Peregrine · Plan for review

The Thread as Container and the Prime Suspect Inference Pass: Follow-Through

This plan finishes the work from the 20 September design night: The Thread as a container, the verb-registry amendment, and the four ecosystems, together with the rest of the Prime Suspect inference pass. It continues prime-suspect-inference-pass.md.

24 September 2026 · Source: .claude/plans/thread-container-followthrough.md · CR-NOTE B7EE1005 · Baseline checked against git, roster-status.json, guide.json, verb-registry.md and the exported cases

76 / 100

Confidence (74 at writing)

Entry steps already done

Recommendations (R1–R10)

Clues whose correct choice names suspects

1Where things stand

Three of the recovery entry's next steps were already done. Two others turned out larger than the entry described. This table sets the baseline so no phase redoes finished work.

Item from the entry	State	Evidence
Re-render the container spec review page	Done	`8bb4e79`; the page carries §9 to §11
Status for The Rookie, Undercover Gauntlet, The Shredder	Done	Shelved 2026-09-21, review cards 27 to 29
Decouple the charge gate; flat quarter-clock chain bonus	Shipped	`e0e98f0`, `ea3ecdb`, `7c619b1`
Hide category until pulled; ask for the chain at the charge	Shipped	`c805cd0`
Cases 003 and 006 turn on a crossing	Shipped	Blind solvers found both crossings
Retire `reading-game`	Decided	The Dossier is the reading game. The slot is still in `guide.json` and has to be removed in-app
Verb-registry §7.6 amendment	Not done	Lives only in `the-thread.md §2`; registry line 62 still reserves `locate` for The Frisk
The Frisk in or out	Open	`unlisted`; Sarah on 2026-09-21: "Unclear"
Names in the answer choices of E01 to E03	Larger	23 clues across all 8 cases have a correct choice naming suspects (27 counting wrong choices)
No-scoring unit test for the container	Not yet possible	No container type exists; scoring writes are inline across the app

Two stale records contradict the settled calls:

ecosystems.json › consequences still reads "Client Cases ship first inside the phone container". That is the deprecated §11 draft wording, which treats a content class as if it were a surface.
The the-thread row in guide.json still carries the 5 September "subpar and conflicting" flag and lists locate, place and connect gestures.

2Dependencies

Phase 0 · Records→No blockers. Can start today

Phase 2 · Registry→Unblocked today if R1 is adopted; otherwise waits on the Frisk call

Phase 3 · Guard→Lands in the same commit as the first shell code, never before

Phase 4a · Choice leak→Scope call, then a two-clue pilot (R3)

Phase 4b / 4c · Rewires→Teen feel pass on 003 and 006 (V1)

Phase 5 · Phone case→Cast, Voicemail and Red String calls; the prototype also waits on V2

Case content changes go to the corpus (guides/peregrine/entries/ps-case*.json) and are re-exported with PressRaven/tools/export-machine.py --write. The exported case files are exporter-owned and are never hand-edited.

3Recommendations summary

Each recommendation is repeated with its reasoning inside the phase it changes. The phase sections below mark them in red callouts.

R1Phase 2

Reserve notice now and leave The Frisk's roster status open.

The registry fix no longer waits on a roster decision.

R2Phase 4a

Scope the leak fix to the 23 correct choices, then make the warning an error.

Fixes what reveals survivors and keeps cases 101 and 102 clean.

R3Phase 4a

Pilot the rewrite on two clues before touching 23.

Tests whether the rewrite changes the SAT skill before the cost is spent.

R4Phase 4b

Check the coin-flip pricing per real case, not only for one parametric shape.

Every rewire gets an automatic pricing check.

R5Phase 2

Add a script that fails on duplicate verb claims in the registry.

§7.6 gains enforcement the way code rules have it.

R6Phase 3

Make a source scan the primary container guard, with a pure navigator type in Core.

Matches how scores are actually written today.

R7Phases 1, 4c

Treat case 101 as a Client case hosting one season trace.

Follows the 13 September rule that the arc rides in on an item.

R8Phase 0

Batch every guide.json edit into one in-app Press Raven session.

The app can't silently restore a hand edit.

R9Phase 5

Set the phone case's size target before authoring it.

The recorded cost becomes comparable to the thirty-case contract.

R10Process

File Peregrine design sessions from the Peregrine checkout.

Prevents a repeat of the 20 September record gap.

P0Phase 0 · Reconcile the records

Success: every record matches the settled calls, and nothing in the corpus still says the container hosts an ecosystem.

