui_schema_drift.py` — the tripwire that keeps the two honest. A new `route_decisions` column must be registered as surfaced or deliberately-not, or the test fails **naming the column**. Five columns had already reached the schema without reaching the dashboard; `ROUTE_DECISIONS_COLUMNS` in `tests/test_route_decisions.py` had itself drifted.
 - `tests/test_tui_warnings.py` — the same idea for warnings. Every class `/metrics` can emit must render in `#warnings-panel`, and every emitted warning must be registered — the second failing with the RAW text, because the point is that nobody knew the class existed. **Its fixture is a coupled system**: adding a seed can silence an existing class (a small-context seed once killed the escalation hazard by dragging the p95 down), which is why both directions are asserted.
 - `router_cli.py` — one-shot `/route` probe (no spend), raw JSON with `--json`. [api](docs/api.md).
@@ -763,6 +764,29 @@ include `request_id` and `session_dir` in `local_energy_observations`, so a
 client report on a local answer resolves to the same `(model_id, provider,
 task_category)` provider-agnostic record as a cloud one.

+## Routing notes
+
+Ranking is quality-first, cost as a tiebreak; cost is never allowed to override
+a real quality gap. The optional `objective.credit_attenuation` block extends
+that tiebreak without changing it: when the block is enabled, a per-provider
+multiplier is applied to a candidate's comparison cost only, producing an
+`effective_cost` that breaks ties. The multiplier is derived from the provider's
+polled account balance (the `balance_url` path, such as OpenRouter), so a low
+prepaid balance can nudge a near-tie toward a healthier provider. The logged
+`est_cost_usd` and the decision history stay as raw catalog estimates. The
+multiplier is 1.0 for providers whose balance comes from per-completion
+`allowance_remaining_usd` telemetry (NeuralWatt), so normal overage readings do
+not bias routing.
+
+Two semantics matter when reading the numbers. `total_balance_usd` is a sum of
+heterogeneous provider-reported readings: OpenRouter's prepaid credits plus
+NeuralWatt's overage allowance, which normally reads near -$0.004. It can be
+negative and it is not a single spendable figure. `credit_attenuation.enabled`
+deliberately lives only in the config file; it is absent from the admin
+persisted-config allowlist and from provider edits. Turning it on or off
+requires editing `config/config.yaml` and `systemctl --user restart
+llm-router.service`, because the dispatcher's `cfg` binds at import time.
+
 ## What's NOT built yet — pick up here

 Built: session-directory attribution, the local energy ledger, local model
--- a/admin/frontend/index.html
+++ b/admin/frontend/index.html
@@ -668,74 +668,113 @@ function renderQuotaModal(quota) {
    el.innerHTML = `<div class="empty">No quota plan configured</div>`;
    return;
  }
-  // ── balance-led rendering ───────────────────────────────────────────
-  if (quota.balance_usd != null && typeof quota.balance_usd === 'number') {
-    const balanceStr = '$' + quota.balance_usd.toFixed(2);
-    const burn = quota.burn_rate_usd_per_hour ?? null;
-    const proj = quota.projected_hours_remaining ?? null;
-    const low = !!quota.runway_low_warning;
-    const hours = proj != null ? '~' + Math.round(proj) + 'h runway' : 'runway unknown';
-    const hourColor = low ? 'var(--tblr-danger)' : !burn ? 'var(--tblr-warning)' : 'var(--tblr-success)';
-    const burnLabel = burn != null ? '$' + (Math.round(burn * 100) / 100).toFixed(2) + '/hr' : (quota.runway_note ? '' : 'n/a');
-
-    // Build detail rows
-    const detailRows = [];
-    if (proj != null) detailRows.push('Projected');
-    if (burn != null && !quota.runway_note) detailRows.push('Burn rate');
-    detailRows.push('Plan');
-    detailRows.push('Metered');
-    detailRows.push('Reset');
-
-    let rowIdx = -1;
-    const runways = ['Low Runway', 'Burn N/A'];
-    const meteredLabel = quota.reset_date
-      ? `Metered (period since ${escapeHtml(quota.reset_date)})`
-      : 'Metered (30d window)';
-
-    el.innerHTML = `
-      <div class="mb-3">
-        <div style="font-size:1.8rem;font-weight:700;color:#4ade80">${escapeHtml(balanceStr)}</div>
-        <div style="font-size:1rem;color:${hourColor};margin-top:2px">${hours}</div>
-        ${low ? '<div class="badge bg-warning text-warning mt-1">Low Runway</div>' : ''}
-      </div>
-      <div class="quota-details">
-        ${burn != null && !quota.runway_note ? '<span class="quota-label">Burn rate</span><span class="quota-value">~' + escapeHtml(burnLabel) + '/hr</span>' : ''}
-        ${quota.runway_note ? '<span class="quota-label">Burn</span><span class="quota-value">' + escapeHtml(quota.runway_note) + '</span>' : ''}
-        ${proj != null ? '<span class="quota-label">Projected</span><span class="quota-value">~' + Math.round(proj) + 'h</span>' : ''}
-        <span class="quota-label">Plan</span><span class="quota-value">${quota.plan_kwh} kWh</span>
-        <span class="quota-label">${escapeHtml(meteredLabel)}</span><span class="quota-value">${quota.metered_kwh_period != null ? quota.metered_kwh_period + ' kWh' : quota.metered_kwh_30d + ' kWh'}</span>
-        <span class="quota-label">Fraction</span><span class="quota-value">${(quota.metered_fraction_of_plan * 100).toFixed(1)}%</span>
-        <span class="quota-label">Reset</span><span class="quota-value">${quota.next_reset_date || 'not configured'}</span>
-      </div>`;
-    return;
-  }
-  // ── fallback: percentage-only rendering ─────────────────────────────
+  const meteredLabel = quota.reset_date
+    ? `Metered (period since ${escapeHtml(quota.reset_date)})`
+    : 'Metered (30d window)';
  const pct = Math.min(quota.metered_fraction_of_plan * 100, 100);
-  const color = pct > 90 ? 'var(--tblr-danger)' : pct > 75 ? 'var(--tblr-warning)' : 'var(--tblr-success)';
+  const heroColor = pct > 90 ? 'var(--tblr-danger)' : pct > 75 ? 'var(--tblr-warning)' : 'var(--tblr-success)';
  const displayPct = (quota.metered_fraction_of_plan * 100).toFixed(1);
  const calls = Number(quota.metered_calls_30d || 0).toLocaleString();
-  const pctInside = pct > 22;
-  const pctStyle = pctInside
-    ? `left:calc(${pct}% - 30px);color:#fff`
-    : `left:calc(${pct}% + 8px);color:${color}`;
+
+  // ── per-provider balance rows (reusing .bar-row / .bars-scroll) ─────
+  const byProvider = quota.by_provider || {};
+  const providers = Object.keys(byProvider).sort();
+  let providerRowsHtml = '';
+  let totalBalance = quota.total_balance_usd;
+  if (providers.length) {
+    providerRowsHtml = providers.map(name => {
+      const p = byProvider[name] || {};
+      const balance = p.balance_usd;
+      const burn = p.burn_rate_usd_per_hour ?? null;
+      const proj = p.projected_hours_remaining ?? null;
+      const low = !!p.runway_low_warning;
+      const balanceAt = p.balance_at || null;
+      const source = p.balance_source || '';
+      const lineParts = [];
+      if (balance != null && typeof balance === 'number') {
+        lineParts.push('$' + balance.toFixed(2));
+      } else {
+        lineParts.push('balance unknown');
+      }
+      if (burn != null && typeof burn === 'number') {
+        lineParts.push('~$' + burn.toFixed(2) + '/hr');
+      }
+      if (proj != null && typeof proj === 'number') {
+        lineParts.push('~' + Math.round(proj) + 'h');
+      }
+      const suffix = formatBalanceStaleness(balanceAt, source);
+      const accent = low ? 'var(--tblr-danger)' : (balance == null || burn == null) ? 'var(--tblr-warning)' : 'var(--tblr-success)';
+      return `<div class="bar-row" title="${escapeHtml(name)}: ${lineParts.join(' · ')}${suffix ? ' · ' + suffix : ''}">
+        <span class="bar-label" style="color:${accent};font-weight:600">${escapeHtml(name)}</span>
+        <span class="bar-val">${lineParts.join(' · ')}${suffix ? ' <span style="color:var(--tblr-secondary);font-size:0.68rem">· ' + escapeHtml(suffix) + '</span>' : ''}</span>
+      </div>`;
+    }).join('');
+  }
+
+  let totalRow = '';
+  if (totalBalance != null && typeof totalBalance === 'number') {
+    totalRow = `<div class="bar-row" style="border-top:1px solid rgba(255,255,255,.09);margin-top:4px;padding-top:8px">
+      <span class="bar-label" style="font-weight:700">Total</span>
+      <span class="bar-val" style="font-weight:700">$${totalBalance.toFixed(2)}</span>
+    </div>`;
+  }

  el.innerHTML = `
-    <div>
-      <div class="d-flex justify-content-between align-items-center mb-2">
-        <div class="text-muted small">${calls} calls (30d)</div>
-        <div class="quota-pct fw-bold" style="color:${color}">${displayPct}%</div>
-      </div>
-      <div class="quota-capsule" role="progressbar" aria-valuemin="0" aria-valuemax="100" aria-valuenow="${pct}">
-        <div class="quota-capsule-fill" style="width:${pct}%;background:${color}"></div>
-        <span class="quota-capsule-pct" style="${pctStyle}">${displayPct}%</span>
-      </div>
-      <div class="quota-details mt-2">
-        <span class="quota-label">Plan</span><span class="quota-value">${quota.plan_kwh} kWh</span>
-        <span class="quota-label">Metered (30d)</span><span class="quota-value">${quota.metered_kwh_30d} kWh</span>
-        <span class="quota-label">Resets</span><span class="quota-value">${quota.next_reset_date || 'not configured'}</span>
-      </div>
-      <div class="quota-note">${escapeHtml(quota.note || '')}</div>
-    </div>`;
+    <div class="mb-3">
+      <div style="font-size:1.8rem;font-weight:700;color:${heroColor}">${displayPct}%</div>
+      <div style="font-size:1rem;color:var(--tblr-secondary);margin-top:2px">${quota.plan_kwh} kWh plan · ${calls} calls (30d)</div>
+    </div>
+    <div class="quota-details mb-3">
+      <span class="quota-label">Plan</span><span class="quota-value">${quota.plan_kwh} kWh</span>
+      <span class="quota-label">${escapeHtml(meteredLabel)}</span><span class="quota-value">${quota.metered_kwh_period != null ? quota.metered_kwh_period + ' kWh' : quota.metered_kwh_30d + ' kWh'}</span>
+      <span class="quota-label">Fraction</span><span class="quota-value">${displayPct}%</span>
+      <span class="quota-label">Reset</span><span class="quota-value">${quota.next_reset_date || 'not configured'}</span>
+    </div>
+    ${providers.length ? `
+    <h6 class="card-title mb-2" style="font-size:.78rem;text-transform:uppercase;letter-spacing:.06em;color:var(--tblr-secondary)">Per-provider balance &amp; runway</h6>
+    <div id="quota-provider-rows" class="bars-scroll" style="max-height:260px">
+      ${providerRowsHtml}
+      ${totalRow}
+    </div>` : ''}
+    <div class="quota-note">${escapeHtml(quota.note || '')}</div>`;
+}
+
+function formatBalanceStaleness(balanceAt, source) {
+  if (!balanceAt) return '';
+  try {
+    const then = new Date(balanceAt);
+    const ms = Date.now() - then.getTime();
+    if (Number.isNaN(ms) || ms < 0) return '';
+    const minutes = Math.floor(ms / 60000);
+    if (minutes < 1) return 'just now';
+    if (minutes < 60) return `sampled ${minutes}m ago`;
+    const hours = Math.floor(minutes / 60);
+    if (hours < 24) return `sampled ${hours}h ago`;
+    const days = Math.floor(hours / 24);
+    return `sampled ${days}d ago`;
+  } catch (_) {
+    return '';
+  }
+}
+
+function _quotaChipAccent(low, hasKnownBurn) {
+  if (low) {
+    return {
+      '--chip-accent': 'var(--tblr-danger)',
+      background: 'linear-gradient(135deg, rgba(239,68,68,.12), rgba(59,130,246,.08))',
+      borderColor: 'rgba(239,68,68,.35)',
+    };
+  }
+  if (!hasKnownBurn) {
+    return {
+      '--chip-accent': 'var(--tblr-warning)',
+      background: 'linear-gradient(135deg, rgba(245,158,11,.12), rgba(59,130,246,.08))',
+      borderColor: 'rgba(245,158,11,.38)',
+    };
+  }
+  return {
+    '--chip-accent': 'var(--tblr-success)',
+  };
 }

 function renderQuotaChip(quota) {
@@ -748,64 +787,55 @@ function renderQuotaChip(quota) {
    return;
  }
  chip.hidden = false;
-  // ── balance-led display ------------------------------------------------
-  if (quota.balance_usd != null) {
-    const burn = quota.burn_rate_usd_per_hour ?? null;
-    const proj = quota.projected_hours_remaining ?? null;
-    const low = !!quota.runway_low_warning;
-    const hours = proj != null ? Math.round(proj) : null;
-    let chipStyle = {};
+  // ── per-provider aggregate display -----------------------------------
+  const byProvider = quota.by_provider || {};
+  const providers = Object.keys(byProvider);
+  const entries = providers.map(name => byProvider[name]).filter(Boolean);
+  const knownBalances = entries.map(p => p.balance_usd).filter(v => v != null && typeof v === 'number');
+  const totalBalance = knownBalances.length ? knownBalances.reduce((a, b) => a + b, 0) : null;

-    if (low) {
-      // Red: runway critically low
-      chipStyle = {
-        '--chip-accent': 'var(--tblr-danger)',
-        background: 'linear-gradient(135deg, rgba(239,68,68,.12), rgba(59,130,246,.08))',
-        borderColor: 'rgba(239,68,68,.35)',
-      };
-    } else if (!burn) {
-      // Amber: burn unknown, not alarming
-      chipStyle = {
-        '--chip-accent': 'var(--tblr-warning)',
-        background: 'linear-gradient(135deg, rgba(245,158,11,.12), rgba(59,130,246,.08))',
-        borderColor: 'rgba(245,158,11,.38)',
-      };
-    } else if (chipStyle._unset !== true) {
-      // Green: healthy runway
-      chipStyle = {
-        '--chip-accent': 'var(--tblr-success)',
-      };
-    }
+  if (totalBalance != null) {
+    const low = entries.some(p => !!p.runway_low_warning);
+    const hasKnownBurn = entries.some(p => p.burn_rate_usd_per_hour != null);
+    const worstProjEntry = entries
+      .filter(p => p.projected_hours_remaining != null)
+      .sort((a, b) => Number(a.projected_hours_remaining) - Number(b.projected_hours_remaining))[0];
+    const hours = worstProjEntry ? Math.round(Number(worstProjEntry.projected_hours_remaining)) : null;
+    const chipStyle = _quotaChipAccent(low, hasKnownBurn);
+    const balanceStr = '$' + totalBalance.toFixed(2);

-    const balanceStr = typeof quota.balance_usd === 'number'
-      ? '$' + quota.balance_usd.toFixed(2)
-      : '$' + String(quota.balance_usd);
+    text.textContent = hours != null ? balanceStr + ' · ~' + hours + 'h' : balanceStr;

-    if (hours != null) {
-      text.textContent = balanceStr + ' · ~' + hours + 'h';
-    } else if (burn != null) {
-      text.textContent = balanceStr;
-    } else {
-      text.textContent = balanceStr;
-    }
-
-    // Apply chip accent colours (border + text tint)
    for (const [k, v] of Object.entries(chipStyle)) {
-      if (k !== '_unset') chip.style.setProperty(k, v);
+      chip.style.setProperty(k, v);
    }
    if (chipStyle['--chip-accent']) {
      text.style.color = chipStyle['--chip-accent'];
    }

-    // Build operator-readable tooltip
-    let title = 'Account balance';
-    if (burn != null) title = '$' + quota.balance_usd.toFixed(2) + ' · ~' + (Math.round(burn * 100) / 100) + '/hr';
-    if (hours != null) title += ' · ~' + hours + 'h runway';
-    text.setAttribute('title', title);
+    // Tooltip lists every provider plus its staleness
+    const tooltipLines = providers.map(name => {
+      const p = byProvider[name] || {};
+      const bal = p.balance_usd;
+      const burn = p.burn_rate_usd_per_hour;
+      const parts = [escapeHtml(name)];
+      if (bal != null && typeof bal === 'number') {
+        parts.push('$' + bal.toFixed(2));
+      } else {
+        parts.push('balance unknown');
+      }
+      if (burn != null && typeof burn === 'number') parts.push('~$' + burn.toFixed(2) + '/hr');
+      const proj = p.projected_hours_remaining;
+      if (proj != null && typeof proj === 'number') parts.push('~' + Math.round(proj) + 'h');
+      const suffix = formatBalanceStaleness(p.balance_at, p.balance_source);
+      if (suffix) parts.push(suffix);
+      return parts.join(' · ');
+    });
+    text.setAttribute('title', tooltipLines.join('\n'));
    fill.style.width = `${Math.min(quota.metered_fraction_of_plan * 100, 100)}%`;
    return;
  }
-  // ── fallback: percentage only ------------------------------------------------
+  // ── fallback: percentage only ----------------------------------------
  const pct = Math.min(quota.metered_fraction_of_plan * 100, 100);
  fill.style.width = `${pct}%`;
  text.textContent = `${(quota.metered_fraction_of_plan * 100).toFixed(1)}%`;
--- a/config/config.yaml
+++ b/config/config.yaml
@@ -104,6 +104,27 @@ objective:
  # peak, far below the dozens/hour a deprecation flare produces.
  rejection_warning_min_count: 6

+  # Credit-aware routing attenuation (OFF BY DEFAULT).
+  # When enabled, a provider configured with a balance_url (e.g. OpenRouter's
+  # prepaid account balance) gets its *comparison cost* inflated inside the
+  # quality-first ranker as its account balance nears zero. Quality bands still
+  # win; this only shifts ties. Providers whose balance comes from per-request
+  # energy telemetry allowance_remaining_usd (e.g. NeuralWatt's overage-billed
+  # subscription) are ALWAYS multiplier 1.0 regardless of their reading, so a
+  # low soft_floor_usd never biases routing toward an attenuated provider just
+  # because a telemetry provider's allowance reads near zero under normal use.
+  #
+  # This knob is deliberately NOT exposed in the admin UI persisted-config
+  # allowlist (_CONFIG_ALLOWLIST in admin.py) or _ProviderUpdateBody; enabling
+  # or tuning it requires editing this file and restarting llm-router.service
+  # (the dispatcher's module-level cfg binds at import).
+  credit_attenuation:
+    enabled: false
+    soft_floor_usd: 5.0          # balance >= this -> multiplier 1.0
+    zero_floor_usd: 0.0          # balance <= this -> max_multiplier
+    max_multiplier: 5.0          # maximum cost-inflation at/below zero floor
+    refresh_seconds: 300         # cache duration for resolved multipliers
+
 context:
  safety_factor: 0.75          # fraction of advertised context treated as usable
  default_output_reserve_tokens: 4096
@@ -735,6 +756,9 @@ dispatch_providers:
  openrouter:
    base_url: "https://openrouter.ai/api/v1"
    api_key_env: "OPENROUTER_API_KEY"
+    # Account-level balance poll; OpenRouter reports per-completion energy via
+    # a separate allowance_remaining field, so this URL is for the prepaid pool.
+    balance_url: "https://openrouter.ai/api/v1/credits"
    has_energy_telemetry: false
    enabled: true

--- a/config/schema.sql
+++ b/config/schema.sql
@@ -307,3 +307,14 @@ CREATE TABLE IF NOT EXISTS local_energy_observations (
 );
 CREATE INDEX IF NOT EXISTS idx_local_energy_model ON local_energy_observations (model_id);
 CREATE INDEX IF NOT EXISTS idx_local_energy_request ON local_energy_observations (request_id);
+
+-- Account-level balance observations per dispatch provider (separate from
+-- energy_observations because this is a prepaid-account pool, not a
+-- per-completion telemetry measure).
+CREATE TABLE IF NOT EXISTS provider_balance_observations (
+    id          INTEGER PRIMARY KEY AUTOINCREMENT,
+    provider    TEXT NOT NULL,
+    balance_usd REAL NOT NULL,
+    observed_at TEXT NOT NULL
+);
+CREATE INDEX IF NOT EXISTS idx_provider_balance ON provider_balance_observations (provider, observed_at);
--- a/docs/admin-portal.md
+++ b/docs/admin-portal.md
@@ -179,7 +179,7 @@ the attributable set, so `POST /outcome` keeps training proficiency normally.

