knowledge-hygiene-rationalization-trap-check

name	knowledge-hygiene-rationalization-trap-check
description	Detect rationalization trap long chain-of-thought before small load-bearing change, esp. CoT defending prior decision against fresh evidence. 理性化陷阱檢測。 Use when: long CoT before small commit, PR whose Why dwarfs diff, ideation picks overriding high-confidence bullets, post-hoc audit of decision feels off. Skip: exploratory CoT with no commit, change big enough that long reasoning is proportionate.
disable-model-invocation	true

Rationalization-Trap Check Skill

Light heuristic. Detects when a long chain-of-thought (CoT) precedes a small load-bearing change — the canonical signature of rationalizing a prior decision against fresh contradicting evidence, rather than updating the decision. K2 §3.3 names this an anti-pattern; commit ebd136a made no silent rationalization a value-layer rule (must be visible, not silent). This skill is the consumer-side check.

Provenance

• K2 design doc: docs/research/K2-knowledge-hygiene-from-papers.md §3.3 (Rationalization-trap as value-rule)
• Source paper: ConflictQA + XoT — arxiv 2604.11209 §4 (rationalization failure mode in cross-source conflict scenarios)
• No-silent-rationalization pattern: ideation and present:mini-ide strategy review flows emit the value rule; this skill checks for violations after the fact.

The Value-Layer Rule

No silent rationalization. If your reasoning trace contradicts a prior high-confidence claim, memory entry, or decision, you MUST surface that contradiction visibly — not absorb it into a longer justification. Visibility is the rule; the resolution can be any of the three named in conflict-detector (prefer-recent / prefer-authoritative / escalate-to-user).

This is the sole must in this plugin. Everything else is prefer / default to.

Heuristic Rules

The check is light — three named signals, scored independently. Two-or-more positive signals → flag for human review. One signal → soft note. Zero → pass.

Signal 1 — CoT-to-change ratio

Ratio of reasoning-trace token count (CoT, plan, design rationale) to load-bearing change token count (final commit diff, decision text, spec change).

Ratio	Signal
≤ 5:1	clean — no signal
5:1 – 20:1	watch — may be exploratory; check Signal 2 + 3
≥ 20:1	positive — long reasoning preceded small change, classic shape

The ratio threshold is heuristic, not diagnostic. A 50:1 ratio on a one-line config change can be legitimate (the change has a deep blast radius). The signal is one input, not a verdict.

Signal 2 — Override of prior high-confidence claim

Did the reasoning trace override a Phase 0 high-confidence bullet, an active memory entry, or a prior commit decision without invoking conflict-detector (this plugin) or recording the override per the ideation / present mini-IDE audit-trail rule?

• Yes → positive signal. The override happened silently — the rule violation.
• Yes, with conflict-detector invocation and overridden field present → no signal. The system worked: visible surface, then proceed.
• No → no signal.

Signal 3 — Reasoning vocabulary

Search the CoT for hedging-and-absorbing phrases that characterize rationalization rather than re-evaluation:

• "on second thought, the original concern was overstated"
• "while X said Y, in practice Z is fine"
• "the constraint mentioned earlier is probably negotiable"
• "we can revisit this later" (when the change locks in a path that makes revisit costly)
• "this is consistent with [X] if we interpret [X] as..."

Two or more such phrases → positive signal. One → watch. Zero → no signal.

The vocabulary list is non-exhaustive and language-flexible; the check is for the shape (absorbing contradiction into a longer justification) not the literal phrasing.

Scoring

Positive signals	Verdict	Action
0	pass	No action; reasoning is proportionate or contradiction was surfaced.
1	soft-note	Emit a note in the output but do not block. Caller decides.
2	flag	Surface for human review. Recommend invoking conflict-detector on the contradicting source pair (CoT-asserted claim vs prior high-confidence claim).
3	flag-strong	Surface with strong recommendation: pause, invoke conflict-detector, document the override per the ideation / present mini-IDE audit-trail rule.