Mark the completed items in prime-suspect-inference-pass.md as done and point its step 5 to this plan
Rewrite the Client Cases line in ecosystems.json › consequences to state the orthogonality rule: the phone container's first tenant, a presentation, with Prime Suspect as the other presentation of Client material
Record in the-thread.md §9 that the Rookie, Gauntlet and Shredder shelving and the Dossier-as-reading-game call were made on 2026-09-21
Re-render the container spec page only if the-thread.md changed in this phase
Prepare the replacement text for the the-thread row (container, owns no verb, may not score) for Sarah to paste in-app R8

R8 · Batch the guide.json edits

Two edits touch guide.json: removing reading-game and rewriting the the-thread row. Press Raven rewrites the file from memory, so a hand edit made while it is running gets restored silently. Do both edits in one in-app session, then confirm with git diff in the corpus that both survived.

P1Phase 1 · Decisions that block later phases (Sarah)

Claude does not make any of these calls. Where a recommendation exists, its reason is stated alongside it.

Decision	Blocks	Recommendation
The Frisk: in or out	Nothing, if R1 is adopted	Keep it open until the V2 mockup pass. The registry fix lands either way (R1)
First thread's cast	Phase 5	None. This is a story call
Voicemail transcript: authored or derived	Phase 5 schema	Leaning authored for the first case. Deriving it needs a punctuation-stripping pipeline that nothing else uses, and one case does not justify building one
Red String: Client Cases or Season Plot	Client contract costing	None. It depends on whether the bank is replayable, and that has not been measured
Case 101 (Halloran's Books): which ecosystem	Phase 4c	A Client case hosting one season trace (R7). The 13 September architecture has the arc ride in on "a fifth [clue] that's a memo that shouldn't exist". A Prime Suspect case with Halloran in one clue follows that pattern. The case follows the Client pair rule, and the one Halloran clue also follows the trace locks
Answer-choice leak: scope	Phase 4a	The 23 correct choices (R2). A wrong choice that names suspects reveals nothing the finding won't name once the clue is read
Has a teen feel pass on 003/006 happened?	Phase 4b	The inference plan records blind solvers but no teen pass. The answer decides whether 4b can start

Make each call in the table above
Remove the reading-game slot from inside Press Raven, together with the R8 batch

P2Phase 2 · Amend the verb registry

Depends on: nothing, if R1 is adopted. Success: verb-registry.md states the station/container rule itself, and no verb is claimed twice.

R1 · Decouple the registry from the roster call

The recovery entry treats the notice/locate split as conditional on The Frisk shipping. The split describes two different acts: salience with no criterion, and search against a stated claim. It holds whether or not either game is ever built. Recording notice as a reserved verb costs one table row and ends the three-way deadlock now. If The Frisk is cut later, the row is struck like any other reservation.

Add the station/container amendment to verb-registry.md §2, including the rules that a container may not score and that a scoring container counts as a station
Rewrite the Thread row (line 73) to say it claims no verb and accepts navigation only. Drop the gesture list, which reads as three verb claims
Add a notice row reserved for The Frisk, strike locate (The Frisk) on line 62, and clear The Lobby row's flag R1
If The Frisk is later cut, strike the notice row and move The Frisk to shelved or cut in roster-status.json
Update the-thread.md §2.3 with the outcome and re-render the review page
Write tools/check-verb-registry.py: parse the verb tables, fail on duplicate claims within a section, and ignore struck rows R5
Confirm the script passes on the amended registry and fails on a copy with a planted duplicate R5

R5 · Enforce the registry

§7.6 is the registry's binding rule, and today the only check is someone reading the tables. The tables are regular enough to parse. A duplicate-claim check would have caught the three-way locate claim when it was introduced.

P3Phase 3 · The container may not score, enforced

Depends on: the first shell code. This phase lands in the same commit as the first container type. A test written before a container exists passes without checking anything.

What the 24 September check found Scoring is written inline, with no seam to test through. NightTake is inserted in NightDebrief.swift:240 and :291 and in CaseSession.swift:187. DrillResult is constructed inside at least four game views (ForensicsView, TheMakeView, ShakedownGameView, KingpinShowdownView). PersonalBests computes from stored drills and writes nothing, so the first draft's plan to spy on it was wrong.

R6 · Make the source scan the primary guard

A behavioral test with spy sinks needs a scoring seam the app does not have. Building one only to test this rule is a larger refactor than the rule warrants. A scan of the shell's source for the symbols that write scores fails on exactly the "reasonable-looking later change" that the-thread.md §2.2 warns about. Pairing it with a pure navigator type in Core keeps the shell's logic testable without a view.

