<!-- REFERENCE COPY — do not edit this file. Edit the installed working copy at AGENTS.md. -->
<!--
  Anti-ceremony (CHECKLIST.md A12):
  - Consumer: every agent working in this repo, and the reviewer grepping for the patterns (A9).
  - Gate: A9 anti-reward-hacking law — the patterns must appear here verbatim.
  - Defect class: agents under pressure optimizing the metric instead of the work (29/44 suite projects needed this written down).
  - Delete when: the project adopts a machine-enforced equivalent (none exists in this kit; keep this file).
-->

# AGENTS.md — agent operating instructions for this project

Installed by the FrankenSuite Starter Kit. Tailor the project-specific
sections; keep the 12 forbidden patterns verbatim and the three
load-bearing rules intact (CHECKLIST.md A9).

## The one rule

"A step you can satisfy by believing you did it is not a step."

## The 12 forbidden reward-hacking patterns

Quoted verbatim from the suite-wide agent law (AGENTS.md Rule 0.5), as
observed in franken_markdown and frankenfs:

1. gate self-weakening ("temporarily" loosening a check to land a change)
2. proof-class inflation (relabeling weak evidence as strong)
3. golden regeneration reflex (regenerating goldens until they pass)
4. commit-stream pumping (many trivial commits to look productive)
5. tautological tests (tests that assert what the code does, not what it should do)
6. easy-lever cherry-picking (only attempting levers known to pass)
7. close-pump abuse (closing beads to move a metric)
8. scope-splitting (splitting work to dodge gate thresholds)
9. spec-editing as progress (editing the spec instead of the code)
10. conformance metastasis (growing the conformance suite instead of fixing failures)
11. dependency smuggling (undeclared dependencies)
12. bench-path hardcoding (special-casing the benchmark input)

Three load-bearing rules sit above the list:
- Never weaken a gate to land a change (CHECKLIST.md B7).
- No self-grading without independent verification.
- Demotions are always allowed (see docs/evidence/demotion-rules.md, rule D4).

## Project-specific instructions

<!-- Fill in: build commands, test commands, repo layout, conventions. -->

## Honesty machinery (how to use it)

- The beads graph (`.beads/issues.jsonl`) is the executable form of the plan;
  prose documents are the rationale of record. Close beads only with a
  `close_reason` that cites evidence (commit, receipt, ledger row).
- Claims go in `registries/claims.tsv` BEFORE the README makes them; set
  `enforce=yes` once the proof exists. The pre-commit hook cross-checks them.
- Falsified hypotheses go in `docs/evidence/NEGATIVE_EVIDENCE.md` with a
  real, testable retry predicate — never "later". The hook lints every row.
- Definition of Done: `docs/definition-of-done.md`. Done needs command
  evidence; blocked needs artifact evidence (error text + command + path).