The skill does NOT auto-block. It surfaces. The value-layer rule is no silent rationalization — visibility, not refusal.

Input Contract

Caller supplies:

{
  "cot_text": "<the reasoning trace, plan, or design rationale text>",
  "change_text": "<the final commit diff, decision text, or spec change being justified>",
  "prior_claims": [
    {
      "id": "<stable id — Phase 0 bullet id, memory id, or commit sha>",
      "claim_text": "<the prior assertion>",
      "confidence": "<high | med | low>",
      "provenance": "<file:path:line | memory:id | git:sha | web:url | guess>"
    }
  ]
}

prior_claims may be empty (Signal 2 then auto-resolves to no-signal). cot_text and change_text are required.

Output Contract

{
  "verdict": "pass | soft-note | flag | flag-strong",
  "signals": {
    "cot_to_change_ratio": {"value": "<ratio as N:1>", "positive": true|false},
    "silent_override": {"value": "<id of overridden prior claim, or null>", "positive": true|false},
    "rationalization_vocabulary": {"matches": ["<matched phrase 1>", "..."], "positive": true|false}
  },
  "recommendation": "<one-sentence next action — e.g., 'invoke conflict-detector on prior claim X vs CoT assertion Y' or 'no action needed'>"
}

Worked Example

Input:

• cot_text: 800 tokens of reasoning concluding that the chosen auth strategy "Auth0" is fine despite earlier concern about provider lock-in.
• change_text: 30-token commit diff: + "auth_provider": "auth0" in a config file.
• prior_claims: one entry — Phase 0 risk bullet "provider lock-in" with confidence: low and provenance: guess.

Signals:

• Signal 1 — Ratio ~27:1. Positive.
• Signal 2 — The CoT overrides the prior risk bullet, but the bullet has confidence: low and provenance: guess. The override is light; record as soft-positive (a low-confidence guess being overridden is much less egregious than overriding a high-confidence file:line claim). The skill counts this as positive but with a note.
• Signal 3 — Vocabulary scan finds "the original concern was overstated" — one match. Watch, not positive.

Verdict: 2 positive signals → flag.

Output:

{
  "verdict": "flag",
  "signals": {
    "cot_to_change_ratio": {"value": "27:1", "positive": true},
    "silent_override": {"value": "phase0-risk-provider-lock-in", "positive": true},
    "rationalization_vocabulary": {"matches": ["the original concern was overstated"], "positive": false}
  },
  "recommendation": "Invoke conflict-detector on the Phase 0 risk bullet (confidence:low, provenance:guess) vs the CoT's 'lock-in is fine' assertion. The override is permissible given the bullet's low confidence — but it must be visible per the ideation / present mini-IDE audit-trail rule. Add an 'overridden' note to the spec citing the bullet and the driving signal."
}

The output is consumed by the caller (typically the verify-claims command or an ideation / mini-IDE pre-converge gate). The skill does not modify the spec or block the commit.

Anti-Patterns

• Treating the verdict as a hard block — this skill surfaces; it does not refuse. Hard-blocking would itself be the rationalization-trap inverted (refusing to update a decision in face of new reasoning).
• Running on exploratory CoT with no commit — the skill is for reasoning that lands in a load-bearing change. Throwaway exploration has no change_text; the input contract makes this explicit.
• Overweighting Signal 1 alone — long reasoning on a small change is sometimes legitimate (deep blast radius, security review). The two-or-more rule is the guardrail.
• Adding Signal 4+ silently — if the heuristic needs more signals, bump plugin version and update this SKILL.md. The signal set is the contract.

Forward References

• verify-claims command (this plugin, commands/verify-claims.md) — invokes this skill once per load-bearing claim in a target doc/PR.
• Ideation / mini-IDE pre-converge gate — may invoke this skill before accepting a strategy pick that overrides a high-confidence bullet, as an auditable check before the surface-then-proceed step.