Put the shell's navigation logic in a pure Core type whose only output is a closed route type, such as a StationRoute enum. The SwiftUI shell renders routes and opens stations, and it holds no game state
Write a Core test that walks every navigation action and asserts each result is a StationRoute. The route type cannot carry points, so the compiler enforces the rest
Write a source-level test that scans the shell's app directory and the navigator file, with comments stripped, and fails on NightTake(, DrillResult(, context.insert or PrimeSuspect.Score R6
Reference both tests from verb-registry.md next to the amendment
Do not split the shell into its own SwiftPM target yet. The scoring types are public, so a separate target would not block access by itself

P4Phase 4 · Prime Suspect content

4a · Stop answer choices from naming survivors

The pattern is the same in every flagged clue. The question asks which suspects fit, so the correct choice is the survivor list, and the clue does the elimination for the player. For example, case 004 E01 asks "which suspects' vehicle-speed estimates fit the collar data?" and the correct answer is "Jon, Ryan, and Amir".

Case	Clues with a correct choice naming suspects	Count
001 Calder Cipher	E02, E03, E04	3
002 Ivory Gate	E02, E03, E04	3
003 Glass That Never Broke	E01	1
004 Vanishing of Scout	E01, E02, E03, E07	4
005 Last Toast	E02, E03	2
006 Dead Circuit	E01, E02, E03, E10	4
007 11:47 Message	E01, E02, E03	3
008 Four Seconds	E01, E02, E03	3

R3 · Pilot before rewriting 23 clues

The proposed rewrite asks for the quantity (a speed, a time, a count) and moves the names into the finding. That can change which SAT skill the item tests. "Which rows satisfy a condition" is data reading, while "what is the average speed" is computation. Pilot it on two clues, one math and one reading-flavored, and check each against its knowledge_grain. A second pattern may fit the data-reading items better: the choices name rows or exhibits, and never people.

R2 · Scope, then lock

Fix only the 23 correct choices. Once the validator reports zero, promote the warning to an error so cases 101 and 102 can't reintroduce the leak.

Add a CaseValidator warning for a correct choice that names a suspect, using the existing word-boundary matcher
Pilot the rewrite on case 004 E01 and one reading-flavored clue, check both against knowledge_grain, and record which pattern fits which item type R3
Rewrite the remaining flagged clues in the corpus with the chosen pattern, without touching clears. Case 001's print-margin clue has a pinned lint test
Re-export, then run PrimeSuspectCaseLintTests and PrimeSuspectContentTests
Promote the warning to an error once it reports zero R2

4b · Rewire the remaining cases to the pair rule

Depends on: V1. The inference plan puts this rewire after the teen feel pass, and the recovery entry dropped that gate. This plan keeps the gate, because five rewires are the expensive half and the feel pass is what justifies them.

R4 · Check the pricing per case

testProvingItBeatsTheCoinFlipFromEveryOrdering checks one parametric shape: a 120-minute clock, 20-minute pulls, and a proof costing one or two extra pulls. The rewired cases differ in clock, clue count and pair position. The claim that careful play beats a coin flip is therefore verified for the shape and not for each case. A test that iterates every exported case, computes its minimum covering set with and without the pair, and applies the same inequality would turn each rewire into an automatic check.

Write the per-case pricing test before the first rewire, and confirm it passes for 003 and 006 R4
005 Last Toast: promote E09+E10 so Gavin is pair-only; edit E03; set clues_to_charge so the gate stays at 3
008 Four Seconds: same shape. Oscar becomes pair-only through E09+E10, with E03 edited
004 Vanishing of Scout: new E01+E04 pair for Jon; edit E02 and E04. Fold in E01's 4a fix
007 11:47 Message: decide whether Vivian's triple can drop to a pair before editing
002 Ivory Gate: last and most expensive; both halves are independently conclusive. Fold in its 4a fix
For each case, edit the prose that asserts the moved elimination (arithmetic, logs, finding, closing), then re-export, lint, and run the pricing test
For each case, run a blind solve and confirm the crossing is found unaided

4c · Cases 9 and 10

Finish ps-case101-hallorans-books (still draft) to the pair rule, under the ecosystem Phase 1 assigns R7
Author case 102 to the pair rule, with choices that pass the 4a check. Set clues_to_charge only if the default of 3 is wrong for its graph
Resolve the con-alias collisions (Idris, Bell, Voss, Graves, Finch, and the reused Luca) before export
Export, lint, run the pricing test, blind-solve, and add both cases to the sequential-unlock list

P5Phase 5 · One Client case for a phone

Depends on: the cast, Voicemail and Red String calls. Success: one complete phone case exists as data, with its cost recorded, before any shell prototype is built.

A phone case has a different shape from a board case. It is told in the first person from one person's messages and logs, and it asks what that person was doing. It cannot be a port of ps-case002 (the-thread.md §11.2).

