Where to resume when the games are framed and you are ready to commission questions. The evaluation loop is built and measured; the bank is ahead of the design. This plan carries what is true today, what has to be decided before anyone authors at volume, and the order the work wants to run in.
The tools are done and proven.
| Tool | What it does | State |
|---|---|---|
tools/evaluate-peregrine.py | Jev pass over an item, writes the verdict onto the entry | 431 items evaluated, 23 seconds, $0.04 |
tools/revise-peregrine.py | Bounded revision loop, dry run by default | Ran on three items, halted on regression |
tools/validate-peregrine.py | Corpus invariants, and what still needs a pass | Reports 431 of 431 carry a verdict |
The authority rule governs all of it. Jev flags and never decides. A flag goes to a Claude review blind to Jev; agreement makes the corpus value definitive, and disagreement waits in evaluation/reconciliation-queue.json for a cross-model tool.
The bank, by game.
| Game | Tier | Items | Target |
|---|---|---|---|
| The Wager | locked | 256 | 45 |
| The Lineup | locked | 175 | 50 |
| Target Practice | locked | 45 | 45 |
| Prime Suspect | locked | 9 | 8 |
| The Dossier | locked | 0 | — |
| Before She Comes Back | locked | 0 | — |
| Go Undercover | refining | 24 | 25 |
| The Lobby · The Statement | refining | 2 each | 10 each |
| Red String | refining | 2 | 2 |
| Beat the Shakedown · Fault in the Record · The Make | refining | 0 | — |
| The Frisk | unlisted | 0 | — |
Two locked games carry no content at all. The two banks that are full are four and five times over their targets.
refining games: which survive, which are cut. Nobody should author against a game that may not ship, which is the reason content paused.
bank_target for every surviving game. Nine of the twenty carry none, so "how much content" has no answer for them.
evaluation/mechanisms-batch-*.json onto their entries. They were written and scored overnight on 21 September and live beside the corpus rather than on it; the guide's content bar wants a wrong-reasoning line on every distractor.
177 items have never been evaluated, because each family needs its own state builder.
bundled-drills (93 items) — carries answer_options, not options
target-practice (45) — claims rather than a stem and four choices
go-undercover (24) — a con's spoken line and a counter
prime-suspect (9) — a case with evidence threads, already validated by its own checker
red-string, the-statement, the-lobby (2 each) — chain lines, which also crash the validator's answer-key check today
--run --send --write as it lands, so the bank never again gets ahead of its measurement.
All fourteen testable dimensions have a negative control and every one discriminates, so a flat reading is evidence rather than an unknown.
distractor_pull, one_defensible_answer, sat_stem_conventions, sat_choice_conventions. Each is measured on 431 items today.
difficulty advisory. It agrees with the corpus on 71% of the bank against a 75% human ceiling, so a threshold above the ceiling means nothing.
export-machine.py should gate on a verdict once thresholds exist, so a reconcile or escalate item never reaches the game. Recorded as D2 in the Jev plan.
skill_code disagreements, waiting on the cross-model tool.
season-documents entries fail validation because their family is undeclared in guide.json, which is 27 of the 27 errors a run reports.
w-writing-023 and the understudy pair were repaired; the same defect may sit in other items whose stem does not describe what they test.
docs/reviews/ and confirm the before-and-after is legible
What this solves. It resumes content work in the order the dependencies actually run: decisions first, then landing work already done, then measurement coverage, then volume. It keeps the rule that nothing is authored against a game that may be cut.
Pros. The expensive tooling is finished and proven, so content work inherits a measured bank from its first item. Two locked and empty games mean volume can start the moment their shapes are specced, without waiting on the twelve refining decisions. Section A is mechanical and recovers work already paid for.
Cons and what is uncertain. Section B is the largest unknown: five state builders against five entry shapes, and the Prime Suspect and Red String shapes may not suit a blind pass at all. The Dossier has no content spec, so "author The Dossier" is a design task wearing an authoring task's clothes. Whether the 479 Lineup distractors need mechanisms is a genuine question about that game's design rather than a gap to fill.
Where this plan is inadequate. It assumes the roster decisions land before authoring. If they do not, Section A and Section B still run, and Section C stalls. It also says nothing about the UI and mechanics work that decides which games survive; that work is the real gate and lives outside this plan.
Confidence: 78. High on sections A, B and D, which are mechanical or already measured. Lower on C, since two of its four items are design work whose shape nobody has settled.