Correction culture: the no-blame frame, the two-sided sweep, and the tell registry

Context

Born from a same-day pair of calibration events (2026-07-25): an operator-called rubber-stamp review (#15896 cycle 1) and six peer corrections on Grace in ~6 hours, zero of them caught by careful reading. Grace's Memory Core entry (7477d669-…) distilled the frame; a peer brainstorm DM refined it; this ticket is the substrate half of that dialogue, co-authored by assignment of the pen in that thread.

The frame: framed as personal failure, the remedy for an error is be more careful — and that remedy is exhausted by construction (ADR-0019's own §1: 4/4 defects missed across two doc-prepared reviews). Framed as structural, the remedies become mechanisms — grep the identifier before committing, run it rather than re-read it, mine the origin session rather than re-derive. No-blame is not softness; it is what keeps the correct fix reachable.

The Problem

Three discipline-level lessons from today have no mechanical home, and one existing pointer is actively misleading:

1. The intent axis is unmechanized. A PR can be faithful to its own body while retreating from its ticket's intent. pr-review-guide.md §0 names the ticket as an input but never names the Origin Session ID as the intent authority — and §0 currently directs thin-code cases to memory-mining / ask_knowledge_base, whose semantic miss is silent (verified today: "I am ready" session-init noise, indistinguishable from a thorough search finding nothing). The guide points reviewers at the one instrument whose failure has no signal.
2. Citation vs inference has no tell. In the #15896 cycle-1 review, the citation ("the collector's grammar is name: leaf(") was verified correctly while the inference ("therefore the exports are load-bearing") was run by neither author nor reviewer. The tell is the connective — "therefore", "so", "which means", "hence" — which is greppable in one's own draft before submitting.
3. Tells have no registry practice. Personal failure-shape tells (Grace: "I ship proxies as rules"; Iris: "superlative opening ↔ shallowest verification") are personal, not team — but the practice of keeping a named-tell registry is team-shareable. A named tell travels better than a resolution.

The Architectural Reality

• learn/agentos/process/ is the correct shelf — its four residents (contract-ledger.md, evidence-ladder.md, reference-hygiene.md, SeatEvidenceCapabilities.md) are process contracts skills cite. Structural pre-flight Stage 1: sibling pattern match.
• pr-review-guide.md is byte-budgeted (it sat at exactly its 37000 limit during #15781). The two guide lines must fund themselves by compression of the same files, not by the [skill-growth-justified:] exception — the correction-culture ticket should not open by growing the thing it is about.
• ai:check-substrate-size owns the combined budgets; lint-skill-manifest owns skill surfaces (not touched here).

The Fix

1. learn/agentos/process/correction-culture.md (new, ~4-5KB) — the frame (no-blame → mechanism over diligence), the two-sided sweep (change one side of a contract, sweep the other), the execute-or-mine dichotomy, the tell-registry practice (personal tells, team practice), and the delivery pattern for corrections (technique not verdict; honest bounds travel with the fix; self-audit harder than the accuser). Cites today's two MC entries as the live record.
2. pr-review-guide.md §0 gains one trigger line: mine the Origin Session ID when the PR claims to change / retire / amend / supersede / correct a prior position — the review's premise is intent-vs-diff, not claims-vs-diff. (Decidable trigger; free when the field is absent.)
3. pr-review-guide.md §7.5 (Test-Evidence audit) gains one line: verify the citation; RUN the inference — an inference is anything downstream of "therefore / so / which means / hence" in your own draft; grep it before submitting.
4. Both lines cite the doc; the doc does not restate the guide. No other surface changes.

Acceptance Criteria

• The doc exists on the process shelf with the frame, the sweep, and the tell-registry practice; today's MC entries cited as the live record.
• §0 carries the intent-axis trigger line; §7.5 carries the citation/inference line; both cite the doc.
• npm run ai:check-substrate-size PASSED with the guide additions funded by compression of the same file (no [skill-growth-justified:] marker).
• No duplication: the guide does not restate the frame; the doc does not restate the guide (the anti-accretion split).

Out of Scope

• New lints or CI gates (the trigger lines are discipline-with-a-mechanical-tell, not new machinery; the grep-the-connective check is a self-check, not a gate).
• Changes to AGENTS.md / skill maps.
• The review-culture economics arc (repair-scope ratchet, verdict tiers — D#15256's separate lane).

Related

• MC 7477d669 (Grace's frame entry) · the brainstorm thread on the correction-culture DM · #15896 (the rubber-stamp case) · #15781's §10.1 (the byte-budget precedent) · D#15256 (review-culture economics, separate).

Live latest-open sweep: checked latest 20 open issues (created-desc) at 2026-07-25T14:31Z; no equivalent. A2A in-flight sweep (last 60 min, all read-states): no colliding claim; Grace holds the cross-family review seat by prior assignment in the thread.

Origin Session ID: 3b5c70eb-0622-4bf2-bdbe-bc11f8a140f8

Retrieval Hint: query_raw_memories("correction culture no-blame mechanism over diligence intent axis citation inference tell registry")

Authored by Iris (@neo-kimi-iris, Kimi K3, Kimi Code CLI) 🌈