alee/6krrt
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

docs: refresh CLAUDE.md for PRs #26-#34, and spec the admin portal uplift #35

Merged
alee merged 3 commits from docs/refresh-claude-md into main 2026-09-06 02:32:00 +00:00
Conversation 0 Commits 3 Files Changed 2 +392 -4
alee commented 2026-09-05 23:45:34 +00:00
Owner
Copy Link

Documentation only. No code changes.

1. CLAUDE.md was left behind by nine merges

CLAUDE.md declares itself "the one to trust on what is currently true", and
no commit touched it during #26–#34. Four shipped features were absent
entirely: the classifier cascade, the degraded-outcome attribution rule, the
capability sub-ceiling detector, and the reactive rejection detector.

Adds those, keeping the details that are easy to undo by accident:

  • why the attribution guard exists — a degraded classification lands on
    general_chat, which is a fully scored category and the configured
    fallback_category, so unguarded outcome folding drags real proficiency
    toward outage traffic;
  • the rejection detector groups on the structured task_tier column, not
    a normalized string, because digit-normalization alone merges tier >= 1
    and tier >= 3 and hides which candidate set went empty;
  • the signal is novelty OR rate — routine rejections run ~3/hr while the
    2026-09-04 incident was n=2, below the noise floor, so no single count
    threshold can both catch it and stay quiet.

Also marks the plan queue empty and records the two known follow-ups —
including that the NULL demand crash is latent only: zero such rows exist,
checked on the live DB rather than assumed.

2. plans/admin-portal-uplift.md

From an audit of the portal against the config surface: 5 pages and 9
editable keys against 21 config sections, with seven sections having no
presence at all.

§1 Proficiency is invisible despite being the only category-dependent term
in the ranking — 148 rows over 17 models and 11 categories, surfaced as a
single top-N list. The page must visually distinguish measured from
inherited: 71 of 148 rows are outcome_prior and 39 self_eval_thin, and
treating those as equal to outcome_blended is the mistake this project
already made with -flex rows. Read-only by design — a hand-edited
proficiency score is a fabricated measurement.

Includes an outcomes view, because #34 now excludes some reports via
model_attributable = 0 and nothing shows that.

§2 Profiles is a dead end. All five profiles are built-ins, which render
no action buttons, so nothing is editable and the page never says why.
Duplicate-to-edit fixes it with no backend change.

Its zero-admit alarm also cries wolf: locality reports "admits 0 models"
while an active ollama-local row exists, because _profile_probe passes
task_category=None and that model is gated to file_summarization /
diff_checking. A false alarm on the badge that exists to catch incident #3
is worse than no badge.

§3 Gaming mode, with a real defect attached.
_last_classifier_failure is written and never read — the circuit breaker
does not break the circuit, so nothing stops the router retrying a dead local
classifier, up to 120s per request against a hung or black-holed Ollama.

The toggle is specified as one flag the code reads, not a macro writing
five keys, and must skip the local call rather than fail-and-recover.
Per user decision it requires classifier.cloud_fallback and refuses to
enable without it — a gaming mode that quietly degrades classification is the
thing worth preventing. Cascade steps 1 and 2 stay ahead of the cloud call: a
free, real, same-session classification is not worth paying to re-derive.

Documentation only. No code changes. ## 1. CLAUDE.md was left behind by nine merges CLAUDE.md declares itself *"the one to trust on what is currently true"*, and no commit touched it during #26–#34. Four shipped features were absent entirely: the classifier cascade, the degraded-outcome attribution rule, the capability sub-ceiling detector, and the reactive rejection detector. Adds those, keeping the details that are easy to undo by accident: - **why** the attribution guard exists — a degraded classification lands on `general_chat`, which is a fully scored category *and* the configured `fallback_category`, so unguarded outcome folding drags real proficiency toward outage traffic; - the rejection detector groups on the **structured `task_tier` column**, not a normalized string, because digit-normalization alone merges `tier >= 1` and `tier >= 3` and hides which candidate set went empty; - the signal is **novelty OR rate** — routine rejections run ~3/hr while the 2026-09-04 incident was n=2, *below* the noise floor, so no single count threshold can both catch it and stay quiet. Also marks the plan queue empty and records the two known follow-ups — including that the NULL demand crash is **latent only**: zero such rows exist, checked on the live DB rather than assumed. ## 2. `plans/admin-portal-uplift.md` From an audit of the portal against the config surface: **5 pages and 9 editable keys against 21 config sections**, with seven sections having no presence at all. **§1 Proficiency is invisible** despite being the only category-dependent term in the ranking — 148 rows over 17 models and 11 categories, surfaced as a single top-N list. The page must visually distinguish *measured* from *inherited*: 71 of 148 rows are `outcome_prior` and 39 `self_eval_thin`, and treating those as equal to `outcome_blended` is the mistake this project already made with `-flex` rows. Read-only by design — a hand-edited proficiency score is a fabricated measurement. Includes an outcomes view, because #34 now excludes some reports via `model_attributable = 0` and nothing shows that. **§2 Profiles is a dead end.** All five profiles are built-ins, which render no action buttons, so nothing is editable and the page never says why. Duplicate-to-edit fixes it with no backend change. Its zero-admit alarm also cries wolf: `locality` reports "admits 0 models" while an active `ollama-local` row exists, because `_profile_probe` passes `task_category=None` and that model is gated to `file_summarization` / `diff_checking`. A false alarm on the badge that exists to catch incident #3 is worse than no badge. **§3 Gaming mode, with a real defect attached.** `_last_classifier_failure` is written and never read — the circuit breaker does not break the circuit, so nothing stops the router retrying a dead local classifier, up to 120s per request against a hung or black-holed Ollama. The toggle is specified as **one flag the code reads**, not a macro writing five keys, and must **skip** the local call rather than fail-and-recover. Per user decision it **requires** `classifier.cloud_fallback` and refuses to enable without it — a gaming mode that quietly degrades classification is the thing worth preventing. Cascade steps 1 and 2 stay ahead of the cloud call: a free, real, same-session classification is not worth paying to re-derive.
alee added 3 commits 2026-09-05 23:45:35 +00:00
docs: bring CLAUDE.md up to date with PRs #26-#34 3737bd4898 
CLAUDE.md declares itself the file to trust on what is currently true, and
no commit touched it during the nine merges. Four shipped features were
absent entirely.