 **Read-only dashboards** — `GET /admin/api/snapshot` exposes data for:

- quota burn against `objective.plan_kwh_per_period`
+- quota: energy burn against `objective.plan_kwh_per_period`, plus per-provider balance, burn rate, and runway under `by_provider`, and a `total_balance_usd` sum (old flat balance/burn/runway keys removed)
 - per-model usage from `energy_observations`
 - live routing decisions from `route_decisions`
 - verdict mix and scoring coverage
--- a/src/config.py
+++ b/src/config.py
@@ -36,6 +36,67 @@ class StrictModel(BaseModel):
    model_config = ConfigDict(extra="forbid")


+class CreditAttenuationConfig(StrictModel):
+    """Optional cost-inflation tiebreak for providers with a polled account balance.
+
+    When enabled, a provider whose prepaid balance is near depletion gets its
+    comparison cost multiplied inside the quality-first ranking. Quality bands
+    still win; the multiplier only breaks ties. Telemetry providers (those
+    billed per-completion via ``allowance_remaining_usd``) are always treated
+    as multiplier 1.0 — their "low" reading is normal overage-invoice noise,
+    not a depleting pool, so attenuation never applies to them.
+    """
+
+    enabled: bool = False
+    soft_floor_usd: float = 5.0
+    zero_floor_usd: float = 0.0
+    max_multiplier: float = 5.0
+    refresh_seconds: int = 300
+
+    @field_validator("soft_floor_usd")
+    @classmethod
+    def soft_floor_positive(cls, v: float) -> float:
+        if v <= 0:
+            raise ValueError("objective.credit_attenuation.soft_floor_usd must be > 0")
+        return v
+
+    @field_validator("zero_floor_usd")
+    @classmethod
+    def zero_floor_non_negative(cls, v: float) -> float:
+        if v < 0:
+            raise ValueError(
+                "objective.credit_attenuation.zero_floor_usd must be >= 0"
+            )
+        return v
+
+    @field_validator("max_multiplier")
+    @classmethod
+    def max_multiplier_above_one(cls, v: float) -> float:
+        if v <= 1.0:
+            raise ValueError(
+                "objective.credit_attenuation.max_multiplier must be > 1.0"
+            )
+        return v
+
+    @field_validator("refresh_seconds")
+    @classmethod
+    def refresh_seconds_positive(cls, v: int) -> int:
+        if v <= 0:
+            raise ValueError(
+                "objective.credit_attenuation.refresh_seconds must be > 0"
+            )
+        return v
+
+    @model_validator(mode="after")
+    def zero_below_soft(self) -> "CreditAttenuationConfig":
+        if self.zero_floor_usd >= self.soft_floor_usd:
+            raise ValueError(
+                "objective.credit_attenuation.zero_floor_usd "
+                "must be < soft_floor_usd"
+            )
+        return self
+
+
 class Objective(StrictModel):
    """What the router optimizes: quality, bounded by cost.

@@ -57,6 +118,7 @@ class Objective(StrictModel):
    rejection_warning_window_hours: Optional[int] = None
    rejection_warning_baseline_hours: Optional[int] = None
    rejection_warning_min_count: Optional[int] = None
+    credit_attenuation: CreditAttenuationConfig = CreditAttenuationConfig()

    @field_validator("quality_tolerance")
    @classmethod
@@ -884,9 +946,27 @@ class LocalDispatchModel(StrictModel):
 class DispatchProvider(StrictModel):
    base_url: str
    api_key_env: str
+    balance_url: Optional[str] = None
    has_energy_telemetry: bool = False
    enabled: bool = True

+    @field_validator("balance_url")
+    @classmethod
+    def balance_url_requires_https(cls, v: Optional[str]) -> Optional[str]:
+        if v is None:
+            return v
+        if urlparse(v).scheme != "https":
+            raise ValueError(
+                f"dispatch_providers[...].balance_url must use https://, got {v!r}"
+            )
+        return v
+
+
+# Names-only balance-parser registry. poller imports config, so config cannot
+# import poller's parser dict; the poller implements the parser and tests pin
+# the two sets equal.
+PROVIDERS_WITH_BALANCE_PARSERS: frozenset[str] = frozenset({"openrouter"})
+

 class LoggingConfig(StrictModel):
    # log_path is gone. Nothing ever wrote a file: the dispatcher logs to
@@ -1230,6 +1310,37 @@ class RouterConfig(StrictModel):
        """
        return self._dispatch_meterable_cache

+    @model_validator(mode="after")
+    def balance_url_on_provider_with_parser_and_no_telemetry(
+        self,
+    ) -> "RouterConfig":
+        """balance_url may only be set for providers with a parser and no telemetry.
+
+        A balance_url asks the poller to query a provider account endpoint.
+        Providers that already report balance per completion via
+        has_energy_telemetry would create two contradictory sources; fail at
+        load rather than silently picking one. A provider whose key has no
+        parser implementation would fail quietly every poll cycle.
+        """
+        for provider, prov_cfg in self.dispatch_providers.items():
+            if prov_cfg.balance_url is None:
+                continue
+            if getattr(prov_cfg, "has_energy_telemetry", False):
+                raise ValueError(
+                    f"dispatch_providers[{provider!r}] has both "
+                    f"balance_url and has_energy_telemetry=true. "
+                    f"Telemetry providers already report balance per completion; "
+                    f"configure one source or the other."
+                )
+            if provider not in PROVIDERS_WITH_BALANCE_PARSERS:
+                valid = sorted(PROVIDERS_WITH_BALANCE_PARSERS)
+                raise ValueError(
+                    f"dispatch_providers[{provider!r}].balance_url is set, "
+                    f"but {provider!r} has no balance parser implementation. "
+                    f"Providers with parsers: {valid}"
+                )
+        return self
+
    @model_validator(mode="after")
    def verifier_model_is_stated_once_the_hosts_differ(self) -> "RouterConfig":
        """A remote classifier must not lend its model name to the verifier.
--- a/src/dispatcher.py
+++ b/src/dispatcher.py
@@ -80,6 +80,7 @@ from metrics import (
    local_energy_summary,
    per_model,
    pinch_summary,
+    quota_balance_and_burn,
    quota_burn,
    recent_decisions,
    scoring_coverage,
@@ -90,6 +91,7 @@ from routing import (
    apply_flex_preference,
    capability_gate_reason,
    cheapest_classifier_candidate,
+    credit_attenuation_multiplier,
    parse_eligible_categories,
    rank_candidates,
    rejection_reason,
@@ -668,6 +670,59 @@ def _resolve_auto_classifier() -> Optional[dict]:
    return picked


+_provider_cost_multipliers_cache: Optional[dict[str, float]] = None
+_provider_cost_multipliers_resolved_at: float = 0.0
+
+
+def _provider_cost_multipliers() -> Optional[dict[str, float]]:
+    """Cached per-provider cost multipliers for credit attenuation.
+
+    Returns ``None`` when the feature is disabled so the routing path performs
+    no database work and behaves identically to the pre-attenuation code.
+
+    When enabled, only providers configured with ``balance_url`` (the polled
+    account-balance path, e.g. OpenRouter) receive a multiplier; telemetry
+    providers such as NeuralWatt, whose ``allowance_remaining_usd`` reflects
+    overage accounting rather than a depleting prepaid pool, are intentionally
+    omitted so their normal near-zero readings never bias routing.
+    """
+    global _provider_cost_multipliers_cache, _provider_cost_multipliers_resolved_at
+    att = cfg.objective.credit_attenuation
+    if not att.enabled:
+        return None
+
+    now = time.time()
+    if (
+        _provider_cost_multipliers_cache is not None
+        and now - _provider_cost_multipliers_resolved_at < att.refresh_seconds
+    ):
+        return _provider_cost_multipliers_cache
+
+    conn = _db()
+    try:
+        balances = quota_balance_and_burn(conn, cfg)
+    finally:
+        conn.close()
+
+    multipliers: dict[str, float] = {}
+    for provider, entry in balances["by_provider"].items():
+        prov_cfg = cfg.dispatch_providers.get(provider)
+        if prov_cfg is not None and getattr(prov_cfg, "balance_url", None):
+            multipliers[provider] = credit_attenuation_multiplier(
+                entry.get("balance_usd"),
+                soft_floor_usd=att.soft_floor_usd,
+                zero_floor_usd=att.zero_floor_usd,
+                max_multiplier=att.max_multiplier,
+            )
+
+    if any(v != 1.0 for v in multipliers.values()):
+        logs.info("credit_attenuation_resolved", provider_multipliers=multipliers)
+
+    _provider_cost_multipliers_cache = multipliers
+    _provider_cost_multipliers_resolved_at = now
+    return multipliers
+
+
 class _ClassifierSkipped(Exception):
    """The local classifier was not attempted at all -- gaming mode, or an
    already-open backoff circuit (see _local_classifier_skip_reason).
@@ -1481,6 +1536,7 @@ def route(
                    model=row["model_id"],
                    reason=rejection_reason(row, **filters),
                )
+    provider_cost_multipliers = _provider_cost_multipliers()
    ranked = rank_candidates(
        eligible,
        quality_tolerance=cfg.objective.quality_tolerance,
@@ -1489,6 +1545,7 @@ def route(
        prompt_tokens=classification.required_context_tokens,
        completion_tokens=cfg.objective.assumed_completion_tokens,
        cache_rate=cfg.objective.assumed_cache_rate,
+        provider_cost_multipliers=provider_cost_multipliers,
    )

    if logs.enabled_for_debug():
@@ -1499,6 +1556,11 @@ def route(
                model=row["model_id"],
                prof=row["proficiency_score"],
                est_usd=row.get("cost"),
+                provider_cost_multipliers=provider_cost_multipliers.get(
+                    row.get("provider")
+                )
+                if provider_cost_multipliers
+                else None,
                kwh=row.get("energy"),
            )

--- a/src/metrics.py
+++ b/src/metrics.py
@@ -10,9 +10,10 @@ optionally a ``RouterConfig`` instance; none rely on module-level globals.
 Functions
 ---------
 quota_burn      — kWh metered in the last 30 d and in the current billing
-                  period, against the plan allowance; also reports the account
-                  credit balance, burn rate and runway from
-                  allowance_remaining_usd
+                  period, against the plan allowance; also reports per-provider
+                  credit balance, burn rate and runway (from
+                  allowance_remaining_usd telemetry or the polled
+                  provider_balance_observations table) plus total_balance_usd
 scoring_coverage — which scoring axes actually have data
 capability_ceilings — vision / json_mode context-window sub-ceilings
 capability_demand_warnings — demand-relative warnings for those sub-ceilings
@@ -96,33 +97,26 @@ def _billing_period_start(billing_reset_day: int, today: Optional[date] = None)
    return date(prev_year, prev_month, billing_reset_day).isoformat()


-def quota_balance_and_burn(conn: sqlite3.Connection, cfg: Any) -> dict:
-    """Latest account balance and its burn/runway from ``allowance_remaining_usd``.
+def _balance_series_stats(
+    samples: list[tuple[datetime, float]],
+    *,
+    latest: Optional[tuple[str, float]],
+    burn_window_hours: float,
+    warning_hours: float,
+    min_samples: int,
+    min_hours: float,
+) -> dict[str, Any]:
+    """Compute balance/burn/runway for one provider-scoped series.

-    Pure read of ``energy_observations`` — no dispatcher import. The burn
-    rate is intentionally conservative: only the most recent
+    ``samples`` are the in-window (observed_at, balance) points, ordered
+    oldest-first; rows with unparseable timestamps are already skipped by the
+    caller. ``latest`` is the provider's most recent balance row regardless of
+    window (``(observed_at, balance_usd)``) and may be None.
+
+    The burn rate is intentionally conservative: only the latest
    monotonically-decreasing balance segment is used, and two guards stop a
    fresh credit top-up from producing a wild extrapolation.
    """
-    # Read optional knobs from config; treat None as unset and use code defaults.
-    # Test configs use SimpleNamespace without these attributes, so getattr
-    # must have a fallback and then a second default when the attr is None.
-    burn_window_hours = getattr(cfg.objective, "quota_burn_window_hours", None)
-    if burn_window_hours is None:
-        burn_window_hours = 24
-
-    warning_hours = getattr(cfg.objective, "quota_runway_warning_hours", None)
-    if warning_hours is None:
-        warning_hours = 6
-
-    min_samples = getattr(cfg.objective, "quota_burn_min_segment_samples", None)
-    if min_samples is None:
-        min_samples = 3
-
-    min_hours = getattr(cfg.objective, "quota_burn_min_segment_hours", None)
-    if min_hours is None:
-        min_hours = 0.5
-
    result: dict[str, Any] = {
        "balance_usd": None,
        "balance_at": None,
@@ -133,49 +127,20 @@ def quota_balance_and_burn(conn: sqlite3.Connection, cfg: Any) -> dict:
        "runway_note": None,
    }

-    # 1. Latest non-NULL allowance across ALL rows (no time-window filter).
-    balance_row = conn.execute(
-        """
-        SELECT observed_at, allowance_remaining_usd
-          FROM energy_observations
-         WHERE allowance_remaining_usd IS NOT NULL
-         ORDER BY observed_at DESC, id DESC
-         LIMIT 1
-        """
-    ).fetchone()
-    if balance_row is not None:
-        result["balance_usd"] = float(balance_row["allowance_remaining_usd"])
-        result["balance_at"] = balance_row["observed_at"]
+    if latest is not None:
+        result["balance_at"] = latest[0]
+        result["balance_usd"] = latest[1]

-    # 2. In-window rows for burn estimation.
-    rows = conn.execute(
-        """
-        SELECT observed_at, allowance_remaining_usd
-          FROM energy_observations
-         WHERE allowance_remaining_usd IS NOT NULL
-           AND julianday(observed_at) > julianday('now', '-' || ? || ' hours')
-         ORDER BY observed_at ASC, id ASC
-        """,
-        (str(burn_window_hours),),
-    ).fetchall()
-
-    if not rows:
+    if not samples:
        result["runway_note"] = (
            "burn estimate unavailable: no decreasing balance samples in the current window"
        )
        return result

-    # 3. Split into monotonically-decreasing segments at every balance INCREASE.
-    #    A top-up (credit jump) starts a new segment; only the latest survives.
+    # Split into monotonically-decreasing segments at every balance INCREASE.
+    # A top-up (credit jump) starts a new segment; only the latest survives.
    segments: list[list[tuple[datetime, float]]] = [[]]
-    for row in rows:
-        raw_ts = row["observed_at"]
-        try:
-            ts = datetime.fromisoformat(raw_ts)
-        except ValueError:
-            # Defensive: malformed timestamp would otherwise break /metrics.
-            continue
-        value = float(row["allowance_remaining_usd"])
+    for ts, value in samples:
        current = segments[-1]
        if not current:
            current.append((ts, value))
@@ -191,7 +156,6 @@ def quota_balance_and_burn(conn: sqlite3.Connection, cfg: Any) -> dict:
        )
        return result

-    # 4. Guarded burn-rate computation.
    if len(latest_segment) < min_samples:
        result["runway_note"] = (
            f"burn estimate unavailable: segment after last balance increase has only "
@@ -225,7 +189,6 @@ def quota_balance_and_burn(conn: sqlite3.Connection, cfg: Any) -> dict:
    burn_rate = round(total_decrease / elapsed_hours, 6)
    result["burn_rate_usd_per_hour"] = burn_rate

-    # 5. Projected runway and warning.
    balance = result["balance_usd"]
    if balance is not None and burn_rate > 0.0:
        projected = balance / burn_rate
@@ -235,6 +198,137 @@ def quota_balance_and_burn(conn: sqlite3.Connection, cfg: Any) -> dict:
    return result


