Outcome contract · v1 · the judge and its criteria · for red-pen

The Outcome Contract

The LLM-as-judge system and the criteria it scores against — matched to the problem frame: the bar, the relationship primitive, the kind-shapes, the three endings. The fast loop we climb with, valid only insofar as it predicts your lived eval. Fifteen criteria across card, board, and act; the taste layer is yours to mark.

Draft for your red-pen · criteria extracted from ratified canon · judge fires on ratification

The two loops, and which this is

Your framing, adopted: the four lived numbers (decision-rate, catch-rate, flip-rate, carry-rate) are the EVAL — the slow truth, anchored in your real reactions, and the only thing that can declare DONE. This contract is the JUDGE — the fast loop: explicit criteria matched to the problem frame, scored by LLM judges on every card, every board, every act, every change. The judge is what we climb with; it is valid only insofar as it predicts your eval. Every disagreement between a judge score and your real reaction is a judge bug, fixed by fixture the same day. If the judge says summit and your lived experience disagrees, your experience wins.

Nothing grades itself — the judge least of all. Before every scoring run it must FAIL the known-bad fixtures and PASS the known-goods, or it is fired and the run voids.

CARD criteria — one relationship strand, rendered

Every criterion is binary, cited (the judge quotes the card verbatim), and severity-tagged: FLOOR auto-fails the card; CEILING counts against the climb score.

C1 · IDENTITY [ceiling] — names its subject person-and-practice, first. “Who is this about?” answered in one glance.
C2 · FIRST-REACTION [ceiling — the master criterion] — simulated-you’s first response is a DECISION. A data-answerable question fails; an error catch floor-fails.
C3 · NO FALSE COMPLETE [ceiling] — carries enough to judge, or names exactly what could flip the call — scored against the relationship’s KIND-SHAPE SLOTS (the judge knows this strand is onboarding-blocked, committed-waiting, live-client… and checks every required-but-missing slot was surfaced). A fabricated absence floor-fails — unknown is never dressed as missing.
C4 · MOVE ARRIVES DONE [ceiling] — one of the three honest endings: got it · built the tool and got it · priced it. Analysis-proposing-analysis fails. Silently judging around a hole floor-fails.
C5 · REGISTER [ceiling] — chief-of-staff voice: plain words, a position, zero machinery vocabulary rendered to you, ever.
C6 · GROUNDING [floor] — every claim traceable to a receipt in her world. One fabrication kills the card.
C7 · CURRENCY [floor] — the card reflects the newest evidence in her own world. A card contradicted by her own newer atom — the Brooke class — fails outright.

BOARD criteria — the court as a whole

B1 · SEAT JUSTICE [ceiling] — ordering matches leverage × action-half-life; a live waiting-on-us ask outranks decayed records. Mechanical once the rank law lands; judged until then.
B2 · COVERAGE [ceiling] — the board-level false complete: no known-live relationship missing. The judge receives the evidence census (ask-bearing strands, unrecorded-live parties) and scores the board’s honesty about what it holds and what it doesn’t.
B3 · SELF-CONSISTENCY [floor] — no duplicate subjects, no card contradicting another, honest counts, no mid-word truncation.
B4 · TRAVERSABILITY [ceiling] — the whole board walkable fast: every card scannable at a glance, sections flippable, your five-minutes-for-everything expectation honored.

ACT criteria — a move carried to the edge of consent

A1 · SEND-READY [ceiling] — the draft is in your voice, sendable without rewrite. The judge asks: would Robert edit more than a word or two?
A2 · BLAST RADIUS [ceiling] — who else is affected, who gets kept in the loop — named on the act.
A3 · FOLLOW-THROUGH MINTED [ceiling] — an accepted act creates its own promise-slot: sent → reply owed → N days silent → re-poke.
A4 · CONSENT [floor] — nothing sends without your tap. The judge verifies the act stopped at the edge.

The judge system — small, no platform

  1. One judge call per unit, three forced-apart lenses in one rubric (identity+register · sufficiency+absence · move+carry). Escalates to independent per-lens judges only where disagreement with YOUR reactions exposes a blind spot.
  2. The two-way anchor law before every run — must fail the known-bads (the nameless board, the duplicate board, the Iron Horse false-complete), must pass the known-goods (the ablation’s winners). Fired if it misses either direction; the run voids.
  3. The fixture corpus is the season’s receipts — the nine ablation cards with blind verdicts, both walks, Blake/Christie/Brooke — labeled per-criterion, growing by the calibration law: every real reaction of yours that disagrees with a judge score becomes a fixture the same day.
  4. Judge upgrades are experiments — a new judge version must beat the old on the fixture corpus before adoption. The judge never grades its own upgrade.
  5. Output — a per-unit scorecard (criterion → pass/fail + verbatim citation) and one board-level score per run, appended to the scoreboard: the curve we climb, visible to both of us weekly.

Your red-pen targets

1 · The criteria themselves

What’s missing, what’s wrong, what isn’t YOUR taste? These fifteen are extracted from the ratified canon — the bar, your walk rulings, the relationship primitive — but the taste layer is yours.

2 · Severities

Which CEILINGs are actually FLOORs for you? (Candidates: C1 identity — is a nameless card ever acceptable? A1 send-ready — is a draft needing rewrite a fail or just a miss?)

3 · C2’s anchor lines

The simulated-first-reaction judge will carry your verbatim recorded reactions as its taste anchors. Which of this week’s should found it? (“How many people are there?” · “we can’t cage her” · the five-in-five-minutes expectation…)

4 · Escalation

One judge escalating to three on disagreement — or always three from day one?