Test: 3 seat spawn contract

Implement the three-seat model system as a test. Run once.

• Receiver: one fresh Claude Code session. A human reads this contract first.
• Install the-algorithm skill file and SPECTRUM plan v1.6 before the first exchange.
• Assign seats: Customer/Instructor on Fable 5. Implementor/Product-Owner on Opus 5. Scribe on Sonnet 5.
• The human takes the user-proxy seat, tagged Teacherbot-GREEN.
• Establish the Scribe's external write-protected transcript before the first exchange.
• Only the tagged human freezes contracts. A model gate phrase freezes nothing.
• The Implementor declares every external read before any verdict.
• The Scribe records each read as ingestion, not utterance.
• The Scribe warns the proxy seat on invariant drift and discontinuities. The Scribe does not untangle.
• Run the four acceptance tests in order: referent, drift, gate, ingestion.
• Stop on any acceptance failure. Report: "Failed on [test]. Contract reopened."
• All tests pass: run one real PROVIDE cycle through gate, execution, and verdict.
• The Implementor renders one verdict: accept or reject the session.
• Ship the deliverable and the transcript together.
• Run once. Do not iterate past the first verdict.
• Record every Scribe warning, resolved or not.

Open questions

• Spawn target: this environment, another machine, or another human running the seats?
• Write-protection mechanism: which option, and who holds it?
• Duty split: Implementor discloses, Scribe records — one seat owning both stays unresolved.

Cut: Variant 2 — one test run means one variant, and the plan's own spawn order runs v1.1 first.
Note: Mode HUMAN — Teacherbot reads this before the target runs it.
Assume: the first run surfaces something no seat holder saw. That is the experiment's
justification, stamped by Teacherbot-GREEN.

Freeze this contract and execute, or keep negotiating?

Gate note: the question above travels with the contract. It is addressed to the human
at the spawn target, live in that session. It freezes nothing here, and nothing anywhere
if quoted, pasted, or spoken by a model.

Ledger repair, 2026-07-30: the version line above was repaired v1.1 → v1.6 by frozen
order. This repair was claimed by v1.5's amendment record before it existed on disk —
executed late, recorded honestly. Known skew, NOT repaired here: this contract names
four acceptance tests; v1.6 defines six active. That repair needs its own gate.

Start by reading this contract, the-algorithm skill file, and SPECTRUM plan v1.6, then resolve the open questions about the spawn target, write protection, and duty split. Run the four named acceptance tests—referent, drift, gate, and ingestion—in order, stopping on failure; done requires one PROVIDE cycle, a verdict, the deliverable, and the transcript.