Peregrine · item shape · everything on file

The Make and Forensics

27 August 2026 · assembled from the corpus, the mechanics catalog, the v1 and v2 specs, the mockups and the two built views

There are more notes than the corpus implies. Both games have a numbered spec section written before either was built, both have mockups, and both have a built view you can read the contract off. The specs and the mechanics catalog disagree about what The Make is, and that disagreement is the thing to settle first.

The Make

Pull a case from an envelope and ID the suspect on as little evidence as you dare. Mechanic on file: minimum-evidence. Order 9, built.

Spec §1.1 — peregrine-spec-v2.md, the fullest source

Reveal: three content checkpoints — opening line, full stem, answer choices — but the reveal itself is a continuous hand-drag. The student pulls the question out of a case envelope and the payout drains with every inch. The call button is always live.
Beat one, the substance rule (binding): the call is the con's behaviour, not his name. The player picks what he is pulling — "treats a correlation as a cause" — phrased in Playbook language. The con card is confirmed after the call, as the earned reveal.
Beat two, the move: after a successful call, first-move selection from Playbook options with trap-procedure distractors. Flat-scored, deliberately not stage-scaled.
Payouts: Stage 1 +150 / Stage 2 +100 / Stage 3 +40, wrong call −100. Marked PLACEHOLDER. Blocking tuning requirement: the schedule must be simulated against a spray-early strategy and shown to be a proper scoring rule.
Ghost: the depth of the player's best clean call on this question, a chalk tick at "34% revealed". Clean calls only.
Gate: at least one typed miss in the active section filter.

The built view — TheMakeView.swift, 637 lines

Reads question.con → con.habit for beat one and con.playbook for the beat-two move options, plus question.correctMove and question.answerOptions.
Pull is a continuous drag, 0…1, monotonic within a question. Payout decays 150 → 40 over a 240pt drag. Answer choices appear at pull ≥ 0.72.
Beat-one options are the correct con's habit plus three other cons' habits, shuffled.
The pulled text falls back through questionText → practiceText → snapStem.

The disagreement

The catalogued mechanic minimum-evidence describes a different game. Its content contract asks for one case with five to eight independent clues, each with a reveal cost, plus candidates, a minimum_sufficient subset and red herrings. The built game reveals one question progressively and asks for a behaviour and a move. Those are not the same unit, and the mechanic's own note that minimum_sufficient "has to be computed against the candidate set rather than asserted" was written against a contract the build never implemented. The catalog also flags minimum-evidence as a candidate merge with catch-the-break, and the built Make is closer to catch-the-break's earliness wager than to a clue set.

Forensics

Four prints, no question shown. Name the job and the plant. Mechanic on file: structure-from-distractors, status seed. Order 10, built.

Spec §1.9 — short, and unusually complete

Flow: four choices displayed as evidence tags, no stem. Call the job (question type, from family options, +40) and the plant (the primary trap distractor, +40, wrong plant −15). The stem then reveals and you solve (+20).
Reading is excluded in v1 — distractors reference an unseen passage, so the deduction collapses. Math and Writing only.
Payouts marked PLACEHOLDER, same proper-scoring-rule check as The Make.

All you've got is the answer sheet — four choices somebody built with intent. Three of them were manufactured to catch a specific kind of mistake. The manufacturing leaves marks.

Read the prints. Call what's being tested and which print is the plant. Then the case file opens — and you walk in already knowing what it wants.
— intro copy, marked CANONICAL in both spec versions

The built view — ForensicsView.swift, 431 lines

Phases: job → plant → solve → verdict → summary.
Reads exactly four things: answerOptions for the evidence cards, correctType for the job, trapAnswer for the plant, and correctAnswer for the solve. The plant is located by string-matching trapAnswer against answerOptions.
The job options come from jobPool, a hardcoded list of eight maths topics inside the view: Linear equations, Systems of equations, Ratios & percents, Quadratics, Geometry, Functions, Exponents, Statistics.
Two mockups exist: forensics.html and forensics-evidence-spread.html.

What a corpus item would have to carry

Field the game reads	The Make	Forensics	Already in the corpus?
Four options, one correct	yes	yes	Yes — universal `options` / `answer`
Which distractor is the primary trap	—	yes, the plant	No — wager items tag traps on several distractors and name no primary
Question type, as a pickable slate	—	yes, the job	Partly — `question_type` exists on 69 items; the slate does not, and the view's vocabulary does not match the corpus enum
A con, for habit and playbook	yes	—	Yes — `con_id` plus the 20-entry con taxonomy with `habit`, `first_move`, `trap_procedures`
Correct first move	yes	—	Yes — the taxonomy's `first_move`, and snap-call's `correct_move`
Progressively revealable body text	yes	the sealed stem	Yes — universal `stem`
A clue set with per-clue reveal costs	only if the catalog wins	—	No — nothing like it exists anywhere

Reading down that column: Forensics needs two fields it does not have, and The Make needs none — unless you decide the catalogued clue-set mechanic is what The Make should become, in which case it needs an item type nothing in the corpus resembles.

The decisions the notes force

Is The Make the built game or the catalogued one? If the built game is right, its mechanic should point at something that describes progressive reveal, and minimum-evidence stays in the catalog as an unbuilt idea. If the catalogued one is right, The Make needs a genuinely new content type and is a bigger build than anything else outstanding.
Does Forensics need its own bank, or does it consume one? It reads the fields an authored Wager or Target Practice item already carries. Declaring consumes: "the-wager" and adding the two missing fields to that bank is far cheaper than a third bank, and the field to declare it exists as of today.
The job vocabulary is a real collision. The view offers "Linear equations, Quadratics, Exponents"; the corpus enum is solve, setup, inference, transition, geometry…; the skill enum is a third vocabulary again. One of the three has to win before a job slate can be authored.
Both payout schedules are PLACEHOLDER with a blocking tuning requirement. Each spec says the schedule must be simulated against a spray-early strategy and shown to be a proper scoring rule. Neither simulation has been run.
Figures change both. Now that figures are in, a Forensics evidence card and a Make envelope may both have to hold one, and that is a decision about the item shape rather than an addition to it.