+def quota_balance_and_burn(conn: sqlite3.Connection, cfg: Any) -> dict:
+    """Per-provider balance, burn and runway.
+
+    Returns ``{"by_provider": {provider: {balance_usd, balance_at,
+    balance_source, burn_window_hours, burn_rate_usd_per_hour,
+    projected_hours_remaining, runway_low_warning, runway_note}},
+    "total_balance_usd": float|None}`` with one entry per
+    ``cfg.dispatch_providers`` key.
+
+    The two-source rule:
+
+    * ``has_energy_telemetry`` providers read
+      ``energy_observations.allowance_remaining_usd`` scoped to their provider.
+    * Every other provider reads ``provider_balance_observations`` scoped to
+      its provider (the polled account-balance table), guarded with
+      ``sqlite3.OperationalError`` so a live DB predating that table degrades
+      to an all-None entry instead of failing ``/metrics``.
+
+    ``balance_source`` is CONFIG-derived and present even on all-None entries,
+    so a consumer can tell a prepaid-pool depletion story from an
+    overage-allowance accounting figure without re-deriving config:
+    ``"telemetry"`` (per-completion allowance), ``"polled"`` (balance_url
+    configured), ``"unconfigured"`` (neither).
+    """
+    # Read optional knobs from config; treat None as unset and use code defaults.
+    # Test configs use SimpleNamespace without these attributes, so getattr
+    # must have a fallback and then a second default when the attr is None.
+    burn_window_hours = getattr(cfg.objective, "quota_burn_window_hours", None)
+    if burn_window_hours is None:
+        burn_window_hours = 24
+
+    warning_hours = getattr(cfg.objective, "quota_runway_warning_hours", None)
+    if warning_hours is None:
+        warning_hours = 6
+
+    min_samples = getattr(cfg.objective, "quota_burn_min_segment_samples", None)
+    if min_samples is None:
+        min_samples = 3
+
+    min_hours = getattr(cfg.objective, "quota_burn_min_segment_hours", None)
+    if min_hours is None:
+        min_hours = 0.5
+
+    providers = getattr(cfg, "dispatch_providers", None) or {}
+    by_provider: dict[str, dict[str, Any]] = {}
+
+    for provider, pc in providers.items():
+        has_telemetry = getattr(pc, "has_energy_telemetry", False)
+        balance_url = getattr(pc, "balance_url", None)
+
+        if has_telemetry:
+            balance_source = "telemetry"
+            latest_sql = """
+                SELECT observed_at, allowance_remaining_usd
+                  FROM energy_observations
+                 WHERE provider = ?
+                   AND allowance_remaining_usd IS NOT NULL
+                 ORDER BY observed_at DESC, id DESC
+                 LIMIT 1
+            """
+            window_sql = """
+                SELECT observed_at, allowance_remaining_usd
+                  FROM energy_observations
+                 WHERE provider = ?
+                   AND allowance_remaining_usd IS NOT NULL
+                   AND julianday(observed_at) > julianday('now', '-' || ? || ' hours')
+                 ORDER BY observed_at ASC, id ASC
+            """
+        else:
+            balance_source = "polled" if balance_url else "unconfigured"
+            latest_sql = """
+                SELECT observed_at, balance_usd
+                  FROM provider_balance_observations
+                 WHERE provider = ?
+                 ORDER BY observed_at DESC, id DESC
+                 LIMIT 1
+            """
+            window_sql = """
+                SELECT observed_at, balance_usd
+                  FROM provider_balance_observations
+                 WHERE provider = ?
+                   AND julianday(observed_at) > julianday('now', '-' || ? || ' hours')
+                 ORDER BY observed_at ASC, id ASC
+            """
+
+        latest: Optional[tuple[str, float]] = None
+        raw_rows: list[sqlite3.Row] = []
+        try:
+            row = conn.execute(latest_sql, (provider,)).fetchone()
+            if row is not None and row[1] is not None:
+                latest = (row["observed_at"], float(row[1]))
+            raw_rows = conn.execute(
+                window_sql,
+                (provider, str(burn_window_hours)),
+            ).fetchall()
+        except sqlite3.OperationalError:
+            # Live DBs that predate the provider_balance_observations table
+            # must not 500 /metrics; report an all-None entry instead.
+            latest = None
+            raw_rows = []
+
+        samples: list[tuple[datetime, float]] = []
+        for row in raw_rows:
+            try:
+                ts = datetime.fromisoformat(row["observed_at"])
+            except ValueError:
+                # Defensive: malformed timestamp would otherwise break /metrics.
+                continue
+            samples.append((ts, float(row[1])))
+
+        entry = _balance_series_stats(
+            samples,
+            latest=latest,
+            burn_window_hours=burn_window_hours,
+            warning_hours=warning_hours,
+            min_samples=min_samples,
+            min_hours=min_hours,
+        )
+        entry["balance_source"] = balance_source
+        by_provider[provider] = entry
+
+    known_balances = [
+        entry["balance_usd"]
+        for entry in by_provider.values()
+        if entry["balance_usd"] is not None
+    ]
+    total_balance_usd = round(sum(known_balances), 6) if known_balances else None
+
+    return {"by_provider": by_provider, "total_balance_usd": total_balance_usd}
+
+
 def quota_burn(
    conn: sqlite3.Connection,
    cfg: Any,
@@ -249,10 +343,11 @@ def quota_burn(
    Returns a dict with ``plan_kwh``, ``metered_kwh_30d``,
    ``metered_kwh_period``, ``metered_fraction_of_plan``,
    ``metered_calls_30d``, ``reset_date`` (the billing-period start),
-    ``window_start_30d`` (the rolling 30-day window start), ``note`` and the
-    balance/burn/runway fields from ``quota_balance_and_burn``.
-    When ``cfg.objective.billing_reset_day`` is set, also returns
-    ``next_reset_date`` — the upcoming billing-period reset day.
+    ``window_start_30d`` (the rolling 30-day window start), ``note``, plus the
+    per-provider ``by_provider`` mapping and ``total_balance_usd`` from
+    ``quota_balance_and_burn``. When ``cfg.objective.billing_reset_day`` is
+    set, also returns ``next_reset_date`` — the upcoming billing-period reset
+    day.
    """
    # This gate removes the report only; it is intentionally not used to refuse
    # or alter request dispatch — routing decisions remain independent of quota.
@@ -262,28 +357,70 @@ def quota_burn(
    plan = cfg.objective.plan_kwh_per_period
    window_start_30d = (datetime.now(timezone.utc).date() - timedelta(days=30)).isoformat()

-    row = conn.execute(
-        """
-        SELECT COALESCE(SUM(energy_kwh), 0) kwh, COUNT(*) n
-          FROM energy_observations
-         WHERE julianday(observed_at) > julianday('now', '-30 days')
-        """
-    ).fetchone()
-    metered_kwh_30d = round(float(row["kwh"]), 5)
-    metered_calls_30d = row["n"]
+    providers = getattr(cfg, "dispatch_providers", None)
+    if providers is None:
+        # SimpleNamespace fixture compatibility: cfg has no dispatch_providers,
+        # so keep the historical unscoped SUM/COUNT behavior.
+        telemetry_providers: Optional[list[str]] = None
+    else:
+        telemetry_providers = [
+            name
+            for name, pc in providers.items()
+            if getattr(pc, "has_energy_telemetry", False)
+        ]
+
+    if telemetry_providers == []:
+        # No provider can meter energy. Both sums are honestly zero; do not
+        # run the queries with an empty IN () clause (SQLite syntax error).
+        metered_kwh_30d = 0.0
+        metered_calls_30d = 0
+    else:
+        if telemetry_providers:
+            telemetry_placeholders = ",".join("?" * len(telemetry_providers))
+            thirty_where = (
+                "julianday(observed_at) > julianday('now', '-30 days') "
+                f"AND provider IN ({telemetry_placeholders})"
+            )
+            thirty_params = tuple(telemetry_providers)
+        else:
+            thirty_where = "julianday(observed_at) > julianday('now', '-30 days')"
+            thirty_params = ()
+        row = conn.execute(
+            f"""
+            SELECT COALESCE(SUM(energy_kwh), 0) kwh, COUNT(*) n
+              FROM energy_observations
+             WHERE {thirty_where}
+            """,
+            thirty_params,
+        ).fetchone()
+        metered_kwh_30d = round(float(row["kwh"]), 5)
+        metered_calls_30d = row["n"]

    reset_day = getattr(cfg.objective, "billing_reset_day", None)
    if reset_day is not None:
        period_start = _billing_period_start(reset_day)
-        period_row = conn.execute(
-            """
-            SELECT COALESCE(SUM(energy_kwh), 0) kwh
-              FROM energy_observations
-             WHERE julianday(observed_at) >= julianday(?)
-            """,
-            (period_start,),
-        ).fetchone()
-        metered_kwh_period = round(float(period_row["kwh"]), 5)
+        if telemetry_providers == []:
+            metered_kwh_period = 0.0
+        else:
+            if telemetry_providers:
+                period_placeholders = ",".join("?" * len(telemetry_providers))
+                period_where = (
+                    "julianday(observed_at) >= julianday(?) "
+                    f"AND provider IN ({period_placeholders})"
+                )
+                period_params = (period_start, *telemetry_providers)
+            else:
+                period_where = "julianday(observed_at) >= julianday(?)"
+                period_params = (period_start,)
+            period_row = conn.execute(
+                f"""
+                SELECT COALESCE(SUM(energy_kwh), 0) kwh
+                  FROM energy_observations
+                 WHERE {period_where}
+                """,
+                period_params,
+            ).fetchone()
+            metered_kwh_period = round(float(period_row["kwh"]), 5)
        metered_fraction_of_plan = round(metered_kwh_period / plan, 4)
        reset_date = period_start
    else:
@@ -302,9 +439,11 @@ def quota_burn(
        "note": "router-metered only; traffic bypassing the router is not counted",
    }

-    # Merge the balance/burn/runway block; quota_balance_and_burn also supplies
-    # the burn_window_hours default.
-    result.update(quota_balance_and_burn(conn, cfg))
+    # Merge the per-provider balance/burn/runway block. The old flat keys are
+    # deliberately gone: every consumer reads by_provider.
+    balance = quota_balance_and_burn(conn, cfg)
+    result["by_provider"] = balance["by_provider"]
+    result["total_balance_usd"] = balance["total_balance_usd"]

    if reset_day is not None:
        result["next_reset_date"] = _next_reset_date(reset_day)
--- a/src/poller.py
+++ b/src/poller.py
@@ -22,16 +22,17 @@ migration.

 from __future__ import annotations

+import os
 import sqlite3
 import sys
 import warnings
 from dataclasses import dataclass
 from datetime import datetime, timezone
-from typing import Optional
+from typing import Callable, Optional

 import requests

-from config import LocalDispatchModel, RouterConfig, load_config
+from config import DispatchProvider, LocalDispatchModel, RouterConfig, load_config
 from tier import apply_tiering

 MODELS_URL = "https://api.neuralwatt.com/v1/models"
@@ -58,6 +59,15 @@ SUFFIX_FAST = "fast"
 SUFFIX_SHORT = "short"
 SERVING_SUFFIXES = frozenset({SUFFIX_FLEX, SUFFIX_FAST, SUFFIX_SHORT})

+# Provider-specific account-balance parsers. The registry lives inside the
+# poller because poller imports config, and config importing this dict would
+# create a circular dependency. Tests pin this set equal to
+# config.PROVIDERS_WITH_BALANCE_PARSERS.
+_BALANCE_PARSERS: dict[str, Callable[[dict], float]] = {
+    "openrouter": lambda payload: payload["data"]["total_credits"]
+    - payload["data"]["total_usage"]
+}
+

 class CatalogTooSmall(requests.RequestException):
    """Raised when the fetched catalog has zero rows — not a transient error,
@@ -67,6 +77,14 @@ class CatalogTooSmall(requests.RequestException):
    exit with code 1 without any extra machinery."""


+def _auth_header(prov_cfg: DispatchProvider) -> Optional[dict[str, str]]:
+    """Return a Bearer header if the provider's API key env var is set."""
+    key = os.environ.get(prov_cfg.api_key_env)
+    if not key:
+        return None
+    return {"Authorization": f"Bearer {key}"}
+
+
 def parse_serving_class(model_id: str) -> tuple[str, str, str]:
    """Derive (latency_class, reasoning_mode, context_variant) from a model id.

@@ -380,6 +398,40 @@ def upsert(conn: sqlite3.Connection, rows: list[ModelRow], cfg: RouterConfig) ->
    conn.commit()


+def _ensure_provider_balance_table(conn: sqlite3.Connection) -> None:
+    """Idempotently create the provider_balance_observations table and index.
+
+    Live router.db files that predate this feature lack the table; the
+    CREATE TABLE IF NOT EXISTS / CREATE INDEX IF NOT EXISTS DDL is safe to
+    re-run on every poll.
+    """
+    conn.execute("""
+        CREATE TABLE IF NOT EXISTS provider_balance_observations (
+            id          INTEGER PRIMARY KEY AUTOINCREMENT,
+            provider    TEXT NOT NULL,
+            balance_usd REAL NOT NULL,
+            observed_at TEXT NOT NULL
+        )
+    """)
+    conn.execute(
+        "CREATE INDEX IF NOT EXISTS idx_provider_balance "
+        "ON provider_balance_observations (provider, observed_at)"
+    )
+    conn.commit()
+
+
+def record_balance(
+    conn: sqlite3.Connection, provider: str, balance_usd: float
+) -> None:
+    """Write a provider account-balance observation row."""
+    conn.execute(
+        "INSERT INTO provider_balance_observations (provider, balance_usd, observed_at) "
+        "VALUES (?, ?, ?)",
+        (provider, balance_usd, datetime.now(timezone.utc).isoformat()),
+    )
+    conn.commit()
+
+
 def _ensure_models_eligible_categories(conn: sqlite3.Connection) -> None:
    """Idempotently add the models.eligible_categories column.

@@ -505,6 +557,7 @@ def main() -> int:
    conn.execute("PRAGMA foreign_keys = ON")

    upsert_local_dispatch_models(conn, cfg)
+    _ensure_provider_balance_table(conn)

    # Map provider keys (dispatch_providers dict keys) to their fetch functions.
    # Only entries with keys matching the key are dispatched; unknown keys
@@ -522,6 +575,26 @@ def main() -> int:
            print(f"[{provider}] skipped (disabled)")
            continue

+        if prov_cfg.balance_url:
+            headers = _auth_header(prov_cfg)
+            if headers is None:
+                print(
+                    f"[{provider}] balance poll skipped: "
+                    f"{prov_cfg.api_key_env} not set",
+                    file=sys.stderr,
+                )
+            else:
+                try:
+                    resp = requests.get(
+                        prov_cfg.balance_url, headers=headers, timeout=REQUEST_TIMEOUT
+                    )
+                    resp.raise_for_status()
+                    balance = _BALANCE_PARSERS[provider](resp.json())
+                    record_balance(conn, provider, balance)
+                    print(f"[{provider}] balance ${balance:.2f}")
+                except (requests.RequestException, KeyError, TypeError, ValueError) as e:
+                    print(f"[{provider}] balance poll FAILED: {e}", file=sys.stderr)
+
        fetcher = FETCHERS.get(provider)
        if fetcher is None:
            print(f"[{provider}] unknown provider — skipping (no fetcher)")
--- a/src/routing.py
+++ b/src/routing.py
@@ -549,6 +549,34 @@ def within_budget(row: dict, max_energy_kwh: float | None) -> bool:
    return energy is None or energy <= max_energy_kwh


+def credit_attenuation_multiplier(
+    balance_usd: float | None,
+    *,
+    soft_floor_usd: float,
+    zero_floor_usd: float,
+    max_multiplier: float,
+) -> float:
+    """Return a cost multiplier that nudges routing away from a low balance.
+
+    ``None`` or a balance at/above ``soft_floor_usd`` returns 1.0: the provider
+    participates normally. A balance at/below ``zero_floor_usd`` returns
+    ``max_multiplier``. Between the two floors the multiplier rises linearly
+    from 1.0 to ``max_multiplier``.
+
+    Unknown balance is treated as healthy. This matches the "NULL does not
+    disqualify" precedent used for ``min_tool_proficiency`` in
+    ``rejection_reason``: a provider whose balance we cannot read is left out
+    of the nudge rather than pushed toward the maximum attenuation. The routing
+    rule is fail-neutral for missing information.
+    """
+    if balance_usd is None or balance_usd >= soft_floor_usd:
+        return 1.0
+    if balance_usd <= zero_floor_usd:
+        return max_multiplier
+    frac = (soft_floor_usd - balance_usd) / (soft_floor_usd - zero_floor_usd)
+    return 1.0 + frac * (max_multiplier - 1.0)
+
+
 def rank_candidates(
    rows: Sequence[dict],
    *,
@@ -557,6 +585,7 @@ def rank_candidates(
    prompt_tokens: int = 0,
    completion_tokens: int = 500,
    cache_rate: float = 0.84,
+    provider_cost_multipliers: dict[str, float] | None = None,
 ) -> list[dict]:
    """Order candidates: best quality first, cheapest among equals.

@@ -588,6 +617,16 @@ def rank_candidates(
    Returns each row plus ``proficiency_score``, ``cost_score`` (reported for
    visibility only, no longer part of the decision) and ``composite``, which
    is now simply the effective quality after the tolerance band.
+
+    ``provider_cost_multipliers`` is an optional per-provider cost inflation
+    map used ONLY inside the quality-band tiebreak.  A provider mapped to
+    ``5.0`` is treated as five times as expensive for ranking, but its
+    ``cost`` and ``cost_score`` remain the honest catalog estimate.  This is
+    the credit-attenuation hook: it can nudge a near-tie toward a provider
+    with a healthier account balance, but it can never override a genuine
+    quality gap because the band is computed before the cost key.  When the
+    argument is ``None`` or omitted the sort falls back to the raw cost and
+    behavior is identical to the pre-attenuation implementation.
    """
    affordable = [r for r in rows if within_budget(r, max_energy_per_request)]

@@ -596,18 +635,35 @@ def rank_candidates(
    # has no price for keeps whatever measured cost it arrived with rather
    # than losing the field — a missing list price is not free.
    estimates = []
+    effective_estimates = []
    for r in affordable:
        est = estimated_cost(r, prompt_tokens, completion_tokens, cache_rate)
-        estimates.append(r.get("cost") if est is None else est)
+        raw = r.get("cost") if est is None else est
+        estimates.append(raw)
+        if provider_cost_multipliers and raw is not None:
+            multiplier = provider_cost_multipliers.get(r.get("provider"), 1.0)
+            effective_estimates.append(raw * multiplier)
+        else:
+            effective_estimates.append(raw)
    # cost_score is retained purely so callers can still see the spread; it
    # does not enter the ordering.
    cost_scores = cost_score(estimates)

    ranked = []
-    for row, c_s, est in zip(affordable, cost_scores, estimates):
+    for row, c_s, raw, effective in zip(
+        affordable, cost_scores, estimates, effective_estimates
+    ):
        p_s = proficiency_score(row.get("proficiency"))
-        ranked.append({**row, "cost": est, "cost_score": c_s,
-                       "proficiency_score": p_s, "composite": p_s})
+        ranked.append(
+            {
+                **row,
+                "cost": raw,
+                "effective_cost": effective,
+                "cost_score": c_s,
+                "proficiency_score": p_s,
+                "composite": p_s,
+            }
+        )

    if not ranked:
        return []
@@ -629,7 +685,9 @@ def rank_candidates(
    ranked.sort(
        key=lambda r: (
            band(r),
-            r["cost"] if r.get("cost") is not None else float("inf"),
+            r["effective_cost"]
+            if r.get("effective_cost") is not None
+            else float("inf"),
            r["model_id"],
        )
    )
--- a/src/tui.py
+++ b/src/tui.py
@@ -65,6 +65,51 @@ def _fmt_runway(hours: float) -> str:
    return f"~{hours / 24:.1f}d"


