feat(audit): cost-trace truthfulness audit script (Theme 5 Layer 5)

[mitwilli-create]

Summary

Theme 5 Layer 5 from the 2026-05-29 task-audit. Standalone script that statistically audits the polish-cost-trace JSONL ledgers for plausibility, sum coherence, and outlier detection. Closes the bug-class catalog entry cost-tracker-implausible-entry-poisoning-pre-flight-check by giving Mitchell a recurring audit lens, not just a write-time check.

Smoke validation

$ node scripts/audit-cost-trace.mjs --dry-run
[load] 897 entries from 7 files
[D1] per-day totals: INFO (grand=$125.6)
[D2] cost bands: PASS (outlier_pct=0)
[D3] outliers: PASS (count=0)
[D4] model plausibility: PASS (flagged=0)
[D5] suspicious file: INFO

Healthy state across the board: no >$5 single-call outliers, no model-cost mismatches (Haiku >$0.50 / Sonnet >$2 / Opus >$5), no entries in data/cost-trace-suspicious.jsonl. Total tracked spend across 7 trace files is $125.60.

Dimensions

#	Dimension	Verdict logic
D1	Per-day totals — last 7 days	INFO (informational)
D2	Cost bands — % calls in [0, $0.50] / [$0.50, $2] / [$2, $5] / >$5	PASS if 0 outliers
D3	Outliers (>$5) — top 20 entries flagged for review	PASS if 0 outliers
D4	Model-cost plausibility — Haiku >$0.50, Sonnet >$2, Opus >$5 flagged	PASS if 0 flagged
D5	Suspicious file — count from data/cost-trace-suspicious.jsonl	PASS if empty

Tunable model bands

MODEL_BANDS constant in the script. Adjust if a vendor changes pricing:

const MODEL_BANDS = {
  'haiku': 0.50, 'flash': 0.50, 'sonnet': 2.00, 'opus': 5.00,
  'gpt-5': 3.00, 'gemini': 1.00, 'grok': 1.00, 'sonar': 0.50,
};

Test plan

• node --check passes
• --dry-run produces audit report in .claude/audit/<DATE>/cost-trace-audit-<DATE>.md
• Default mode invokes Sonnet 4.6 synthesis (~$0.05-0.15)
• --council triggers 7-model adjudication
• Output is gitignored

🤖 Generated with Claude Code