Recipe, not skill. Cairntir keeps its three-skill core. Finalization Mode combines Quality evidence with a Crucible stress-test and a hard stop budget.
Run Finalization Mode when drafting a roadmap and again when its implementation enters the final phase. The roadmap is not executable until its ending is clear.
Extract and freeze:
COMPLETE, BLOCKED, or EXHAUSTED.Send only the thesis, issue, and test parameters to an impartial tester before coding. The tester authors the acceptance tests and freezes their hashes. The coder may consume tests and failure evidence, but test paths remain outside the coder’s write scope. A changed test hash invalidates the run instead of becoming a repair attempt.
The independent result vocabulary is PASS, FAIL, or INCONCLUSIVE. Honest
red is useful: it identifies a gap the coder can repair. False green is a
contract breach. Test deletion, selective-suite substitution, weakened
assertions, expectation rebaselining, and fixture-specific production logic all
invalidate the run and remain visible in the terminal evidence.
For each round:
Do not repeat an identical check against identical state. Do not invent work after every acceptance item passes. Do not spend the verification reserve on implementation.
COMPLETE: the independent tester returns PASS, every acceptance item has
fresh evidence, and all non-goals held.BLOCKED: name the unmet item, retained evidence, and smallest external
action that can unblock it.EXHAUSTED: name the round or resource limit reached and preserve the last
reproducible state. Further work requires a new roadmap or explicit amendment.Before reporting COMPLETE, use Quality to audit the evidence and Crucible to
ask whether the passing condition represents the requested outcome or merely a
score the agent learned to game.