+def _quota_lead_values(rows: dict) -> tuple[Optional[float], Optional[float], Optional[float], bool, Optional[str]]:
+    providers = sorted(
+        {label[: -len(" balance_usd")] for label in rows if label.endswith(" balance_usd")}
+    )
+    eligible = [
+        p for p in providers if rows.get(f"{p} balance_source") == "polled"
+    ]
+    candidates = []
+    for provider in eligible:
+        balance_value = rows.get(f"{provider} balance_usd")
+        balance = _numeric_suffix_value(balance_value)
+        burn_value = rows.get(f"{provider} burn_rate_usd_per_hour")
+        burn = _numeric_suffix_value(burn_value)
+        runway_value = rows.get(f"{provider} projected_hours_remaining")
+        runway = _numeric_suffix_value(runway_value)
+        low = bool(rows.get(f"{provider} runway_low_warning"))
+        note = rows.get(f"{provider} runway_note")
+        candidates.append((provider, balance, burn, runway, low, note))
+
+    if not candidates:
+        return (None, None, None, False, None)
+
+    def _key(item):
+        _provider, balance, _burn, runway, _low, _note = item
+        runway_sort = runway if isinstance(runway, (int, float)) else float("inf")
+        balance_sort = balance if isinstance(balance, (int, float)) else float("inf")
+        return (runway_sort, balance_sort, item[0])
+
+    _, balance, burn, runway, low, note = min(candidates, key=_key)
+    return (balance, burn, runway, low, note)
+
+
+def _numeric_suffix_value(value: object) -> Optional[float]:
+    if value is None:
+        return None
+    if isinstance(value, (int, float)):
+        return float(value)
+    if isinstance(value, str):
+        try:
+            return float(value.split()[0])
+        except (IndexError, ValueError):
+            return None
+    return None
+
+
 def _format_quota_lead(
    balance: Optional[float],
    burn: Optional[float],
@@ -616,21 +661,10 @@ class DashboardApp(App):
        lead = self.query_one("#quota-lead", Static)
        note = self.query_one("#quota-note", Static)
        if rows:
-            balance = rows.get("balance_usd")
-            burn = rows.get("burn_rate_usd_per_hour")
-            lead.update(
-                _format_quota_lead(
-                    balance,
-                    burn,
-                    rows.get("projected_hours_remaining"),
-                    bool(rows.get("runway_low_warning")),
-                )
-            )
-            # Explain an absent burn rate rather than leaving a bare "n/a".
+            balance, burn, runway, low, note_text = _quota_lead_values(rows)
+            lead.update(_format_quota_lead(balance, burn, runway, low))
            if balance is not None and burn is None:
-                note.update(
-                    str(rows.get("runway_note") or "burn estimate unavailable")
-                )
+                note.update(str(note_text or "burn estimate unavailable"))
                note.display = True
            else:
                note.update("")
--- a/src/tui_model.py
+++ b/src/tui_model.py
@@ -55,13 +55,44 @@ def build_model(data: dict) -> dict:
            {"label": "calls", "value": quota.get("metered_calls_30d")},
            {"label": "reset_date", "value": quota.get("reset_date")},
            {"label": "next_reset_date", "value": quota.get("next_reset_date")},
-            {"label": "balance_usd", "value": quota.get("balance_usd")},
-            {"label": "burn_rate_usd_per_hour", "value": quota.get("burn_rate_usd_per_hour")},
-            {"label": "projected_hours_remaining", "value": quota.get("projected_hours_remaining")},
-            {"label": "runway_low_warning", "value": quota.get("runway_low_warning")},
-            {"label": "runway_note", "value": quota.get("runway_note")},
            {"label": "window_start_30d", "value": quota.get("window_start_30d")},
+            {"label": "total_balance_usd", "value": quota.get("total_balance_usd")},
        ]
+        for provider, entry in sorted((quota.get("by_provider") or {}).items()):
+            balance = entry.get("balance_usd")
+            balance_at = entry.get("balance_at")
+            balance_label = f"{provider} balance_usd"
+            if balance is None:
+                balance_value = None
+            elif balance_at:
+                balance_value = f"{balance} @ {balance_at}"
+            else:
+                balance_value = str(balance)
+            quota_rows.append({"label": balance_label, "value": balance_value})
+            quota_rows.append(
+                {"label": f"{provider} balance_source", "value": entry.get("balance_source")}
+            )
+            quota_rows.append(
+                {
+                    "label": f"{provider} burn_rate_usd_per_hour",
+                    "value": entry.get("burn_rate_usd_per_hour"),
+                }
+            )
+            quota_rows.append(
+                {
+                    "label": f"{provider} projected_hours_remaining",
+                    "value": entry.get("projected_hours_remaining"),
+                }
+            )
+            quota_rows.append(
+                {
+                    "label": f"{provider} runway_low_warning",
+                    "value": entry.get("runway_low_warning"),
+                }
+            )
+            quota_rows.append(
+                {"label": f"{provider} runway_note", "value": entry.get("runway_note")}
+            )
    else:
        quota_rows = []

--- a/tests/test_admin_frontend.py
+++ b/tests/test_admin_frontend.py
@@ -176,10 +176,12 @@ def test_quota_modal_has_not_configured_fallback():


 def test_admin_quota_chip_references_balance_and_runway():
-    """The index page references the new balance / runway quota fields."""
+    """The index page references per-provider balance / runway quota fields and the provider rows container."""
    index_path = ROOT / "admin" / "frontend" / "index.html"
    html = index_path.read_text()
    assert "balance_usd" in html and "runway_low_warning" in html
+    assert 'id="quota-provider-rows"' in html
+    assert "balance_at" in html


 def test_builtin_profile_cards_offer_duplicate(admin_client):
--- a/tests/test_balance_config.py
+++ b/tests/test_balance_config.py
@@ -0,0 +1,141 @@
+"""QA for DispatchProvider.balance_url + fail-closed validators."""
+
+from pathlib import Path
+
+import pytest
+from pydantic import ValidationError
+
+from config import (
+    PROVIDERS_WITH_BALANCE_PARSERS,
+    DispatchProvider,
+    RouterConfig,
+    _merge_overlay,
+    load_config,
+)
+
+
+CONFIG_PATH = Path("config/config.yaml")
+
+
+def _base_cfg_dict() -> dict:
+    """Return the real base config dict so mutants start from a valid tree."""
+    raw = load_config(CONFIG_PATH).model_dump()
+    # The real config currently has only "neuralwatt" in dispatch_providers.
+    return raw
+
+
+def _provider_dict(**overrides) -> dict:
+    base = {"base_url": "https://example.com/v1", "api_key_env": "EXAMPLE_KEY"}
+    base.update(overrides)
+    return base
+
+
+class TestDispatchProviderBalanceUrl:
+    def test_none_is_allowed(self):
+        p = DispatchProvider(base_url="https://x/v1", api_key_env="X")
+        assert p.balance_url is None
+
+    def test_https_url_is_allowed(self):
+        p = DispatchProvider(
+            base_url="https://x/v1",
+            api_key_env="X",
+            balance_url="https://x/credits",
+        )
+        assert p.balance_url == "https://x/credits"
+
+    def test_http_url_rejected(self):
+        with pytest.raises(ValidationError) as exc_info:
+            DispatchProvider(
+                base_url="https://x/v1",
+                api_key_env="X",
+                balance_url="http://insecure.example/credits",
+            )
+        assert "https://" in str(exc_info.value)
+        assert "http://insecure.example/credits" in str(exc_info.value)
+
+    def test_unknown_provider_key_fails(self):
+        """(a) balance_url on a provider without a parser is an error."""
+        raw = _base_cfg_dict()
+        raw["dispatch_providers"]["unknown_provider"] = _provider_dict(
+            balance_url="https://unknown.example/credits",
+        )
+        with pytest.raises(ValidationError) as exc_info:
+            RouterConfig(**raw)
+        err = str(exc_info.value)
+        assert "unknown_provider" in err
+        assert "no balance parser implementation" in err
+        assert "neuralwatt" not in err
+
+    def test_telemetry_provider_with_balance_url_fails(self):
+        """(b) balance_url + has_energy_telemetry is a double-source error."""
+        raw = _base_cfg_dict()
+        raw["dispatch_providers"]["neuralwatt"]["balance_url"] = (
+            "https://neuralwatt.example/credits"
+        )
+        raw["dispatch_providers"]["neuralwatt"]["has_energy_telemetry"] = True
+        with pytest.raises(ValidationError) as exc_info:
+            RouterConfig(**raw)
+        err = str(exc_info.value)
+        assert "neuralwatt" in err
+        assert "has_energy_telemetry=true" in err
+        assert "Telemetry providers already report balance per completion" in err
+
+    def test_openrouter_balance_url_loads(self):
+        """(c) openrouter with a balance URL loads fine."""
+        raw = _base_cfg_dict()
+        raw["dispatch_providers"]["openrouter"] = _provider_dict(
+            balance_url="https://openrouter.ai/api/v1/credits",
+        )
+        cfg = RouterConfig(**raw)
+        assert cfg.dispatch_providers["openrouter"].balance_url == (
+            "https://openrouter.ai/api/v1/credits"
+        )
+
+    def test_http_balance_url_rejected_at_provider_level(self):
+        """(d) http:// balance URL fails the field validator."""
+        raw = _base_cfg_dict()
+        raw["dispatch_providers"]["openrouter"] = _provider_dict(
+            balance_url="http://openrouter.example/credits",
+        )
+        with pytest.raises(ValidationError) as exc_info:
+            RouterConfig(**raw)
+        assert "https://" in str(exc_info.value)
+
+
+class TestRegistryAndRealConfig:
+    def test_registry_is_openrouter_only(self):
+        assert PROVIDERS_WITH_BALANCE_PARSERS == frozenset({"openrouter"})
+
+    def test_real_config_still_loads(self):
+        cfg = load_config(CONFIG_PATH)
+        assert "neuralwatt" in cfg.dispatch_providers
+        assert cfg.dispatch_providers["neuralwatt"].balance_url is None
+
+
+class TestAdminRoundTripPreservation:
+    """(e2) _merge_overlay preserves base-only balance_url through provider edits."""
+
+    def test_overlay_preserves_balance_url(self):
+        base = {
+            "dispatch_providers": {
+                "openrouter": _provider_dict(
+                    balance_url="https://openrouter.ai/api/v1/credits",
+                ),
+            },
+        }
+        overlay = {
+            "dispatch_providers": {
+                "openrouter": {
+                    "api_key_env": "OPENROUTER_API_KEY_EDITED",
+                    "timeout_seconds": 30,
+                },
+            },
+        }
+        merged = _merge_overlay(base, overlay)
+        assert merged["dispatch_providers"]["openrouter"]["balance_url"] == (
+            "https://openrouter.ai/api/v1/credits"
+        )
+        assert (
+            merged["dispatch_providers"]["openrouter"]["api_key_env"]
+            == "OPENROUTER_API_KEY_EDITED"
+        )
--- a/tests/test_credit_attenuation_config.py
+++ b/tests/test_credit_attenuation_config.py
@@ -0,0 +1,80 @@
+"""QA for Objective.credit_attenuation config block."""
+
+from pathlib import Path
+
+import pytest
+from pydantic import ValidationError
+
+from config import (
+    CreditAttenuationConfig,
+    Objective,
+    RouterConfig,
+    load_config,
+)
+
+
+CONFIG_PATH = Path("config/config.yaml")
+
+class TestCreditAttenuationConfigDefaults:
+    def test_default_constructed_disabled(self):
+        """(a) Objective() default constructs with credit_attenuation disabled."""
+        obj = Objective()
+        assert obj.credit_attenuation.enabled is False
+        assert obj.credit_attenuation.soft_floor_usd == 5.0
+        assert obj.credit_attenuation.zero_floor_usd == 0.0
+        assert obj.credit_attenuation.max_multiplier == 5.0
+        assert obj.credit_attenuation.refresh_seconds == 300
+
+
+class TestCreditAttenuationValidators:
+    def test_zero_floor_eq_soft_floor_rejected(self):
+        """(b) zero_floor_usd == soft_floor_usd raises."""
+        with pytest.raises(ValidationError) as exc_info:
+            CreditAttenuationConfig(soft_floor_usd=5.0, zero_floor_usd=5.0)
+        assert "zero_floor_usd" in str(exc_info.value)
+        assert "soft_floor_usd" in str(exc_info.value)
+
+    def test_max_multiplier_not_above_one_rejected(self):
+        """(c) max_multiplier=1.0 raises."""
+        with pytest.raises(ValidationError) as exc_info:
+            CreditAttenuationConfig(max_multiplier=1.0)
+        assert "max_multiplier" in str(exc_info.value)
+
+    def test_refresh_seconds_zero_rejected(self):
+        """(d) refresh_seconds=0 raises."""
+        with pytest.raises(ValidationError) as exc_info:
+            CreditAttenuationConfig(refresh_seconds=0)
+        assert "refresh_seconds" in str(exc_info.value)
+
+    def test_soft_floor_zero_rejected(self):
+        """(e) soft_floor_usd=0 raises."""
+        with pytest.raises(ValidationError) as exc_info:
+            CreditAttenuationConfig(soft_floor_usd=0.0)
+        assert "soft_floor_usd" in str(exc_info.value)
+
+    def test_zero_floor_negative_rejected(self):
+        """(f) zero_floor_usd<0 raises."""
+        with pytest.raises(ValidationError) as exc_info:
+            CreditAttenuationConfig(zero_floor_usd=-1.0)
+        assert "zero_floor_usd" in str(exc_info.value)
+
+
+class TestCreditAttenuationRealConfig:
+    def test_real_config_has_attenuation_disabled(self):
+        """(g) real load_config returns credit_attenuation.enabled False."""
+        cfg = load_config(Path("config/config.yaml"))
+        assert cfg.objective.credit_attenuation.enabled is False
+        assert cfg.objective.credit_attenuation.soft_floor_usd == 5.0
+        assert cfg.objective.credit_attenuation.zero_floor_usd == 0.0
+        assert cfg.objective.credit_attenuation.max_multiplier == 5.0
+        assert cfg.objective.credit_attenuation.refresh_seconds == 300
+
+
+class TestCreditAttenuationExtraForbid:
+    def test_unknown_key_inside_block_raises(self):
+        """(h) unknown key inside credit_attenuation raises via extra-forbid."""
+        raw = load_config(CONFIG_PATH).model_dump()
+        raw["objective"]["credit_attenuation"]["hard_floor_usd"] = 1.0
+        with pytest.raises(ValidationError) as exc_info:
+            RouterConfig(**raw)
+        assert "hard_floor_usd" in str(exc_info.value)
--- a/tests/test_credit_attenuation_routing.py
+++ b/tests/test_credit_attenuation_routing.py
@@ -0,0 +1,421 @@
+"""Dispatcher-level credit attenuation wiring tests.
+
+These tests exercise the cached resolver and its call sites on the real
+``dispatcher.route_endpoint`` TestClient path. They are intentionally
+end-to-end for the wiring: pure ``rank_candidates`` cases live in
+``tests/test_routing.py``.
+"""
+
+from __future__ import annotations
+
+import sqlite3
+from datetime import datetime, timedelta, timezone
+from pathlib import Path
+
+import pytest
+from starlette.testclient import TestClient
+
+import dispatcher
+
+ROOT = Path(__file__).resolve().parent.parent
+SCHEMA_SQL = (ROOT / "config" / "schema.sql").read_text()
+
+
+def _hours_ago(hours: float) -> str:
+    return (datetime.now(timezone.utc) - timedelta(hours=hours)).isoformat()
+
+
+def _make_db(tmp_path: Path) -> sqlite3.Connection:
+    conn = sqlite3.connect(str(tmp_path / "test.db"))
+    conn.row_factory = sqlite3.Row
+    conn.executescript(SCHEMA_SQL)
+    return conn
+
+
+def _seed_two_provider_models(conn: sqlite3.Connection) -> None:
+    """Two providers, two models: near-identical profile so the tiebreak decides.
+
+    OpenRouter is priced slightly BELOW NeuralWatt (0.29 vs 0.30 per 1M), so
+    the raw-cost order picks or-cheap. The attenuation multipliers must
+    overcome that gap to flip the selection to neuralwatt — which is what
+    makes tests (b)/(c)/(c2) observable rather than tautological: with the
+    wiring broken (multipliers never reaching rank_candidates), openrouter
+    wins and tests (c)/(c2) FAIL.
+    """
+    rows = [
+        # (model_id, provider, tier, context_window,
+        #  cost_per_1m_prompt, cost_per_1m_completion)
+        (
+            "nw-cheap",
+            "neuralwatt",
+            2,
+            131072,
+            0.30,
+            0.30,
+        ),
+        (
+            "or-cheap",
+            "openrouter",
+            2,
+            131072,
+            0.29,
+            0.29,
+        ),
+    ]
+    for model_id, provider, tier, ctx, prompt_cost, completion_cost in rows:
+        conn.execute(
+            """
+            INSERT INTO models (
+                model_id, provider, base_model_id, tier, context_window,
+                effective_context_window, max_output_tokens,
+                cost_per_1m_prompt, cost_per_1m_completion,
+                supports_vision, supports_json_mode,
+                latency_class, reasoning_mode, context_variant,
+                access_level, availability, last_updated
+            ) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?,
+                      1, 'standard', 'default', 'full', 'public', 'active',
+                      '2026-08-22T00:00:00+00:00')
+            """,
+            (
+                model_id,
+                provider,
+                model_id,
+                tier,
+                ctx,
+                ctx,
+                16384,
+                prompt_cost,
+                completion_cost,
+                1,
+            ),
+        )
+        conn.execute(
+            """
+            INSERT INTO proficiency
+                (model_id, provider, category, blended_score, source, last_updated)
+            VALUES (?, ?, 'coding_general', 0.80, 'self_eval_thin',
+                    '2026-01-01T00:00:00+00:00')
+            """,
+            (model_id, provider),
+        )
+    conn.commit()
+
+
+def _seed_provider_balance(conn: sqlite3.Connection, provider: str, balance: float) -> None:
+    conn.execute(
+        "INSERT INTO provider_balance_observations (provider, balance_usd, observed_at) "
+        "VALUES (?, ?, ?)",
+        (provider, balance, _hours_ago(0.1)),
+    )
+    conn.commit()
+
+
+def _seed_energy_allowance(
+    conn: sqlite3.Connection,
+    provider: str,
+    model_id: str,
+    allowance_remaining_usd: float,
+) -> None:
+    conn.execute(
+        "INSERT INTO energy_observations "
+        "(model_id, provider, task_category, completion_tokens, energy_kwh, "
+        "cost_usd, carbon_g_co2eq, attribution_ratio, observed_at, "
+        "allowance_remaining_usd) "
+        "VALUES (?, ?, 'coding_general', 100, 5.0e-05, 0.001, "
+        "2.4e-03, 0.25, ?, ?)",
+        (model_id, provider, _hours_ago(0.1), allowance_remaining_usd),
+    )
+    conn.commit()
+
+
+def _route_response(client: TestClient, task: str = "fix a bug") -> dict:
+    resp = client.post(
+        "/route",
+        json={
+            "task": task,
+            "task_category": "coding_general",
+            "task_tier": 2,
+            "required_context_tokens": 1000,
+        },
+    )
+    assert resp.status_code == 200, resp.text
+    return resp.json()
+
+
+@pytest.fixture(autouse=True)
+def _reset_provider_multiplier_cache():
+    """Every dispatcher-attenuation test starts with a cold cache.
+
+    This also prevents the resolver cache from leaking between tests.
+    """
+    setattr(dispatcher, "_provider_cost_multipliers_cache", None)
+    setattr(dispatcher, "_provider_cost_multipliers_resolved_at", 0.0)
+    yield
+    setattr(dispatcher, "_provider_cost_multipliers_cache", None)
+    setattr(dispatcher, "_provider_cost_multipliers_resolved_at", 0.0)
+
+
+@pytest.fixture
+def route_client(tmp_path, monkeypatch):
+    """A TestClient wired to a temp DB seeded with two cross-provider models."""
+    conn = _make_db(tmp_path)
+    _seed_two_provider_models(conn)
+    conn.close()
+
+    monkeypatch.setattr(dispatcher.cfg.database, "path", str(tmp_path / "test.db"))
+    monkeypatch.setattr(dispatcher.cfg.verification, "local_llm_enabled", False)
+    monkeypatch.setattr(dispatcher.cfg.routing, "require_vision", False)
+    monkeypatch.setattr(dispatcher.cfg.exploration, "enabled", False)
+    monkeypatch.setenv("NEURALWATT_API_KEY", "test-key")
+    monkeypatch.setenv("OPENROUTER_API_KEY", "test-key")
+
+    with TestClient(dispatcher.app) as client:
+        yield client
+
+
+# (a) OFF (default): the resolver never touches the DB, even if a call would raise.
+
+def test_credit_attenuation_off_by_default_no_db_read(
+    route_client, monkeypatch
+):
+    """With attenuation disabled, quota_balance_and_burn is never called."""
+    calls = []
+
+    def explode(*args, **kwargs):
+        calls.append((args, kwargs))
+        raise RuntimeError("resolver should not have been called")
+
+    monkeypatch.setattr(dispatcher, "quota_balance_and_burn", explode)
+
+    decision = _route_response(route_client)
+    assert decision["selected"] is not None
+    assert calls == []
+
+
+# (b) ON + all balances healthy → selection matches the no-attenuation baseline.
+
+def test_credit_attenuation_on_healthy_balances_matches_baseline(
+    route_client, monkeypatch
+):
+    """Healthy OpenRouter balance keeps the raw-cost order (OpenRouter first).
+
+    OpenRouter is priced slightly cheaper, so the OFF baseline selects it.
+    Enabling attenuation with a healthy balance must NOT change that — the
+    multiplier is 1.0 and the winner stays or-cheap.
+    """
+    db_path = dispatcher.cfg.database.path
+    conn = sqlite3.connect(db_path)
+    conn.row_factory = sqlite3.Row
+    _seed_provider_balance(conn, "openrouter", 50.0)
+    conn.close()
+
+    # OFF baseline: the cheaper model wins.
+    decision = _route_response(route_client)
+    assert decision["selected"]["provider"] == "openrouter"
+    assert decision["selected"]["model_id"] == "or-cheap"
+
+    # ON + healthy balance: multiplier 1.0, same winner.
+    monkeypatch.setattr(dispatcher.cfg.objective.credit_attenuation, "enabled", True)
+    monkeypatch.setattr(
+        dispatcher.cfg.objective.credit_attenuation, "soft_floor_usd", 5.0
+    )
+    decision = _route_response(route_client)
+    assert decision["selected"]["provider"] == "openrouter"
+
+
+# (c) ON + synthetic low OpenRouter balance flips the tiebreak toward the
+#     healthier NeuralWatt provider.
+
+def test_credit_attenuation_on_low_openrouter_balance_flips_to_healthy_provider(
+    route_client, monkeypatch
+):
+    """A near-zero OpenRouter balance must flip the pick to NeuralWatt.
+
+    OpenRouter is priced cheaper than NeuralWatt, so the only way neuralwatt
+    can win is a >1.0 multiplier applied to or-cheap's comparison cost inside
+    route(). If the dispatcher stopped passing provider_cost_multipliers into
+    rank_candidates, openrouter (raw-cheaper) would win and this test FAILS.
+    """
+    monkeypatch.setattr(dispatcher.cfg.objective.credit_attenuation, "enabled", True)
+    monkeypatch.setattr(
+        dispatcher.cfg.objective.credit_attenuation, "soft_floor_usd", 5.0
+    )
+    monkeypatch.setattr(
+        dispatcher.cfg.objective.credit_attenuation, "zero_floor_usd", 0.0
+    )
+    monkeypatch.setattr(
+        dispatcher.cfg.objective.credit_attenuation, "max_multiplier", 5.0
+    )
+
+    db_path = dispatcher.cfg.database.path
+    conn = sqlite3.connect(db_path)
+    conn.row_factory = sqlite3.Row
+    # 0.01 is just above the zero floor: multiplier ≈ 4.99, well past the
+    # ~3% raw-price gap between or-cheap and nw-cheap.
+    _seed_provider_balance(conn, "openrouter", 0.01)
+    conn.close()
+
+    decision = _route_response(route_client)
+    assert decision["selected"]["provider"] == "neuralwatt"
+    assert decision["selected"]["model_id"] == "nw-cheap"
+
+
+# (c2) Billing-shape guard: low NeuralWatt allowance (telemetry provider) is
+#      ignored; OpenRouter stays attenuated.
+
+def test_credit_attenuation_ignores_telemetry_provider_allowance(
+    route_client, monkeypatch
+):
+    """NeuralWatt's per-completion allowance is never attenuated.
+
+    OpenRouter is raw-cheaper, so neuralwatt winning here requires BOTH:
+    OpenRouter attenuated (low polled balance → ~5x comparison cost) AND
+    NeuralWatt left at multiplier 1.0 despite its negative telemetry
+    allowance. If the balance_url billing-shape guard broke and telemetry
+    providers were attenuated too, nw-cheap (~1.5 effective) would lose to
+    or-cheap (~1.45 effective) and this test FAILS. If the multiplier
+    passthrough broke, or-cheap wins raw and this test FAILS.
+    """
+    monkeypatch.setattr(dispatcher.cfg.objective.credit_attenuation, "enabled", True)
+    monkeypatch.setattr(
+        dispatcher.cfg.objective.credit_attenuation, "soft_floor_usd", 5.0
+    )
+    monkeypatch.setattr(
+        dispatcher.cfg.objective.credit_attenuation, "zero_floor_usd", 0.0
+    )
+    monkeypatch.setattr(
+        dispatcher.cfg.objective.credit_attenuation, "max_multiplier", 5.0
+    )
+
+    db_path = dispatcher.cfg.database.path
+    conn = sqlite3.connect(db_path)
+    conn.row_factory = sqlite3.Row
+    # NeuralWatt reading near zero (mirrors live overage-invoice noise).
+    _seed_energy_allowance(conn, "neuralwatt", "nw-cheap", -0.004)
+    # OpenRouter prepaid pool nearly empty.
+    _seed_provider_balance(conn, "openrouter", 0.01)
+    conn.close()
+
+    decision = _route_response(route_client)
+    assert decision["selected"]["provider"] == "neuralwatt"
+    assert decision["selected"]["model_id"] == "nw-cheap"
+
+
+# (d) Resolver caches: monkeypatch the binding; only the first call resolves,
+#     the immediate second call returns cached value.
+
+def test_credit_attenuation_resolver_caches_within_refresh_window(
+    route_client, monkeypatch
+):
+    """Two route calls within refresh_seconds resolve balances once."""
+    monkeypatch.setattr(dispatcher.cfg.objective.credit_attenuation, "enabled", True)
+
+    calls = []
+
+    def counting_quota(*args, **kwargs):
+        calls.append((args, kwargs))
+        return {
+            "by_provider": {
+                "openrouter": {
+                    "balance_usd": 10.0,
+                    "balance_source": "polled",
+                },
+                "neuralwatt": {
+                    "balance_usd": 10.0,
+                    "balance_source": "telemetry",
+                },
+            },
+            "total_balance_usd": 20.0,
+        }
+
+    monkeypatch.setattr(dispatcher, "quota_balance_and_burn", counting_quota)
+
+    _route_response(route_client)
+    _route_response(route_client)
+    assert len(calls) == 1
+
+
+# (e) Only balance_url providers appear in the multiplier dict.
+
+def test_provider_cost_multipliers_only_for_balance_url_providers(
+    route_client, monkeypatch
+):
+    """The resolver returns a dict containing only providers with balance_url."""
+    monkeypatch.setattr(dispatcher.cfg.objective.credit_attenuation, "enabled", True)
+
+    def fake_quota(*args, **kwargs):
+        return {
+            "by_provider": {
+                "openrouter": {
+                    "balance_usd": 1.0,
+                    "balance_source": "polled",
+                },
+                "neuralwatt": {
+                    "balance_usd": -0.004,
+                    "balance_source": "telemetry",
+                },
+            },
+            "total_balance_usd": 0.996,
+        }
+
+    monkeypatch.setattr(dispatcher, "quota_balance_and_burn", fake_quota)
+
+    multipliers = dispatcher._provider_cost_multipliers()
+    assert "openrouter" in multipliers
+    assert "neuralwatt" not in multipliers
+
+
+# (f) Disabled feature returns None.
+
+def test_provider_cost_multipliers_none_when_disabled(
+    route_client, monkeypatch
+):
+    """When credit_attenuation.enabled is False the resolver returns None."""
+    monkeypatch.setattr(dispatcher.cfg.objective.credit_attenuation, "enabled", False)
+
+    def explode(*args, **kwargs):
+        raise RuntimeError("should not resolve when disabled")
+
+    monkeypatch.setattr(dispatcher, "quota_balance_and_burn", explode)
+    assert dispatcher._provider_cost_multipliers() is None
+
+
+# (g) Stale cache refreshes outside refresh_seconds.
+
+def test_credit_attenuation_resolver_refreshes_after_refresh_seconds(
+    route_client, monkeypatch
+):
+    """A long-past resolution timestamp triggers a fresh DB read."""
+    monkeypatch.setattr(dispatcher.cfg.objective.credit_attenuation, "enabled", True)
+    monkeypatch.setattr(
+        dispatcher.cfg.objective.credit_attenuation, "refresh_seconds", 60
+    )
+
+    calls = []
+
+    def counting_quota(*args, **kwargs):
+        calls.append((args, kwargs))
+        return {
+            "by_provider": {
+                "openrouter": {
+                    "balance_usd": 10.0,
+                    "balance_source": "polled",
+                },
+                "neuralwatt": {
+                    "balance_usd": 10.0,
+                    "balance_source": "telemetry",
+                },
+            },
+            "total_balance_usd": 20.0,
+        }
+
+    monkeypatch.setattr(dispatcher, "quota_balance_and_burn", counting_quota)
+
+    # Warm the cache, then age it past the refresh window.
+    dispatcher._provider_cost_multipliers()
+    assert len(calls) == 1
+    dispatcher._provider_cost_multipliers_resolved_at = (
+        dispatcher._provider_cost_multipliers_resolved_at - 120
+    )
+    dispatcher._provider_cost_multipliers()
+    assert len(calls) == 2
--- a/tests/test_metrics.py
+++ b/tests/test_metrics.py
@@ -37,6 +37,7 @@ from metrics import (
    local_energy_summary,
    per_model,
    pinch_summary,
+    quota_balance_and_burn,
    quota_burn,
    recent_decisions,
    rejection_warnings,
@@ -223,20 +224,43 @@ def test_quota_burn_empty_db(tmp_path):
    assert result["metered_calls_30d"] == 0


