Flow Audit Loop

3) Run discovery audit

• Use gpt/bundle_ACF-XXX.md as the sole input.
• Focus on: missing fail-closed behavior, missing chokepoints/order, missing ATs, missing RejectReasonCode tokens.
• Keep findings minimal and evidence-based.

4) Patch specs/CONTRACT.md (minimal edits)

• Add explicit fail-closed rules for missing/unparseable inputs.
• Add new AT-### blocks where coverage is missing.
• Update RejectReasonCode registry when introducing new rejection tokens.
• Keep edits localized to the relevant section.

5) Regenerate bundle and run checkers

python3 scripts/extract_contract_excerpts.py \
  --contract specs/CONTRACT.md \
  --flows specs/flows/ARCH_FLOWS.yaml \
  --flow-id ACF-XXX \
  --bundle \
  --out gpt/bundle_ACF-XXX.md \
  --line-numbers

python3 scripts/check_contract_crossrefs.py --contract specs/CONTRACT.md --strict --check-at --include-bare-section-refs
python3 scripts/check_arch_flows.py --contract specs/CONTRACT.md --flows specs/flows/ARCH_FLOWS.yaml --strict

6) Repeat until clean

• If the audit returns findings, apply patches and loop again.
• If no findings: declare “ACF-XXX audit clean.”

AT numbering

• Pick the next unused AT id with:

rg -n "AT-\\d+" specs/CONTRACT.md | tail -n 5

• Use the next available number.

Output format (audit results)

• Scorecard: step closure, ordering/chokepoints, fail-closed, contradictions, AT coverage.
• Findings: severity + evidence + minimal patch text + proposed AT (if needed).
• After fixes: confirm bundle regeneration + checker results.