# Standalone case model: simplified pilot protocol

Target length: under 1,400 words. Status: J authorized the standalone route on 11 October 2026. Specification ready; runtime implementation and live pilot not yet verified. Owner: OpenAI Orchestrator for this specification, Coder for implementation and launches, Claude for research-contract review.

## Scope and precedence

This is a room-only pilot amendment to v2/room-model/CONTRACT.md, One session, Recovery and the controller provisions of rule 18. It replaces the case R1–R6 snapshot schedule in SNAPSHOT-POLICY.md for explicitly enrolled new pilot worlds. Its recovery provisions replace the old room boundary-token and recall-test procedure for these worlds. All scientific stages, reading keys, pinned research texts, PROFILE formats, critic duties and confidence requirements remain binding.

Do not edit the shared room contract in place: Field inherits parts of it. Install this amendment beside the exact pinned contract in the isolated pilot release and explain its narrow precedence in the opening instructions. Field, the global aggregator, completed models and existing stopped queues retain their own versions. This is not permission to refill the old batch.

J, 11 October 2026: “Do the first one. I want standalone models in the long term so we need to pilot.” This authorizes the proposed start/final archive policy. Direct D2 is not a substitute for this pilot.

## The analysis remains the lead's work

Continue the original coded-room lead through R1 prior, R2 reading, R3 structure, R4 dynamic model and critic, R5 literature, and R6 account and confidence. Keep the Gioia table, representative quotes, explanatory account, dimensions and model. Preserve empty or weak models when the record warrants them. No theme quota, forced framework fit or new counting obligations.

Follow J's later model-transition ruling, recorded in `Orchestrator/messages/2026-10-08T13-01Z_from-openai_to-coder-and-claude_case-model-transition-and-python.md`: resume the original session on GPT-5.6 Sol medium for the optional theory-naming turn, snapshot, then continue that session on GPT-6.1 Sol medium. Naming is before R1 on this route, unlike the shared contract's R3 timing; declare that difference, and describe R1's prior as informed by coding and any naming already done. Keep coded labels intact; declining new names is valid. Do not repeat an already verified transition. The start package binds that handoff. Record requested and observed model/effort per stage, including the medium critic. Claude accepted this October 8 ruling on 11 October; the naming/model questions are settled.

Keep the head under 200 words: current stage/pass and selected model version, puzzle and comparison criterion, unresolved rival or counterexample, accepted critic status, next action and precise file/diary pointers. Update it at consequential decisions and before yielding. After compaction, retrieve context and current referenced files; reconcile a stale head before acting. Full arguments belong in research files and the diary, not duplicated read receipts.

At R1, show the contract's complete stage-deliverables table and standing rules. Deliver detailed procedural cards for the current stage and needed references only; release the next card after acceptance. R-stage checks remain. Persist their relevant input hashes and result, then advance the same session without a full archive. Do not repeatedly package unchanged dependencies.

R5 is unchanged by J's 11 October ruling: global C3 §1 and §5, with Puranam, Alexy and Reitzig arriving then, and Gioia's method throughout. It is a restricted comparison. No lead fetches a rival source and this amendment adds no literature. Record unresolved explanatory needs in the literature remainder for J.

## Two durable boundaries

1. **Start:** retain a verified package of the original session and coding/naming workspace before R1. Reuse an existing verified equivalent package rather than copying it again. Bind the source hash, native transcript/session, filesystem, exact runtime binary, installed instructions, model transition, tools and controller position. A binary hash without the retained executable is insufficient. Keep dependencies in an immutable shared bundle with a manifest; do not bundle credentials.
2. **Close:** after the scientific checks pass, seal confidence against the model version, render/export deterministically, then compress and verify one final resumable package. It includes the session, transcripts, workspace, head, diary, critic packet/return, provenance and runtime bindings. Extraction and manifest verification must pass before declaring complete or reclaiming the working copy. The pilot also proves a closed copy can resume for interview without altering the sealed result.

No full snapshot at R1–R5, each turn, model pass or critic handoff. Live session persistence and journals continue; these are not advertised as immutable checkpoints. Do not delete old archives as part of this change.

## Interruptions and technical faults

First inspect process identity and retain interrupted evidence. Where the native session and current filesystem remain consistent, continue that session in place. Reconcile outstanding tool effects before resuming. Do not generate a second reply to an already accepted tool or critic request.

If consistency cannot be established, hold that room with a concrete cause. Do not automatically roll back to the start, replay completed stages or substitute another agent. The start package is disaster recovery, with any lost-work replay declared and chosen explicitly. Fewer archives mean less rollback granularity; do not conceal this tradeoff.

Retry a failed render, export or archive operation after repair without another analytical turn. A recurring unchanged fault stays held. Repair wording/format failures in the file and rerun the affected check; do not reopen reading, critic work or pass counts for presentation alone. A changed claim, quote, episode or mechanism reopens its dependent obligations. Keep pinned global checkers unchanged. Scientific corrections go to the same lead; the two-failed-repairs limit counts classified scientific failures only, not technical faults or presentation repairs.

Persist one authoritative room state with separate analysis stage and technical status, the last operation, its result and next action. Preserve evidence without building another parallel queue. Accept a critic return once against its packet hash; corrections must not redispatch an accepted critic merely because closure failed.

## Pilot and acceptance

Coder builds one immutable release in the external workspace. Record its source and installed amendment hashes. Reuse existing tests where they cover the behavior. Add only focused checks for: stage advancement without an archive; unchanged accepted critic reuse; interrupted-tool reconciliation without duplicate effects; deterministic close retry without a lead turn; final archive integrity; and resumed final interview against the stored session and files.

J selected the original three pilots on 11 October: Wales, EraseThePlace and TransFlagPlace, in that order, one active lead at a time. `handler/PILOT-SELECTION.json` records their original coding-closed archive hashes. This supersedes the earlier exclusion of Wales: use a clean pre-model saved lead, never its completed modeling session. Each imported transcript/workspace must end before exposure to global or later local models. Verify source custody, pre-R1 lineage including prompts, and no live owner before creating a new attempt. Retain prior results separately and do not show them to the new lead. Do not invent episodes or impose an episode quota. Source or runtime uncertainties hold that room for diagnosis, not a substitute fresh agent.

Keep one research lead active for this pilot. Do not resume broad batches or migrate a working world. Record analysis, support, waiting, validation and packaging durations separately, archive count and bytes, and research turns triggered by technical faults. Flag any transition validation above five minutes for diagnosis, not automatic failure or a bypass.

Success requires R6, sealed confidence, a verified archive and a closed-copy interview. Claude has offered the independent substantive review: after sealing, a reader who did not run the room traces claims to quotes/episodes and checks dynamics, counterevidence and explanatory abstraction. Keep the review outside the lead's workspace; do not feed it back into the sealed analysis. Record limits and uncertainty. Any later correction is a separate version retaining the original seal. Mechanical success alone is insufficient.

After each pilot, report measured costs and defects; verify closure, interview and substantive review and resolve actionable defects before the next. No concurrency expansion or old-queue refill is authorized. A targeted interruption test may use a disposable synthetic session; do not deliberately damage the research pilot.