-# --- quota_burn balance / burn / runway tests --------------------------------
+# Internal mirror of the provider_balance_observations DDL (schema.sql owns the
+# canonical version). Local creation keeps these tests hermetic against a
+# transiently older schema.sql in this shared worktree.
+_PROVIDER_BALANCE_TABLE_SQL = """
+CREATE TABLE IF NOT EXISTS provider_balance_observations (
+    id INTEGER PRIMARY KEY AUTOINCREMENT,
+    provider TEXT NOT NULL,
+    balance_usd REAL NOT NULL,
+    observed_at TEXT NOT NULL
+);
+CREATE INDEX IF NOT EXISTS idx_provider_balance
+    ON provider_balance_observations (provider, observed_at);
+"""
+
+
+def _telemetry_provider_cfg(**objective_overrides) -> SimpleNamespace:
+    """Build a cfg whose only dispatch provider is telemetry-backed."""
+    return SimpleNamespace(
+        objective=SimpleNamespace(**objective_overrides),
+        dispatch_providers={
+            "neuralwatt": SimpleNamespace(has_energy_telemetry=True),
+        },
+    )
+
+
+# --- quota_burn balance / burn / runway tests (per-provider) ------------------


 def test_quota_burn_with_top_up_ignores_credit_jump(tmp_path):
    """A credit top-up splits the window; burn uses only the latest segment.

-    Series before top-up: 3.00 → 2.00 → 1.00. Then balance jumps to 10.00 and
-    resumes decreasing 10.00 → 9.25 → 8.50. Burn must come from the post-top-up
+    Series before top-up: 3.00 -> 2.00 -> 1.00. Then balance jumps to 10.00 and
+    resumes decreasing 10.00 -> 9.25 -> 8.50. Burn must come from the post-top-up
    segment (1.50 USD / 1.5 h = 1.0 USD/h), not from the overall MAX-MIN which
    would invent a phantom burn of ~7.00 USD.
    """
-    cfg = SimpleNamespace(
-        objective=SimpleNamespace(plan_kwh_per_period=6.25, billing_reset_day=6)
-    )
+    cfg = _telemetry_provider_cfg(plan_kwh_per_period=6.25, billing_reset_day=6)
    conn = _make_db(tmp_path)
    now = _now()
    rows = [
@@ -257,19 +281,20 @@ def test_quota_burn_with_top_up_ignores_credit_jump(tmp_path):
    conn.commit()

    result = quota_burn(conn, cfg)
-    assert result["balance_usd"] == pytest.approx(8.50)
-    assert result["balance_at"] == rows[-1][2].isoformat()
-    assert result["burn_window_hours"] == 24
-    assert result["burn_rate_usd_per_hour"] == pytest.approx(1.0)
-    assert result["projected_hours_remaining"] == pytest.approx(8.5)
-    assert result["runway_low_warning"] is False
+    entry = result["by_provider"]["neuralwatt"]
+    assert entry["balance_usd"] == pytest.approx(8.50)
+    assert entry["balance_at"] == rows[-1][2].isoformat()
+    assert entry["balance_source"] == "telemetry"
+    assert entry["burn_window_hours"] == 24
+    assert entry["burn_rate_usd_per_hour"] == pytest.approx(1.0)
+    assert entry["projected_hours_remaining"] == pytest.approx(8.5)
+    assert entry["runway_low_warning"] is False
+    assert result["total_balance_usd"] == pytest.approx(8.50)


 def test_quota_burn_with_top_up_immediately_before_window_end_returns_none(tmp_path):
-    """Top-up with only a short final tail → burn is unavailable, never wild."""
-    cfg = SimpleNamespace(
-        objective=SimpleNamespace(plan_kwh_per_period=6.25, billing_reset_day=6)
-    )
+    """Top-up with only a short final tail -> burn is unavailable, never wild."""
+    cfg = _telemetry_provider_cfg(plan_kwh_per_period=6.25, billing_reset_day=6)
    conn = _make_db(tmp_path)
    now = _now()
    rows = [
@@ -290,18 +315,17 @@ def test_quota_burn_with_top_up_immediately_before_window_end_returns_none(tmp_p
    conn.commit()

    result = quota_burn(conn, cfg)
-    assert result["burn_rate_usd_per_hour"] is None
-    assert result["runway_note"] is not None
-    assert "burn estimate unavailable" in result["runway_note"]
-    assert result["projected_hours_remaining"] is None
-    assert result["runway_low_warning"] is False
+    entry = result["by_provider"]["neuralwatt"]
+    assert entry["burn_rate_usd_per_hour"] is None
+    assert entry["runway_note"] is not None
+    assert "burn estimate unavailable" in entry["runway_note"]
+    assert entry["projected_hours_remaining"] is None
+    assert entry["runway_low_warning"] is False


 def test_quota_burn_all_null_allowance_degrades_gracefully(tmp_path):
-    """All allowance_remaining_usd NULL → no balance, no burn, no false alarm."""
-    cfg = SimpleNamespace(
-        objective=SimpleNamespace(plan_kwh_per_period=6.25, billing_reset_day=6)
-    )
+    """All allowance_remaining_usd NULL -> no balance, no burn, no false alarm."""
+    cfg = _telemetry_provider_cfg(plan_kwh_per_period=6.25, billing_reset_day=6)
    conn = _make_db(tmp_path)
    now = _now()
    rows = [
@@ -319,20 +343,21 @@ def test_quota_burn_all_null_allowance_degrades_gracefully(tmp_path):

    result = quota_burn(conn, cfg)
    assert result is not None
-    assert result["balance_usd"] is None
-    assert result["balance_at"] is None
-    assert result["burn_rate_usd_per_hour"] is None
-    assert result["projected_hours_remaining"] is None
-    assert result["runway_low_warning"] is False
-    assert result["runway_note"] is not None
+    entry = result["by_provider"]["neuralwatt"]
+    assert entry["balance_usd"] is None
+    assert entry["balance_at"] is None
+    assert entry["balance_source"] == "telemetry"
+    assert entry["burn_rate_usd_per_hour"] is None
+    assert entry["projected_hours_remaining"] is None
+    assert entry["runway_low_warning"] is False
+    assert entry["runway_note"] is not None
+    assert result["total_balance_usd"] is None
    assert result["metered_kwh_30d"] == pytest.approx(5.0)


 def test_quota_burn_runway_warning_when_below_threshold(tmp_path):
    """Low balance and positive burn below warning threshold triggers warning."""
-    cfg = SimpleNamespace(
-        objective=SimpleNamespace(plan_kwh_per_period=6.25, billing_reset_day=6)
-    )
+    cfg = _telemetry_provider_cfg(plan_kwh_per_period=6.25, billing_reset_day=6)
    conn = _make_db(tmp_path)
    now = _now()
    rows = [
@@ -351,10 +376,56 @@ def test_quota_burn_runway_warning_when_below_threshold(tmp_path):
    conn.commit()

