/corrections-audit

name: corrections-audit
description: "Use to analyze correction trends, surface recurring patterns, and graduate repeat corrections to guardrails or anti-patterns."
metadata:
  instruction_budget: "40"
  framework_dependency: "mycelium"
  framework_dependency_note: "This skill is designed to run within the Mycelium framework (https://github.com/haabe/mycelium). Standalone use will skip the canvas state, theory gates, and harness behavior the skill assumes. Install: /plugin install mycelium@haabe-mycelium."

Corrections Audit Skill

Analyze corrections.md for trends, recurring patterns, and actionable insights.

Attribution: run the script, do not count by hand

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/check_correction_attribution.py" --root . --snapshot

**Added 2026-08-03.** This audit's headline number — *who caught the mistake* — was computed by hand and had been produced exactly twice, six weeks apart, each time because a human remembered to look. It is the only number that answers the question the correction loop exists to answer: **more harness, or more context.**

The script reports the escape rate **and its denominator, every time**. At the time of writing 14 of 72 entries carry a catcher, so the rate covers 19% of the corpus — quoting it without that is a claim about the whole wearing borrowed clothes, which is the failure this repo spent 2026-08-02/03 removing from its own checks. If it prints `NO RATE AVAILABLE`, that is not 0%: it means nothing in the corpus says who caught anything.

`--snapshot` appends the reading to `.claude/evals/metrics/corrections/<date>.json`, the same layout the other metric adapters use. **The level is not the signal.** Trend `caught_by_hook_or_check` rising in absolute terms; the ratio can also be improved by logging fewer user-caught mistakes, which is why this script never gates on it.

**Attribution is a HARD RULE at write time**, not a suggestion — see `engine/agent-operating-contract.md`. It was advisory until 2026-08-03 and 72 of 100 entries carry no catcher as a result. Do not backfill those in bulk: who caught a mistake six weeks ago is not recoverable by inference, and a guessed catcher corrupts the only number this loop produces.

**For entries written before the rule, add one phrase per new entry** — `caught by user`, `caught by hook`, `caught by review`, `self-caught`. The parser reads the prose forms already present rather than requiring a new field, so nothing existing needs rewriting.

When to Use

• Loop 2 (Incremental) cadence: after every 3+ corrections are logged
• When the same correction category appears 3+ times
• During `/mycelium:diamond-assess` if corrections gate has findings
• Before starting a new diamond at the same scale as a previously corrected one

Workflow

1. **Load corrections AND warnings AND clusters**: Read `.claude/memory/corrections.md`, `.claude/memory/warnings-log.md`, AND `.claude/memory/cluster-instances.md` (the cluster log graduated 2026-05-08 — canonical record of recurring-pattern instances and their graduation status; without it, "the cluster has graduated N times" has no auditable backing).

• If corrections + warnings both empty AND no clusters logged: report "No corrections, warnings, or clusters to audit" and stop
• Treat all three as inputs to the same pattern analysis. Corrections capture agent-introduced failures; warnings capture framework-state debt; cluster-instances capture cross-cluster pattern accumulation with explicit graduation criteria. Same recurring-pattern shape, different vantage points.

2. **Categorize by frequency**:

• Group corrections by `Category` (bias, security, engineering, process, communication)
• Group by `Scope` (discovery, delivery, orchestration, quality)
• Count occurrences per group

3. **Detect recurring patterns**:

• [ ] Same category appears 3+ times -> candidate for guardrail graduation
• [ ] Same scope appears 3+ times -> candidate for domain-level CLAUDE.md update
• [ ] Same mistake repeats after prevention was documented -> prevention strategy failed, needs escalation

4. **Check origin distribution** (APEX alignment):

• Count corrections by `Origin` (ai-generated, human-written, ai-assisted)
• If ai-generated corrections dominate (>60%): flag for prompt/context improvement, BUT see `detection_origin` cross-check below before acting on this interpretation
• If human-written corrections dominate (>60%): flag for process/training improvement
• If ai-assisted is high: check if the AI contribution or the human contribution caused the issue

4b. **Cross-check with detection_origin** (when field is present — see .claude/memory/README.md):

• Count corrections by `Detection_origin` if present (user / agent_self / hook / evaluator / eval_runner / external_review)
• **Critical disambiguation**: if Origin is heavily ai-generated AND detection_origin is heavily `user`, the apparent AI-quality signal is actually a HARNESS-DETECTION GAP. The AI is generating failures and the user is the only entity catching them. The right intervention is more harness checks (hooks, evaluators), NOT more AI context.
• If detection_origin is dominantly `user` (>70%): flag for harness-detection gap. Suggest where new hooks or evaluators could catch the failure modes earlier.
• If detection_origin is well-distributed across mechanisms: harness coverage is healthy; trust the Origin signal at face value.
• Surfaced 2026-05-03 (mycelium-roadmap dogfood): without this cross-check, the audit's "100% ai-generated → improve prompt context" framing would have driven the wrong intervention. Real signal was "AI generates, user catches" — fixed by shipping the framework-guard hook (harness-detection layer), not by improving prompts.

5. **Root-cause recurring corrections** (5 Whys): For each correction that appears 3+ times, apply 5 Whys to find the systemic root:

• Why did this happen? -> Why did that happen? -> .