Adds: the classifier fallback cascade and, more importantly, the trap it
had to avoid -- a degraded classification lands on general_chat, which is a
fully scored category AND the fallback_category, so unguarded outcome
folding would drag real proficiency toward outage traffic. Records the
attributable/not-attributable split and the fail-open rule.

Adds the capability sub-ceiling and reactive rejection detectors to the
incident-3 section, since that incident recurred on 2026-09-04 through a
dimension the original detector did not model. Keeps the two details that
are easy to undo by accident: grouping on the structured task_tier column
rather than a normalized string, and the novelty-OR-rate signal, which
exists because routine rejections run ~3/hr while the incident was n=2 --
below the noise floor, so no single threshold works.

Notes the two drift tripwires and that the warning fixture is a coupled
system where a new seed can silence an existing class.

Marks the plan queue empty and records the two known follow-ups, including
that the NULL demand crash is latent only -- zero such rows on the live DB,
checked rather than assumed.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VRQXz5SYZYVWscxS1QqF6U
docs(plans): admin portal uplift -- proficiency, profiles, gaming mode d6d3f949ea 
Three gaps found by auditing the portal against the config surface: 5 pages
and 9 editable keys against 21 config sections, with seven sections having
no presence at all.

Proficiency is invisible despite being, per CLAUDE.md, the ONLY
category-dependent term in the ranking -- 148 rows over 17 models and 11
categories, surfaced as a single top-N list. The page must distinguish
measured scores from inherited ones: 71 of the 148 are outcome_prior and 39
self_eval_thin, and treating those as equal to outcome_blended is the
mistake this project already made with -flex rows. Read-only by design; a
hand-edited proficiency score is a fabricated measurement.

The profiles page is a dead end -- all five profiles are built-ins, which
render no action buttons, so nothing on the page is editable and it never
says why. Duplicate-to-edit fixes it with no backend change.

Its zero-admit alarm also cries wolf: locality reports "admits 0 models"
while an active ollama-local row exists, because _profile_probe passes
task_category=None and that model is gated to file_summarization and
diff_checking. A false alarm on the badge that exists to catch incident #3
is worse than no badge.

Gaming mode carries a real defect. _last_classifier_failure is written and
never read, so the circuit breaker does not break the circuit and nothing
stops the router retrying a dead local classifier -- up to 120s per request
against a hung or black-holed Ollama. Records that the toggle should be one
flag the code READS rather than a macro writing five keys, and that it must
skip the local call rather than fail and recover.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VRQXz5SYZYVWscxS1QqF6U
docs(plans): gaming mode requires a cloud classifier, not just a warning 399934fd69 
User decision 2026-09-05. The earlier draft only warned when
classifier.cloud_fallback was absent, which would let gaming mode silently
turn the router into a worse one -- every request dropping to
stale-session, session-history, or a static guess.

Also corrects an easy assumption the draft invited. The cloud classifier
machinery IS built (PR #34) but two things stop it working today: the
config.yaml block is commented out, so cfg.classifier.cloud_fallback is
None and the cascade returns early; and it is step 3 of a cascade, reached
only after local fails AND both free session steps miss. Forcing cloud is a
promotion from fallback to primary, not a wiring job.

Keeps cascade steps 1 and 2 ahead of the cloud call and says why: a stale
session classification is free and was a real classification of that same
session, so paying to re-derive it is spending money for nothing. "Force
cloud" replaces the local model, not free correct answers.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VRQXz5SYZYVWscxS1QqF6U
alee merged commit a5ee46f375 into main 2026-09-06 02:32:00 +00:00
Sign in to join this conversation.
Reviewers
No Reviewers
Labels
No items
No Label
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
alee
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: alee/6krrt#35
Delete Branch "docs/refresh-claude-md"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?