    result = quota_burn(conn, cfg)
-    assert result["burn_rate_usd_per_hour"] == pytest.approx(0.4)
-    assert result["projected_hours_remaining"] == pytest.approx(1.5)
-    assert result["runway_low_warning"] is True
-    assert result["burn_window_hours"] == 24
+    entry = result["by_provider"]["neuralwatt"]
+    assert entry["burn_rate_usd_per_hour"] == pytest.approx(0.4)
+    assert entry["projected_hours_remaining"] == pytest.approx(1.5)
+    assert entry["runway_low_warning"] is True
+    assert entry["burn_window_hours"] == 24
+
+
+def test_quota_burn_min_samples_guard_per_provider(tmp_path):
+    """A post-top-up segment with only 2 samples cannot compute a burn rate."""
+    cfg = _telemetry_provider_cfg(plan_kwh_per_period=6.25, billing_reset_day=6)
+    conn = _make_db(tmp_path)
+    now = _now()
+    conn.execute(
+        "INSERT INTO energy_observations "
+        "(model_id, provider, energy_kwh, allowance_remaining_usd, observed_at) "
+        "VALUES (?, 'neuralwatt', 0.001, ?, ?)",
+        ("m", 10.0, (now - timedelta(hours=2)).isoformat()),
+    )
+    conn.execute(
+        "INSERT INTO energy_observations "
+        "(model_id, provider, energy_kwh, allowance_remaining_usd, observed_at) "
+        "VALUES (?, 'neuralwatt', 0.001, ?, ?)",
+        ("m", 9.0, (now - timedelta(minutes=30)).isoformat()),
+    )
+    conn.commit()
+
+    result = quota_burn(conn, cfg)
+    entry = result["by_provider"]["neuralwatt"]
+    assert entry["burn_rate_usd_per_hour"] is None
+    assert "need 3" in entry["runway_note"]
+
+
+def test_quota_burn_min_hours_guard_per_provider(tmp_path):
+    """A 3-sample segment spanning only 10 minutes cannot compute a burn rate."""
+    cfg = _telemetry_provider_cfg(plan_kwh_per_period=6.25, billing_reset_day=6)
+    conn = _make_db(tmp_path)
+    now = _now()
+    for value, minutes_ago in ((10.0, 10), (9.0, 7), (8.0, 1)):
+        conn.execute(
+            "INSERT INTO energy_observations "
+            "(model_id, provider, energy_kwh, allowance_remaining_usd, observed_at) "
+            "VALUES (?, 'neuralwatt', 0.001, ?, ?)",
+            ("m", value, (now - timedelta(minutes=minutes_ago)).isoformat()),
+        )
+    conn.commit()
+
+    result = quota_burn(conn, cfg)
+    entry = result["by_provider"]["neuralwatt"]
+    assert entry["burn_rate_usd_per_hour"] is None
+    assert "need at least 0.5 h" in entry["runway_note"]


 def test_quota_burn_billing_period_kwh_excludes_rolling_window(tmp_path):
@@ -363,9 +434,7 @@ def test_quota_burn_billing_period_kwh_excludes_rolling_window(tmp_path):
    Rows are pinned on either side of each boundary so the expected values can
    be computed from first principles without relying on a helper.
    """
-    cfg = SimpleNamespace(
-        objective=SimpleNamespace(plan_kwh_per_period=6.25, billing_reset_day=6)
-    )
+    cfg = _telemetry_provider_cfg(plan_kwh_per_period=6.25, billing_reset_day=6)
    conn = _make_db(tmp_path)
    now = _now()
    today = now.date()
@@ -406,20 +475,19 @@ def test_quota_burn_billing_period_kwh_excludes_rolling_window(tmp_path):
    )

    result = quota_burn(conn, cfg)
+    assert "by_provider" in result
+    assert "total_balance_usd" in result
    assert "metered_kwh_period" in result
    assert "window_start_30d" in result
    assert result["metered_kwh_period"] == pytest.approx(expected_period)
    assert result["metered_kwh_30d"] == pytest.approx(expected_rolling)
    assert result["metered_fraction_of_plan"] == pytest.approx(expected_period / 6.25)
+    assert result["by_provider"]["neuralwatt"]["balance_source"] == "telemetry"


-def test_quota_burn_includes_balance_and_runway_keys(tmp_path):
-    """quota_burn returns the full balance/burn/runway key set, and the
-    standalone quota_balance_and_burn helper matches it.
-    """
-    cfg = SimpleNamespace(
-        objective=SimpleNamespace(plan_kwh_per_period=6.25, billing_reset_day=6)
-    )
+def test_quota_burn_per_provider_shape_matches_standalone_helper(tmp_path):
+    """quota_burn carries by_provider/total_balance_usd and no flat keys."""
+    cfg = _telemetry_provider_cfg(plan_kwh_per_period=6.25, billing_reset_day=6)
    conn = _make_db(tmp_path)
    now = _now()
    rows = [
@@ -438,50 +506,35 @@ def test_quota_burn_includes_balance_and_runway_keys(tmp_path):
    conn.commit()

    result = quota_burn(conn, cfg)
-    for key in (
-        "plan_kwh",
-        "metered_kwh_30d",
-        "metered_kwh_period",
-        "metered_fraction_of_plan",
-        "metered_calls_30d",
-        "note",
-        "reset_date",
-        "window_start_30d",
+    flat_balance_keys = {
        "balance_usd",
        "balance_at",
-        "burn_window_hours",
        "burn_rate_usd_per_hour",
        "projected_hours_remaining",
        "runway_low_warning",
        "runway_note",
-        "next_reset_date",
-    ):
-        assert key in result, f"missing key {key!r}"
-
-    from metrics import quota_balance_and_burn
-
-    balance_only = quota_balance_and_burn(conn, cfg)
-    for key in (
-        "balance_usd",
-        "balance_at",
        "burn_window_hours",
-        "burn_rate_usd_per_hour",
-        "projected_hours_remaining",
-        "runway_low_warning",
-        "runway_note",
-    ):
-        assert key in balance_only, f"missing key {key!r}"
-        assert balance_only[key] == result[key]
+    }
+    for key in flat_balance_keys:
+        assert key not in result, f"flat key {key!r} must not be top-level in quota_burn"
+
+    assert isinstance(result["by_provider"], dict)
+    assert "neuralwatt" in result["by_provider"]
+    entry = result["by_provider"]["neuralwatt"]
+    for key in flat_balance_keys | {"balance_source"}:
+        assert key in entry, f"per-provider key {key!r} missing"
+    assert result["total_balance_usd"] == pytest.approx(0.60)
+
+    standalone = quota_balance_and_burn(conn, cfg)
+    assert standalone["by_provider"]["neuralwatt"] == entry
+    assert standalone["total_balance_usd"] == result["total_balance_usd"]


 def test_quota_burn_reset_date_is_billing_period_start(tmp_path):
    """reset_date is the billing-period start; window_start_30d is rolling 30 d."""
-    cfg = SimpleNamespace(
-        objective=SimpleNamespace(plan_kwh_per_period=6.25, billing_reset_day=6)
-    )
+    cfg = _telemetry_provider_cfg(plan_kwh_per_period=6.25, billing_reset_day=6)
    conn = _make_db(tmp_path)
-    now = datetime.now(timezone.utc)
-    today = now.date()
+    today = _now().date()
    reset_day = 6
    if today.day >= reset_day:
        expected_period_start = date(today.year, today.month, reset_day).isoformat()
@@ -495,6 +548,8 @@ def test_quota_burn_reset_date_is_billing_period_start(tmp_path):
    result = quota_burn(conn, cfg)
    assert result["reset_date"] == expected_period_start
    assert result["window_start_30d"] == expected_rolling_start
+    assert "by_provider" in result
+    assert "neuralwatt" in result["by_provider"]

    cfg_unconfigured = SimpleNamespace(
        objective=SimpleNamespace(plan_kwh_per_period=6.25)
@@ -503,6 +558,340 @@ def test_quota_burn_reset_date_is_billing_period_start(tmp_path):
    assert "reset_date" in result_unconfigured
    assert result_unconfigured["reset_date"] is None
    assert result_unconfigured["window_start_30d"] == expected_rolling_start
+    assert result_unconfigured["by_provider"] == {}
+    assert result_unconfigured["total_balance_usd"] is None
+
+
+# --- per-provider QA cases -----------------------------------------------------
+
+
+def test_quota_balance_interleaved_two_providers_is_independent(tmp_path):
+    """THE regression test: interleaved two-provider series compute separately.
+
+    Two providers' rows are interleaved in time so a global view would mix
+    them. energy_observations also carries poison allowance rows for
+    openrouter (genuinely possible: dispatch rows carry foreign-shaped
+    allowance data) that are NEWER than every neuralwatt row — a global
+    "latest allowance" read would report 80.0 for everyone. Each provider's
+    entry must instead trace to its own source rows, with DIFFERENT burn
+    rates so a merged/globally-latest regression cannot pass silently.
+    """
+    cfg = SimpleNamespace(
+        objective=SimpleNamespace(plan_kwh_per_period=6.25, billing_reset_day=6),
+        dispatch_providers={
+            "neuralwatt": SimpleNamespace(has_energy_telemetry=True),
+            "openrouter": SimpleNamespace(has_energy_telemetry=False),
+        },
+    )
+    conn = _make_db(tmp_path)
+    now = _now()
+
+    # NeuralWatt telemetry series: burn 1.0 USD/h, balance 2.0, runway 2 h.
+    for bal, at in (
+        (5.0, now - timedelta(hours=4)),
+        (4.0, now - timedelta(hours=3)),
+        (3.0, now - timedelta(hours=2)),
+        (2.0, now - timedelta(hours=1)),
+    ):
+        conn.execute(
+            "INSERT INTO energy_observations "
+            "(model_id, provider, energy_kwh, allowance_remaining_usd, observed_at) "
+            "VALUES (?, 'neuralwatt', 0.001, ?, ?)",
+            ("m", bal, at.isoformat()),
+        )
+
+    # Poison: openrouter allowance rows in energy_observations NEWER than every
+    # neuralwatt row. A global-latest implementation would report 80.0.
+    for bal, at in (
+        (100.0, now - timedelta(minutes=50)),
+        (90.0, now - timedelta(minutes=35)),
+        (80.0, now - timedelta(minutes=10)),
+    ):
+        conn.execute(
+            "INSERT INTO energy_observations "
+            "(model_id, provider, energy_kwh, allowance_remaining_usd, observed_at) "
+            "VALUES (?, 'openrouter', 0.001, ?, ?)",
+            ("m", bal, at.isoformat()),
+        )
+
+    # OpenRouter's real source, offset 30 min from the neuralwatt rows so the
+    # series visibly interleave: burn 4.0 USD/h, balance 8.0, runway 2 h.
+    for bal, at in (
+        (20.0, now - timedelta(minutes=210)),
+        (16.0, now - timedelta(minutes=150)),
+        (12.0, now - timedelta(minutes=90)),
+        (8.0, now - timedelta(minutes=30)),
+    ):
+        conn.execute(
+            "INSERT INTO provider_balance_observations "
+            "(provider, balance_usd, observed_at) "
+            "VALUES ('openrouter', ?, ?)",
+            (bal, at.isoformat()),
+        )
+    conn.commit()
+
+    result = quota_balance_and_burn(conn, cfg)
+    by_provider = result["by_provider"]
+
+    nw = by_provider["neuralwatt"]
+    assert nw["balance_usd"] == pytest.approx(2.0)
+    assert nw["balance_at"] == (now - timedelta(hours=1)).isoformat()
+    assert nw["burn_rate_usd_per_hour"] == pytest.approx(1.0)
+    assert nw["projected_hours_remaining"] == pytest.approx(2.0)
+    assert nw["runway_low_warning"] is True
+    assert nw["balance_source"] == "telemetry"
+
+    orr = by_provider["openrouter"]
+    assert orr["balance_usd"] == pytest.approx(8.0)
+    assert orr["balance_at"] == (now - timedelta(minutes=30)).isoformat()
+    assert orr["burn_rate_usd_per_hour"] == pytest.approx(4.0)
+    assert orr["projected_hours_remaining"] == pytest.approx(2.0)
+    assert orr["runway_low_warning"] is True
+
+    # The two providers must differ where their series differ; the poison
+    # rows (80.0, or any burn computed over a mixed series) must not appear.
+    assert nw["balance_usd"] != orr["balance_usd"]
+    assert nw["burn_rate_usd_per_hour"] != orr["burn_rate_usd_per_hour"]
+    assert result["total_balance_usd"] == pytest.approx(10.0)
+
+
+def test_quota_balance_polled_series_equivalence(tmp_path):
+    """A polled series computes the same burn/runway as telemetry would."""
+    conn = _make_db(tmp_path)
+    now = _now()
+    for bal, hours_ago in ((5.0, 3), (4.0, 2), (3.0, 1)):
+        conn.execute(
+            "INSERT INTO provider_balance_observations "
+            "(provider, balance_usd, observed_at) "
+            "VALUES ('openrouter', ?, ?)",
+            (bal, (now - timedelta(hours=hours_ago)).isoformat()),
+        )
+        conn.execute(
+            "INSERT INTO energy_observations "
+            "(model_id, provider, energy_kwh, allowance_remaining_usd, observed_at) "
+            "VALUES (?, 'neuralwatt', 0.001, ?, ?)",
+            ("m", bal, (now - timedelta(hours=hours_ago)).isoformat()),
+        )
+    conn.commit()
+
+    polled_cfg = SimpleNamespace(
+        objective=SimpleNamespace(plan_kwh_per_period=6.25),
+        dispatch_providers={
+            "openrouter": SimpleNamespace(
+                has_energy_telemetry=False,
+                balance_url="https://openrouter.ai/api/v1/credits",
+            ),
+        },
+    )
+    telemetry_cfg = _telemetry_provider_cfg(plan_kwh_per_period=6.25)
+
+    polled = quota_balance_and_burn(conn, polled_cfg)["by_provider"]["openrouter"]
+    telemetry = quota_balance_and_burn(conn, telemetry_cfg)["by_provider"]["neuralwatt"]
+
+    assert polled["balance_usd"] == telemetry["balance_usd"] == pytest.approx(3.0)
+    assert polled["balance_at"] == telemetry["balance_at"]
+    assert polled["burn_rate_usd_per_hour"] == pytest.approx(1.0)
+    assert polled["burn_rate_usd_per_hour"] == telemetry["burn_rate_usd_per_hour"]
+    assert polled["projected_hours_remaining"] == telemetry["projected_hours_remaining"]
+    assert polled["runway_low_warning"] == telemetry["runway_low_warning"]
+    assert polled["balance_source"] == "polled"
+    assert telemetry["balance_source"] == "telemetry"
+
+
+def test_quota_balance_top_up_guard_per_provider(tmp_path):
+    """A top-up for one provider resets only that provider's segment."""
+    cfg = SimpleNamespace(
+        objective=SimpleNamespace(plan_kwh_per_period=6.25, billing_reset_day=6),
+        dispatch_providers={
+            "neuralwatt": SimpleNamespace(has_energy_telemetry=True),
+            "openrouter": SimpleNamespace(has_energy_telemetry=False),
+        },
+    )
+    conn = _make_db(tmp_path)
+    now = _now()
+
+    # NeuralWatt: top-up to 4.0 at -60m; the post-top-up segment (4.0 -> 2.0
+    # over 1 h) burns 2.0 USD/h. Without the reset, the flat 3.0 -> 2.0 span
+    # would report 0.5 USD/h.
+    for bal, at in (
+        (3.0, now - timedelta(hours=2)),
+        (2.0, now - timedelta(minutes=90)),
+        (4.0, now - timedelta(minutes=60)),  # top-up
+        (3.0, now - timedelta(minutes=40)),
+        (2.6, now - timedelta(minutes=20)),
+        (2.0, now),
+    ):
+        conn.execute(
+            "INSERT INTO energy_observations "
+            "(model_id, provider, energy_kwh, allowance_remaining_usd, observed_at) "
+            "VALUES (?, 'neuralwatt', 0.001, ?, ?)",
+            ("m", bal, at.isoformat()),
+        )
+
+    # OpenRouter: monotonic decrease, burn (8 - 2) / 2 h = 3.0 USD/h.
+    for bal, at in (
+        (8.0, now - timedelta(hours=2)),
+        (5.0, now - timedelta(hours=1)),
+        (2.0, now),
+    ):
+        conn.execute(
+            "INSERT INTO provider_balance_observations "
+            "(provider, balance_usd, observed_at) "
+            "VALUES ('openrouter', ?, ?)",
+            (bal, at.isoformat()),
+        )
+    conn.commit()
+
+    result = quota_balance_and_burn(conn, cfg)
+    assert result["by_provider"]["neuralwatt"]["burn_rate_usd_per_hour"] == pytest.approx(2.0)
+    assert result["by_provider"]["openrouter"]["burn_rate_usd_per_hour"] == pytest.approx(3.0)
+
+
+def test_quota_balance_no_data_provider_all_none(tmp_path):
+    """A configured provider with no rows gets an all-None entry and a note.
+
+    total_balance_usd sums only the providers that report a balance.
+    """
+    cfg = SimpleNamespace(
+        objective=SimpleNamespace(plan_kwh_per_period=6.25, billing_reset_day=6),
+        dispatch_providers={
+            "neuralwatt": SimpleNamespace(has_energy_telemetry=True),
+            "openrouter": SimpleNamespace(has_energy_telemetry=False),
+        },
+    )
+    conn = _make_db(tmp_path)
+    conn.execute(
+        "INSERT INTO energy_observations "
+        "(model_id, provider, energy_kwh, allowance_remaining_usd, observed_at) "
+        "VALUES (?, 'neuralwatt', 0.001, ?, ?)",
+        ("m", 5.0, _now().isoformat()),
+    )
+    conn.commit()
+
+    result = quota_balance_and_burn(conn, cfg)
+    assert result["by_provider"]["neuralwatt"]["balance_usd"] == pytest.approx(5.0)
+    openrouter = result["by_provider"]["openrouter"]
+    assert openrouter["balance_usd"] is None
+    assert openrouter["runway_note"] is not None
+    assert result["total_balance_usd"] == pytest.approx(5.0)
+
+
+def test_quota_balance_total_none_when_no_provider_reports(tmp_path):
+    """total_balance_usd is None when no provider reports a balance."""
+    cfg = SimpleNamespace(
+        objective=SimpleNamespace(plan_kwh_per_period=6.25),
+        dispatch_providers={
+            "openrouter": SimpleNamespace(has_energy_telemetry=False),
+        },
+    )
+    conn = _make_db(tmp_path)
+    result = quota_balance_and_burn(conn, cfg)
+    assert result["by_provider"]["openrouter"]["balance_usd"] is None
+    assert result["total_balance_usd"] is None
+
+
+def test_quota_balance_providers_absent_empty(tmp_path):
+    """A cfg without dispatch_providers yields an empty by_provider mapping."""
+    cfg = SimpleNamespace(objective=SimpleNamespace(plan_kwh_per_period=6.25))
+    conn = _make_db(tmp_path)
+    result = quota_balance_and_burn(conn, cfg)
+    assert result["by_provider"] == {}
+    assert result["total_balance_usd"] is None
+
+
+def test_quota_balance_source_correctness(tmp_path):
+    """balance_source reflects the config-derived source, even all-None."""
+    cfg = SimpleNamespace(
+        objective=SimpleNamespace(plan_kwh_per_period=6.25),
+        dispatch_providers={
+            "neuralwatt": SimpleNamespace(has_energy_telemetry=True),
+            "openrouter": SimpleNamespace(
+                has_energy_telemetry=False,
+                balance_url="https://openrouter.ai/api/v1/credits",
+            ),
+            "bare": SimpleNamespace(has_energy_telemetry=False),
+        },
+    )
+    conn = _make_db(tmp_path)
+    result = quota_balance_and_burn(conn, cfg)
+    by_provider = result["by_provider"]
+    assert by_provider["neuralwatt"]["balance_source"] == "telemetry"
+    assert by_provider["openrouter"]["balance_source"] == "polled"
+    assert by_provider["bare"]["balance_source"] == "unconfigured"
+    # Present on all-None entries, not just on providers with data.
+    for provider in ("neuralwatt", "openrouter", "bare"):
+        assert "balance_source" in by_provider[provider]
+
+
+def test_quota_balance_missing_provider_balance_table_guard(tmp_path):
+    """A live DB predating provider_balance_observations must not 500 /metrics.
+
+    The T1 schema landed after live DBs were created; the guard degrades the
+    polled provider to an all-None entry while telemetry providers keep
+    working.
+    """
+    cfg = SimpleNamespace(
+        objective=SimpleNamespace(plan_kwh_per_period=6.25),
+        dispatch_providers={
+            "neuralwatt": SimpleNamespace(has_energy_telemetry=True),
+            "openrouter": SimpleNamespace(has_energy_telemetry=False),
+        },
+    )
+    conn = _make_db(tmp_path)
+    now = _now()
+    for bal, hours_ago in ((5.0, 3), (4.0, 2), (3.0, 1)):
+        conn.execute(
+            "INSERT INTO energy_observations "
+            "(model_id, provider, energy_kwh, allowance_remaining_usd, observed_at) "
+            "VALUES (?, 'neuralwatt', 0.001, ?, ?)",
+            ("m", bal, (now - timedelta(hours=hours_ago)).isoformat()),
+        )
+    conn.commit()
+    conn.execute("DROP TABLE provider_balance_observations")
+    conn.commit()
+
+    result = quota_balance_and_burn(conn, cfg)
+    assert result["by_provider"]["neuralwatt"]["balance_usd"] == pytest.approx(3.0)
+    openrouter = result["by_provider"]["openrouter"]
+    assert openrouter["balance_usd"] is None
+    assert openrouter["runway_note"] is not None
+
+
+def test_quota_balance_malformed_timestamp_skipped(tmp_path):
+    """A malformed observed_at degrades to fewer samples, never a crash.
+
+    With the malformed row (6.0) skipped, the neuralwatt series holds only two
+    valid samples, so the min-samples guard fires. If the malformed row were
+    counted, three samples over 2 h would produce a 4.0 USD/h burn.
+    """
+    cfg = _telemetry_provider_cfg(plan_kwh_per_period=6.25)
+    conn = _make_db(tmp_path)
+    now = _now()
+    conn.execute(
+        "INSERT INTO energy_observations "
+        "(model_id, provider, energy_kwh, allowance_remaining_usd, observed_at) "
+        "VALUES (?, 'neuralwatt', 0.001, ?, ?)",
+        ("m", 10.0, (now - timedelta(hours=3)).isoformat()),
+    )
+    conn.execute(
+        "INSERT INTO energy_observations "
+        "(model_id, provider, energy_kwh, allowance_remaining_usd, observed_at) "
+        "VALUES (?, 'neuralwatt', 0.001, ?, ?)",
+        ("m", 6.0, ""),
+    )
+    conn.execute(
+        "INSERT INTO energy_observations "
+        "(model_id, provider, energy_kwh, allowance_remaining_usd, observed_at) "
+        "VALUES (?, 'neuralwatt', 0.001, ?, ?)",
+        ("m", 2.0, (now - timedelta(hours=1)).isoformat()),
+    )
+    conn.commit()
+
+    result = quota_balance_and_burn(conn, cfg)
+    entry = result["by_provider"]["neuralwatt"]
+    assert entry["balance_usd"] == pytest.approx(2.0)
+    assert entry["burn_rate_usd_per_hour"] is None
+    assert entry["runway_note"] is not None


