Docs: sweep retired reaction-classifier narration from the samskara docs#382
Merged
Conversation
The evaluation-judge redesign retired the live 1-10 minute reaction classifier, the wall-clock decay sweep, and the sqrt-N cohort reinforcement, but samskara.md still narrated all three in present tense in roughly a dozen places - the prompt inventory listed a REACTION_PROMPT that no longer exists, the data model omitted the verdict column and described was_confirmed as classifier-written, the driver-split rationale explained a tail phase that is not a phase anymore, and two gotchas (sqrt-N reinforcement, the 15-minute neutral backdate) described mechanisms that are gone. Rewritten against current reality: verdict is the authoritative fire judgment, was_confirmed is the legacy shadow, evaluation rides its own day-gated cron, and the Observability calibration note now explains pending fires in judge terms. Also: append the 2026-07-03 hosted-judge executions to the samskara-decay QA results log (the pre-fix truncation fail and the post-budget-fix pass), and capture the "max_completion_tokens pays for the reasoning pass" constraint as a CLAUDE.md convention - it has now bitten two features (second-thoughts reviewer, samskara judge) and is repo-wide scope.
The samskara tab's Activity note and the per-message cohort pill
description both still explained resolution in terms of the retired
live reaction classifier ("only the turn right after an instinct
fires can confirm it"). Reworded to the next-day judge: fires resolve
in a batch after the conversation settles, recent fires are always
pending, and one-question conversations are skipped by design.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
SYNOPSIS
Docs-only hygiene: bring the samskara docs (dev + user) in line with the evaluation-judge era, log today's hosted QA runs, and capture the reasoning-token-budget convention in CLAUDE.md.
PURPOSE
The evaluation-judge redesign retired the live reaction classifier, the wall-clock decay sweep, and sqrt-N cohort reinforcement - but
docs/dev/samskara.mdstill narrated all three in present tense in ~a dozen places, and two user-manual panels explained resolution via the retired classifier. A stale dev doc misleads the next session with full confidence.DESCRIPTION
docs/dev/samskara.md: prompt inventory drops the nonexistentREACTION_PROMPT(five prompts, not six); data model gains theverdictcolumn and describeswas_confirmedas the judge-written legacy shadow; driver-split rationale now says evaluation rides its own day-gated cron; deleted the retired sqrt-N reinforcement gotcha; rewrote the "neutral has no boolean state" gotcha (the 15-minute backdate is gone - filter onverdict); Observability calibration note explains pending fires in judge terms (~88% not-engaged is recall working, not a defect)docs/qa/use-cases/samskara-decay.md: append-only results rows for today's two hosted-judge executions (truncation fail @ a1c3424, post-budget-fix pass @ 09a25f3)CLAUDE.md: new "Venice sub-completions on reasoning models" convention -max_completion_tokenspays for the thinking pass, which scales with input size; has bitten second-thoughts (Fix the second-thoughts reviewer: non-reasoning model + tolerant parse #370) and the samskara judge (Samskara judge: budget the reasoning pass, pin low effort #378)docs/user/samskara.md+docs/user/chat.md: Activity note and cohort-pill description reworded to the next-day-judge modelNotes:
samskara_apply_evaluation- intentional removal, not lost documentation (git preserves it).Generated by Claude Code