R9 · Size before authoring

"Most expensive by unit" is a measured claim only if the unit is fixed before the case is written. Setting the target first also gives the shell its first answer to the 13 September rubric question on unit size and word ceiling.

Set the size target (item count, word ceiling per item, number of distinct stations) in the-thread.md §6 R9
Draft the content model in the-thread.md §6: message, voice memo with transcript, photo, and log, with the station each routes to
Author the case in the corpus under guides/peregrine/design/ until an exporter target exists
Record the cost (hours, words, items) against the size target and the thirty-case Client contract in ecosystems.json
Hold the shell prototype until V2 is done

P6Process

R10 · Keep the record in the right project

Nothing from 20 September reached Peregrine's Chronicle, because the design session ran from the PressRaven checkout. When a Peregrine design session starts from PressRaven, run /cr-entry from the Peregrine checkout at the end.

Add one line to Peregrine's CLAUDE.md saying Peregrine design sessions file their entry from the Peregrine checkout R10

VVerification (Sarah)

V1. Teen feel pass on cases 003 and 006: does crossing two findings feel like deduction or like a longer subtraction, and does the ending still read as a tally? This gates 4b
V2. Teen feel pass on the existing Thread and Frisk mockups before any content is commissioned. This gates the Phase 5 prototype and informs the Frisk call
Review the registry amendment and the notice row
Open the re-rendered container spec page and confirm §9 records the 2026-09-21 roster calls
After the 4a pilot, play both pilot clues on device and confirm the choices no longer reveal who survives
After 4b, play one rewired case start to finish and confirm the Case Closed line names the pair
After the R8 batch, run git diff in the corpus and confirm both guide.json edits survived

CConfidence and how to raise it

74→76

The score is my confidence that this plan is the best route to finishing the continuation. It was 74 when first written. The 24 September checks corrected Phase 3's design, and that moves it to 76.

Why 74

The score is a weighted view of the plan's parts. Each row's bar shows the score at writing. A green bar shows where the checks changed it.

Part	Weight	Score	Reason
Baseline	20%	92	Every claim was checked against git, `roster-status.json`, `guide.json` and the registry. The margin covers records outside those sources
Sequencing	20%	85	Follows the inference plan's own gate. Unknown: whether a teen pass already happened
Phase 2 · Registry	10%	85	Mechanical once R1 is accepted
Phase 3 · Guard	15%	60 → 75	The sink list was assumed, and the check found it partly wrong, which led to R6. Remaining gap: no shell code exists to test the design against
Phase 4a · Leak	15%	55	The rewrite pattern is my proposal. It hasn't been tried on a single clue, and it may change what an item tests
Phase 4b/4c · Rewires	10%	75	The method is proven on 003 and 006. Case 007's triple and case 002's double-conclusive halves are unresolved
Phase 5 · Phone case	10%	55	No content model yet; size and cost unknown

What would raise it

Action	Who	Moves	Effect
Pilot the 4a rewrite on two clues and check `knowledge_grain`	Claude	4a: 55 → 75	+3
Say whether a teen feel pass on 003/006 has happened	Sarah	Sequencing: 85 → 92	+1.5
Draft the phone-case content model and size target (R9)	Claude, after the cast call	Phase 5: 55 → 70	+1.5
Write the R4 per-case pricing test	Claude	4b/4c: 75 → 85	+1
Answer the Phase 1 calls	Sarah	Removes branches in Phases 2 and 5	+1
Settle 007's triple and 002's pair shape on paper	Claude	4b/4c	+0.5

Ceiling: about 85 from planning alone. The largest open question is whether crossing two findings feels like deduction to a sixteen-year-old, and only V1 can answer it. If V1 comes back negative, stop 4b and switch to the reserved "offer one of two each pull" lever from the inference plan.

RFinal review

Pros

The plan starts from the repo's actual state. It drops three finished items and flags two stale corpus lines. It keeps the inference plan's gate on the expensive rewire, and it designs the container guard around how scores are actually written.

Cons

Most phases wait on decisions. With R1 adopted, Phase 0, Phase 2, the 4a warning and pilot, and the R4 test can start today. Phases 4b, 4c and 5 cannot. The 4a fix touches 23 clues across every case.

Edge cases

Rewriting a question toward a quantity can change the skill it tests, which is why R3 pilots first. Case 101's Halloran clue has to satisfy both the pair rule and the season's trace locks. The source guard has to strip comments, or a doc comment could trip it.

Rendered 24 September 2026 from .claude/plans/thread-container-followthrough.md. The Markdown file is the source of truth.