 def test_local_energy_summary_reset_date_is_billing_period_start(tmp_path):
--- a/tests/test_metrics_endpoint.py
+++ b/tests/test_metrics_endpoint.py
@@ -226,24 +226,27 @@ def test_metrics_contains_no_session_dir(seeded_client):
    assert "session_dir" not in body


-def test_metrics_quota_carries_balance_and_runway_keys(seeded_client):
-    """GET /metrics quota section includes the new balance/burn/runway keys
-    alongside the existing kWh/keys. With no allowance seeded, kWh fields are
-    present and balance/burn values are None or empty, but keys must exist.
-    """
+def test_metrics_quota_carries_per_provider_balance_and_total(seeded_client):
+    """GET /metrics quota carries by_provider + total_balance_usd, no flat keys."""
    resp = seeded_client.get("/metrics")
    assert resp.status_code == 200
    data = resp.json()
    assert data["quota"] is not None
    quota = data["quota"]
-    legacy_keys = (
+    for key in (
        "plan_kwh",
        "metered_kwh_30d",
+        "metered_kwh_period",
        "metered_fraction_of_plan",
        "metered_calls_30d",
        "note",
-    )
-    new_keys = (
+        "window_start_30d",
+    ):
+        assert key in quota, f"missing quota key {key!r}"
+
+    assert "by_provider" in quota
+    assert "total_balance_usd" in quota
+    flat_balance_keys = (
        "balance_usd",
        "balance_at",
        "burn_window_hours",
@@ -251,11 +254,37 @@ def test_metrics_quota_carries_balance_and_runway_keys(seeded_client):
        "projected_hours_remaining",
        "runway_low_warning",
        "runway_note",
-        "window_start_30d",
-        "metered_kwh_period",
    )
-    for key in legacy_keys + new_keys:
-        assert key in quota, f"missing quota key {key!r}"
+    for key in flat_balance_keys:
+        assert key not in quota, f"flat key {key!r} must not be top-level in quota"
+
+    for provider, entry in quota["by_provider"].items():
+        for key in flat_balance_keys + ("balance_source",):
+            assert key in entry, f"per-provider key {key!r} missing for {provider}"
+
+
+def test_metrics_quota_survives_missing_provider_balance_table(tmp_path, monkeypatch):
+    """A live DB predating provider_balance_observations must not 500 /metrics."""
+    conn = _make_db(tmp_path)
+    _seed_models(conn)
+    _seed_decision(conn)
+    conn.execute("DROP TABLE provider_balance_observations")
+    conn.commit()
+    conn.close()
+
+    monkeypatch.setattr(dispatcher.cfg.database, "path", str(tmp_path / "test.db"))
+    monkeypatch.setattr(dispatcher.cfg.verification, "local_llm_enabled", False)
+    monkeypatch.setattr(dispatcher.cfg.routing, "require_vision", False)
+    monkeypatch.setattr(dispatcher.cfg.local_energy, "enabled", False)
+    monkeypatch.setenv("NEURALWATT_API_KEY", "test-key")
+
+    with TestClient(dispatcher.app) as client:
+        resp = client.get("/metrics")
+    assert resp.status_code == 200
+    quota = resp.json()["quota"]
+    assert "by_provider" in quota
+    for entry in quota["by_provider"].values():
+        assert entry["balance_usd"] is None


 def test_metrics_local_energy_omitted_when_disabled(seeded_client):
--- a/tests/test_multi_provider_poller.py
+++ b/tests/test_multi_provider_poller.py
@@ -244,20 +244,13 @@ def test_one_provider_fails_other_succeeds(tmp_db, monkeypatch):
    """Failure in one provider's fetch does not abort the other."""
    connect, cfg = tmp_db

-    call_count = 0
-
-    def _failing_get(url, timeout):
-        nonlocal call_count
-        call_count += 1
-        if "openrouter" in url:
-            # OpenRouter fetch goes first (dict iteration order)
-            if call_count == 1:
-                return _fake_response(
-                    {"data": [_openrouter_model("openai/gpt-6-astra")]}
-                )
-            # Second neuralwatt call
-            return _fake_response({"data": [_catalog_item("model_a")]})
-        # NeuralWatt fails
+    def _failing_get(url, timeout, **kwargs):
+        if "/credits" in url:
+            # OpenRouter account-balance poll happens before the catalog fetch.
+            return _fake_response({"data": {"total_credits": 50.0, "total_usage": 0.0}})
+        if "openrouter.ai" in url:
+            return _fake_response({"data": [_openrouter_model("openai/gpt-6-astra")]})
+        # NeuralWatt catalog fails
        raise requests.RequestException("network error")

    monkeypatch.setattr(poller, "load_config", lambda p: cfg)
@@ -674,3 +667,183 @@ def test_per_provider_mark_stale_scoped(tmp_path):
    assert stale_or == 0, "openrouter row should NOT be stale when scoped to neuralwatt"

    conn.close()
+
+
+# ---------------------------------------------------------------------------
+# Balance polling tests
+# ---------------------------------------------------------------------------
+
+
+OPENROUTER_BALANCE_URL = REAL_CFG.dispatch_providers["openrouter"].balance_url
+
+
+def _url_keyed_get(url: str, timeout: object, **kwargs: object) -> MagicMock:
+    """Dispatch mocked HTTP responses by URL for balance + catalog tests."""
+    if "/credits" in url:
+        return _fake_response(
+            {"data": {"total_credits": 50.0, "total_usage": 7.0}}
+        )
+    if "openrouter.ai" in url:
+        return _fake_response({"data": [_openrouter_model("openai/gpt-6-astra")]})
+    if "neuralwatt" in url:
+        return _fake_response({"data": [_catalog_item("nw-model-a")]})
+    raise RuntimeError(f"unexpected URL in test mock: {url}")
+
+
+def test_balance_poll_success_records_row(tmp_db, monkeypatch):
+    """An OpenRouter balance poll writes one provider_balance_observations row."""
+    connect, cfg = tmp_db
+
+    monkeypatch.setattr(poller, "load_config", lambda p: cfg)
+    monkeypatch.setattr(poller.requests, "get", _url_keyed_get)
+    monkeypatch.setenv("OPENROUTER_API_KEY", "sk-or-test")
+
+    exit_code = poller.main()
+    assert exit_code == 0
+
+    conn = connect()
+    try:
+        row = conn.execute(
+            "SELECT provider, balance_usd FROM provider_balance_observations "
+            "WHERE provider = ?",
+            ("openrouter",),
+        ).fetchone()
+        assert row is not None, "expected a balance row for openrouter"
+        assert row["provider"] == "openrouter"
+        assert row["balance_usd"] == 43.0
+    finally:
+        conn.close()
+
+
+def test_balance_parser_registry_matches_config_registry():
+    """poller._BALANCE_PARSERS keys equal config.PROVIDERS_WITH_BALANCE_PARSERS."""
+    from config import PROVIDERS_WITH_BALANCE_PARSERS
+
+    assert set(poller._BALANCE_PARSERS) == PROVIDERS_WITH_BALANCE_PARSERS
+
+
+def test_balance_poll_failure_is_isolated(tmp_db, monkeypatch):
+    """A balance-poll exception does not abort the rest of the poll run."""
+    connect, cfg = tmp_db
+
+    def _get(url, timeout, **kwargs):
+        if "/credits" in url:
+            raise requests.ConnectionError("balance endpoint unreachable")
+        if "openrouter.ai" in url:
+            return _fake_response({"data": [_openrouter_model("openai/gpt-6-astra")]})
+        if "neuralwatt" in url:
+            return _fake_response({"data": [_catalog_item("nw-model-a")]})
+        raise RuntimeError(f"unexpected URL: {url}")
+
+    monkeypatch.setattr(poller, "load_config", lambda p: cfg)
+    monkeypatch.setattr(poller.requests, "get", _get)
+    monkeypatch.setenv("OPENROUTER_API_KEY", "sk-or-test")
+
+    import io
+
+    stderr_capture = io.StringIO()
+    monkeypatch.setattr("sys.stderr", stderr_capture)
+
+    exit_code = poller.main()
+    assert exit_code == 0
+    assert "[openrouter] balance poll FAILED" in stderr_capture.getvalue()
+
+    conn = connect()
+    try:
+        # NeuralWatt catalog still upserted
+        assert conn.execute(
+            "SELECT COUNT(*) FROM models WHERE provider='neuralwatt'"
+        ).fetchone()[0] == 1
+        # No balance row because the poll failed
+        assert conn.execute(
+            "SELECT COUNT(*) FROM provider_balance_observations"
+        ).fetchone()[0] == 0
+    finally:
+        conn.close()
+
+
+def test_balance_poll_skipped_when_api_key_missing(tmp_db, monkeypatch):
+    """Without OPENROUTER_API_KEY the balance poll is skipped but the run continues."""
+    connect, cfg = tmp_db
+
+    monkeypatch.setattr(poller, "load_config", lambda p: cfg)
+    monkeypatch.setattr(poller.requests, "get", _url_keyed_get)
+    monkeypatch.delenv("OPENROUTER_API_KEY", raising=False)
+
+    import io
+
+    stderr_capture = io.StringIO()
+    monkeypatch.setattr("sys.stderr", stderr_capture)
+
+    exit_code = poller.main()
+    assert exit_code == 0
+    assert "[openrouter] balance poll skipped: OPENROUTER_API_KEY not set" in stderr_capture.getvalue()
+
+    conn = connect()
+    try:
+        assert conn.execute(
+            "SELECT COUNT(*) FROM provider_balance_observations"
+        ).fetchone()[0] == 0
+    finally:
+        conn.close()
+
+
+def test_balance_parser_drift_caught_not_fatal(tmp_db, monkeypatch):
+    """A malformed credits response is logged and the run continues."""
+    connect, cfg = tmp_db
+
+    def _get(url, timeout, **kwargs):
+        if "/credits" in url:
+            return _fake_response({"data": {}})
+        if "openrouter.ai" in url:
+            return _fake_response({"data": [_openrouter_model("openai/gpt-6-astra")]})
+        if "neuralwatt" in url:
+            return _fake_response({"data": [_catalog_item("nw-model-a")]})
+        raise RuntimeError(f"unexpected URL: {url}")
+
+    monkeypatch.setattr(poller, "load_config", lambda p: cfg)
+    monkeypatch.setattr(poller.requests, "get", _get)
+    monkeypatch.setenv("OPENROUTER_API_KEY", "sk-or-test")
+
+    import io
+
+    stderr_capture = io.StringIO()
+    monkeypatch.setattr("sys.stderr", stderr_capture)
+
+    exit_code = poller.main()
+    assert exit_code == 0
+    assert "[openrouter] balance poll FAILED" in stderr_capture.getvalue()
+
+
+def test_disabled_provider_with_balance_url_gets_no_http_call(tmp_db, monkeypatch):
+    """An enabled=False provider with balance_url configured is skipped entirely."""
+    connect, cfg = tmp_db
+
+    or_cfg = REAL_CFG.dispatch_providers["openrouter"]
+    disabled_or = or_cfg.model_copy(update={"enabled": False})
+    cfg = cfg.model_copy(
+        update={
+            "dispatch_providers": {
+                "neuralwatt": REAL_CFG.dispatch_providers["neuralwatt"],
+                "openrouter": disabled_or,
+            }
+        }
+    )
+
+    monkeypatch.setattr(poller, "load_config", lambda p: cfg)
+    call_log: list[str] = []
+
+    def _get(url, timeout, **kwargs):
+        call_log.append(url)
+        if "neuralwatt" in url:
+            return _fake_response({"data": [_catalog_item("nw-model-a")]})
+        if "openrouter.ai" in url:
+            return _fake_response({"data": [_openrouter_model("openai/gpt-6-astra")]})
+        raise RuntimeError(f"unexpected URL: {url}")
+
+    monkeypatch.setattr(poller.requests, "get", _get)
+
+    exit_code = poller.main()
+    assert exit_code == 0
+    assert not any("/credits" in u for u in call_log)
+    assert not any("openrouter.ai" in u for u in call_log)
--- a/tests/test_provider_balance.py
+++ b/tests/test_provider_balance.py
@@ -0,0 +1,86 @@
+"""Tests for the provider_balance_observations table and poller helpers.
+
+All tests seed a throwaway SQLite DB from schema.sql and assert on queried
+rows, not on stdout strings.
+"""
+
+from __future__ import annotations
+
+import sqlite3
+from datetime import datetime, timezone
+from pathlib import Path
+
+import pytest
+
+import poller
+
+ROOT = Path(__file__).resolve().parent.parent
+SCHEMA_SQL = (ROOT / "config" / "schema.sql").read_text()
+
+
+def _make_db() -> sqlite3.Connection:
+    conn = sqlite3.connect(":memory:")
+    conn.execute("PRAGMA foreign_keys = ON")
+    conn.executescript(SCHEMA_SQL)
+    return conn
+
+
+def _table_columns(conn: sqlite3.Connection, table: str) -> list[str]:
+    return [row[1] for row in conn.execute(f"PRAGMA table_info({table})")]
+
+
+def test_provider_balance_observations_schema() -> None:
+    """QA (a): schema.sql creates the expected columns."""
+    conn = _make_db()
+    assert _table_columns(conn, "provider_balance_observations") == [
+        "id",
+        "provider",
+        "balance_usd",
+        "observed_at",
+    ]
+
+
+def test_record_balance_round_trip() -> None:
+    """QA (b): record_balance inserts UTC ISO8601 with correct values."""
+    conn = _make_db()
+    before = datetime.now(timezone.utc)
+    poller.record_balance(conn, "openrouter", 42.5)
+    after = datetime.now(timezone.utc)
+
+    row = conn.execute(
+        "SELECT provider, balance_usd, observed_at "
+        "FROM provider_balance_observations"
+    ).fetchone()
+    assert row is not None
+    assert row[0] == "openrouter"
+    assert row[1] == pytest.approx(42.5)
+    observed = datetime.fromisoformat(row[2])
+    assert observed.tzinfo is not None
+    assert before <= observed <= after
+
+
+def test_ensure_provider_balance_table_idempotent() -> None:
+    """QA (c): _ensure_provider_balance_table runs twice on an empty DB."""
+    conn = sqlite3.connect(":memory:")
+    conn.execute("PRAGMA foreign_keys = ON")
+    poller._ensure_provider_balance_table(conn)
+    poller._ensure_provider_balance_table(conn)
+    assert _table_columns(conn, "provider_balance_observations") == [
+        "id",
+        "provider",
+        "balance_usd",
+        "observed_at",
+    ]
+    idx = conn.execute(
+        "SELECT name FROM sqlite_master WHERE type='index' AND name='idx_provider_balance'"
+    ).fetchone()
+    assert idx is not None
+
+
+def test_record_balance_fails_without_table() -> None:
+    """QA (d): no hidden auto-create; missing table raises OperationalError."""
+    conn = sqlite3.connect(":memory:")
+    conn.execute("PRAGMA foreign_keys = ON")
+    # Table intentionally not created.
+    with pytest.raises(sqlite3.OperationalError):
+        poller.record_balance(conn, "openrouter", 1.0)
--- a/tests/test_routing.py
+++ b/tests/test_routing.py
@@ -10,6 +10,7 @@ from config import FlexPreference
 from routing import (
    apply_flex_preference,
    capability_gate_reason,
+    credit_attenuation_multiplier,
    get_flex_sibling,
    is_eligible,
    rank_candidates,
@@ -470,6 +471,64 @@ def test_reasons_are_single_tokens():
        assert " " not in reason


+# --- credit attenuation multiplier ----------------------------------------
+
+
+def test_credit_attenuation_unknown_balance_is_healthy():
+    assert credit_attenuation_multiplier(
+        None, soft_floor_usd=5.0, zero_floor_usd=0.0, max_multiplier=5.0
+    ) == 1.0
+
+
+def test_credit_attenuation_above_soft_floor_is_healthy():
+    assert credit_attenuation_multiplier(
+        10.0, soft_floor_usd=5.0, zero_floor_usd=0.0, max_multiplier=5.0
+    ) == 1.0
+
+
+def test_credit_attenuation_at_soft_floor_is_healthy():
+    assert credit_attenuation_multiplier(
+        5.0, soft_floor_usd=5.0, zero_floor_usd=0.0, max_multiplier=5.0
+    ) == 1.0
+
+
+def test_credit_attenuation_at_zero_floor_is_max():
+    assert credit_attenuation_multiplier(
+        0.0, soft_floor_usd=5.0, zero_floor_usd=0.0, max_multiplier=5.0
+    ) == 5.0
+
+
+def test_credit_attenuation_below_zero_floor_is_max():
+    assert credit_attenuation_multiplier(
+        -1.0, soft_floor_usd=5.0, zero_floor_usd=0.0, max_multiplier=5.0
+    ) == 5.0
+
+
+def test_credit_attenuation_midpoint_linear():
+    # balance 2.5 is halfway between soft=5 and zero=0 with max=5.
+    # frac = (5 - 2.5) / (5 - 0) = 0.5
+    # result = 1.0 + 0.5 * (5 - 1.0) = 3.0
+    assert credit_attenuation_multiplier(
+        2.5, soft_floor_usd=5.0, zero_floor_usd=0.0, max_multiplier=5.0
+    ) == pytest.approx(3.0)
+
+
+def test_credit_attenuation_multiplier_range_independence():
+    # Same fractional position with max_multiplier=2.0 should yield 1.5.
+    assert credit_attenuation_multiplier(
+        2.5, soft_floor_usd=5.0, zero_floor_usd=0.0, max_multiplier=2.0
+    ) == pytest.approx(1.5)
+
+
+def test_credit_attenuation_equal_floors_is_safe():
+    # With soft == zero, the guard order (>= soft first, <= zero second)
+    # means the interpolation branch is unreachable regardless of balance.
+    # This construction-safety test hits the <= zero branch.
+    assert credit_attenuation_multiplier(
+        0.0, soft_floor_usd=5.0, zero_floor_usd=5.0, max_multiplier=5.0
+    ) == 5.0
+
+
 # --- capability_gate_reason (the extracted flag rule) -----------------------
 #
 # rejection_reason delegates its vision/json-mode arm to this function, and
@@ -1018,3 +1077,92 @@ def test_through_is_eligible_via_filters():
        task_category="summarization",
    ) is False

+
+# QA cases (a)-(g) for provider_cost_multipliers effective_cost tiebreak.
+
+def _ranked_ids(rows, **kwargs):
+    return [r["model_id"] for r in rank_candidates(rows, **kwargs)]
+
+
+def test_provider_cost_multiplier_flips_tiebreak_toward_healthy_provider():
+    rows = [
+        _row(model_id="A", provider="provA", cost=0.10, proficiency=0.80),
+        _row(model_id="B", provider="provB", cost=0.10, proficiency=0.80),
+    ]
+    assert _ranked_ids(rows) == ["A", "B"]
+    assert _ranked_ids(
+        rows, provider_cost_multipliers={"provA": 5.0}
+    ) == ["B", "A"]
+
+
+def test_provider_cost_multiplier_cannot_override_quality_band_gap():
+    rows = [
+        _row(model_id="A", provider="provA", cost=0.10, proficiency=0.95),
+        _row(model_id="B", provider="provB", cost=0.01, proficiency=0.80),
+    ]
+    assert _ranked_ids(
+        rows, provider_cost_multipliers={"provA": 5.0}, quality_tolerance=0.1
+    ) == ["A", "B"]
+
+
+def test_ranked_rows_carry_raw_cost_and_effective_cost():
+    rows = [
+        _row(model_id="A", provider="provA", cost=0.10, proficiency=0.80),
+    ]
+    ranked = rank_candidates(rows, provider_cost_multipliers={"provA": 5.0})
+    assert len(ranked) == 1
+    assert ranked[0]["cost"] == pytest.approx(0.10)
+    assert ranked[0]["cost_score"] == pytest.approx(1.0)
+    assert ranked[0]["effective_cost"] == pytest.approx(0.50)
+
+
+def test_provider_multiplier_applies_to_fallback_cost_when_estimate_missing():
+    rows = [
+        _row(
+            model_id="A",
+            provider="provA",
+            cost=0.20,
+            proficiency=0.80,
+            prompt_cost_1m=None,
+            completion_cost_1m=None,
+        ),
+    ]
+    ranked = rank_candidates(
+        rows,
+        prompt_tokens=0,
+        completion_tokens=0,
+        provider_cost_multipliers={"provA": 3.0},
+    )
+    assert ranked[0]["cost"] == pytest.approx(0.20)
+    assert ranked[0]["effective_cost"] == pytest.approx(0.60)
+
+
+def test_provider_cost_multiplier_unknown_provider_ignored():
+    rows = [
+        _row(model_id="A", provider="provA", cost=0.10, proficiency=0.80),
+    ]
+    ranked = rank_candidates(
+        rows, provider_cost_multipliers={"provUnknown": 99.0}
+    )
+    assert ranked[0]["effective_cost"] == pytest.approx(0.10)
+
+
+def test_provider_cost_multiplier_defaults_to_one_for_unmapped_provider():
+    rows = [
+        _row(model_id="A", provider="provA", cost=0.10, proficiency=0.80),
+        _row(model_id="B", provider="provB", cost=0.10, proficiency=0.80),
+    ]
+    assert _ranked_ids(
+        rows, provider_cost_multipliers={"provA": 5.0}
+    ) == ["B", "A"]
+
+
+def test_provider_cost_multiplier_empty_dict_matches_no_argument():
+    rows = [
+        _row(model_id="A", provider="provA", cost=0.10, proficiency=0.80),
+        _row(model_id="B", provider="provB", cost=0.10, proficiency=0.80),
+    ]
+    assert _ranked_ids(rows) == _ranked_ids(
+        rows, provider_cost_multipliers={}
+    )
+
--- a/tests/test_tui.py
+++ b/tests/test_tui.py
@@ -35,13 +35,19 @@ def _fixture() -> dict:
            "metered_calls_30d": 18,
            "reset_date": "2026-07-26",
            "note": "router-metered only",
-            "balance_usd": 8.50,
-            "balance_at": "2026-08-23T09:58:00+00:00",
-            "burn_window_hours": 24,
-            "burn_rate_usd_per_hour": 1.0,
-            "projected_hours_remaining": 8.5,
-            "runway_low_warning": False,
-            "runway_note": None,
+            "total_balance_usd": 8.50,
+            "by_provider": {
+                "neuralwatt": {
+                    "balance_usd": 8.50,
+                    "balance_at": "2026-08-23T09:58:00+00:00",
+                    "balance_source": "telemetry",
+                    "burn_window_hours": 24,
+                    "burn_rate_usd_per_hour": 1.0,
+                    "projected_hours_remaining": 8.5,
+                    "runway_low_warning": False,
+                    "runway_note": None,
+                }
+            },
            "window_start_30d": "2026-07-24",
            "metered_kwh_period": 0.5,
        },
@@ -56,13 +62,19 @@ def _fixture() -> dict:
                "metered_calls_30d": 18,
                "reset_date": "2026-07-26",
                "note": "router-metered only",
-                "balance_usd": 8.50,
-                "balance_at": "2026-08-23T09:58:00+00:00",
-                "burn_window_hours": 24,
-                "burn_rate_usd_per_hour": 1.0,
-                "projected_hours_remaining": 8.5,
-                "runway_low_warning": False,
-                "runway_note": None,
+                "total_balance_usd": 8.50,
+                "by_provider": {
+                    "neuralwatt": {
+                        "balance_usd": 8.50,
+                        "balance_at": "2026-08-23T09:58:00+00:00",
+                        "balance_source": "telemetry",
+                        "burn_window_hours": 24,
+                        "burn_rate_usd_per_hour": 1.0,
+                        "projected_hours_remaining": 8.5,
+                        "runway_low_warning": False,
+                        "runway_note": None,
+                    }
+                },
                "window_start_30d": "2026-07-24",
                "metered_kwh_period": 0.5,
            },
@@ -227,21 +239,33 @@ def test_build_model_quota_panel_includes_reset_date():
    assert by_label["window_start_30d"] == "2026-07-24"


-def test_build_model_quota_panel_includes_balance_and_runway_keys():
-    """The new quota payload keys must reach the TUI data model so the
-    quota panel can surface balance, burn rate, and runway alongside the
-    existing plan/metered/reset rows."""
+def test_build_model_quota_panel_includes_total_balance_and_provider_rows():
+    """The new per-provider quota shape must reach the TUI data model."""
    m = build_model(_fixture())
    rows = m["quota"]
    by_label = {r["label"]: r["value"] for r in rows}
-    assert by_label["balance_usd"] == 8.50
-    assert by_label["burn_rate_usd_per_hour"] == 1.0
-    assert by_label["projected_hours_remaining"] == 8.5
-    assert by_label["runway_low_warning"] is False
+    assert by_label["total_balance_usd"] == 8.50
+    assert by_label["neuralwatt balance_usd"] == "8.5 @ 2026-08-23T09:58:00+00:00"
+    assert by_label["neuralwatt balance_source"] == "telemetry"
+    assert by_label["neuralwatt burn_rate_usd_per_hour"] == 1.0
+    assert by_label["neuralwatt projected_hours_remaining"] == 8.5
+    assert by_label["neuralwatt runway_low_warning"] is False
    assert by_label["window_start_30d"] == "2026-07-24"
    assert by_label["reset_date"] == "2026-07-26"


+def test_build_model_quota_panel_drops_flat_balance_keys():
+    """Old flat balance keys must not leak into the TUI row list."""
+    m = build_model(_fixture())
+    labels = {r["label"] for r in m["quota"]}
+    assert "balance_usd" not in labels
+    assert "balance_at" not in labels
+    assert "burn_rate_usd_per_hour" not in labels
+    assert "projected_hours_remaining" not in labels
+    assert "runway_low_warning" not in labels
+    assert "runway_note" not in labels
+
+
 def test_build_model_per_model_lists_seeded_models():
    m = build_model(_fixture())
    rows = m["per_model"]
@@ -1720,6 +1744,115 @@ def test_format_quota_lead_states():
    assert "[rgb(200,80,80)]runway ~1.5h" in low


+def test_quota_lead_values_binds_only_to_polled_providers():
+    rows = {
+        "openrouter balance_usd": "50.0 @ 2026-08-23T10:00:00+00:00",
+        "openrouter balance_source": "polled",
+        "openrouter burn_rate_usd_per_hour": 5.0,
+        "openrouter projected_hours_remaining": 10.0,
+        "openrouter runway_low_warning": False,
+        "openrouter runway_note": None,
+        "neuralwatt balance_usd": "-0.004 @ 2026-08-23T10:00:00+00:00",
+        "neuralwatt balance_source": "telemetry",
+        "neuralwatt burn_rate_usd_per_hour": 1.0,
+        "neuralwatt projected_hours_remaining": -0.004,
+        "neuralwatt runway_low_warning": True,
+        "neuralwatt runway_note": "allowance reads negative",
+    }
+    balance, burn, runway, low, note = tui._quota_lead_values(rows)
+    assert balance == 50.0
+    assert burn == 5.0
+    assert runway == 10.0
+    assert low is False
+    assert note is None
+
+
+def test_quota_lead_values_uses_lowest_runway_among_polled():
+    rows = {
+        "openrouter balance_usd": 50.0,
+        "openrouter balance_source": "polled",
+        "openrouter burn_rate_usd_per_hour": 5.0,
+        "openrouter projected_hours_remaining": 10.0,
+        "openrouter runway_low_warning": False,
+        "openrouter runway_note": None,
+        "provider-b balance_usd": 20.0,
+        "provider-b balance_source": "polled",
+        "provider-b burn_rate_usd_per_hour": 5.0,
+        "provider-b projected_hours_remaining": 3.0,
+        "provider-b runway_low_warning": True,
+        "provider-b runway_note": "low",
+    }
+    balance, burn, runway, low, note = tui._quota_lead_values(rows)
+    assert balance == 20.0
+    assert runway == 3.0
+    assert low is True
+    assert note == "low"
+
+
+def test_quota_lead_values_falls_back_to_lowest_balance_when_no_runway():
+    rows = {
+        "openrouter balance_usd": 50.0,
+        "openrouter balance_source": "polled",
+        "openrouter burn_rate_usd_per_hour": None,
+        "openrouter projected_hours_remaining": None,
+        "openrouter runway_low_warning": False,
+        "openrouter runway_note": "recent top-up",
+        "provider-b balance_usd": 20.0,
+        "provider-b balance_source": "polled",
+        "provider-b burn_rate_usd_per_hour": None,
+        "provider-b projected_hours_remaining": None,
+        "provider-b runway_low_warning": False,
+        "provider-b runway_note": None,
+    }
+    balance, burn, runway, low, note = tui._quota_lead_values(rows)
+    assert balance == 20.0
+    assert burn is None
+    assert runway is None
+
+
+def test_quota_lead_values_returns_none_for_all_telemetry_providers():
+    rows = {
+        "neuralwatt balance_usd": -0.004,
+        "neuralwatt balance_source": "telemetry",
+        "neuralwatt burn_rate_usd_per_hour": 1.0,
+        "neuralwatt projected_hours_remaining": -0.004,
+        "neuralwatt runway_low_warning": True,
+        "neuralwatt runway_note": "negative allowance",
+    }
+    assert tui._quota_lead_values(rows) == (None, None, None, False, None)
+
+
+def test_quota_lead_values_returns_none_for_old_flat_payload():
+    rows = {
+        "balance_usd": 8.5,
+        "burn_rate_usd_per_hour": 1.0,
+        "projected_hours_remaining": 8.5,
+        "runway_low_warning": False,
+        "runway_note": None,
+    }
+    assert tui._quota_lead_values(rows) == (None, None, None, False, None)
+
+
+def test_quota_lead_values_sorted_name_tiebreak():
+    rows = {
+        "openrouter balance_usd": 50.0,
+        "openrouter balance_source": "polled",
+        "openrouter burn_rate_usd_per_hour": 1.0,
+        "openrouter projected_hours_remaining": 5.0,
+        "openrouter runway_low_warning": False,
+        "openrouter runway_note": None,
+        "provider-a balance_usd": 50.0,
+        "provider-a balance_source": "polled",
+        "provider-a burn_rate_usd_per_hour": 1.0,
+        "provider-a projected_hours_remaining": 5.0,
+        "provider-a runway_low_warning": False,
+        "provider-a runway_note": None,
+    }
+    balance, burn, runway, low, note = tui._quota_lead_values(rows)
+    assert balance == 50.0
+    assert runway == 5.0
+
+
 def test_quota_panel_burn_unavailable_renders_note_verbatim():
    """When burn is None the note renders VERBATIM — never blank, 0, or None.

@@ -1730,11 +1863,12 @@ def test_quota_panel_burn_unavailable_renders_note_verbatim():
    data = _fixture()
    note = "burn estimate unavailable: segment after last balance increase spans only 9 minutes"
    for block in (data["quota"], data["coverage"]["quota"]):
-        block["balance_usd"] = 7.5
-        block["burn_rate_usd_per_hour"] = None
-        block["projected_hours_remaining"] = None
-        block["runway_low_warning"] = False
-        block["runway_note"] = note
+        block["by_provider"]["neuralwatt"]["balance_source"] = "polled"
+        block["by_provider"]["neuralwatt"]["balance_usd"] = 7.5
+        block["by_provider"]["neuralwatt"]["burn_rate_usd_per_hour"] = None
+        block["by_provider"]["neuralwatt"]["projected_hours_remaining"] = None
+        block["by_provider"]["neuralwatt"]["runway_low_warning"] = False
+        block["by_provider"]["neuralwatt"]["runway_note"] = note
    stub.payload = data
    app = tui.DashboardApp(fetcher=stub)

@@ -1758,11 +1892,12 @@ def test_quota_panel_note_hidden_when_burn_is_available():
    stub = _StubFetcher()
    data = _fixture()
    for block in (data["quota"], data["coverage"]["quota"]):
-        block["balance_usd"] = 12.0
-        block["burn_rate_usd_per_hour"] = 0.5
-        block["projected_hours_remaining"] = 24.0
-        block["runway_low_warning"] = False
-        block["runway_note"] = None
+        block["by_provider"]["neuralwatt"]["balance_source"] = "polled"
+        block["by_provider"]["neuralwatt"]["balance_usd"] = 12.0
+        block["by_provider"]["neuralwatt"]["burn_rate_usd_per_hour"] = 0.5
+        block["by_provider"]["neuralwatt"]["projected_hours_remaining"] = 24.0
+        block["by_provider"]["neuralwatt"]["runway_low_warning"] = False
+        block["by_provider"]["neuralwatt"]["runway_note"] = None
    stub.payload = data
    app = tui.DashboardApp(fetcher=stub)

--- a/tests/test_tui_warnings.py
+++ b/tests/test_tui_warnings.py
@@ -64,6 +64,9 @@ CFG = SimpleNamespace(
        default_flex_preference=SimpleNamespace(value="auto"),
    ),
    objective=SimpleNamespace(plan_kwh_per_period=6.25, billing_reset_day=None),
+    dispatch_providers={
+        "neuralwatt": SimpleNamespace(has_energy_telemetry=True, balance_url=None)
+    },
    escalation=SimpleNamespace(enabled=True),
    classifier=SimpleNamespace(degraded_warn_min=20, degraded_warn_threshold=0.5),
 )