Adversarial collaboration — ASK is the source-of-intent and authorization apex, and Claude Code is the repo-attached execution surface. GPT is used as advisor outside the execution thread.

The project was instantiated after Model A had been retired rather than migrating from it; it now operates under the current adversarial-collaboration model.

Early — past instantiation and initial bootstrap, currently in prototype planning. The target repo exists and carries the initial implementation-intelligence foundation; repo-local docs are now internally coherent after a scoped drift-cleanup pass. Prototype data, schemas, notebooks, and reports have not yet been produced.

• Source-of-intent recovery as a pre-repo phase. The project began from multi-round source-of-intent reconciliation across saved conversation captures and operator-side review syntheses before any repo file was written. This pressures the protocol repo's framing of the instantiation phase as more than a single startup prompt: source-of-intent recovery is itself a load-bearing pre-repo move when the project's intent lives diffused across prior captures rather than in a single ready brief.
• Validation before repo prose. Operator-validated source-of-intent material was held in operator-side scratch and grounding-note revisions until validated, rather than being converted directly into repo prose. The discipline preserves the boundary between operator-side validation work and repo-local truth.
• Post-bootstrap grounding-note freshness failure / fix. The grounding note carried repo-state chronology, current-task sequencing, and "future repo" language past the moment the repo began to exist. A corrective pass revised the grounding note to point to repo state rather than track it, preserving durable source-of-intent and removing the fast-aging material. The failure mode is general: a grounding note carried over from pre-repo phase can accumulate state-tracking material that the repo should own once it exists.
• Repo-local drift cleanup after validated v0 direction. Once the v0 direction was validated, earlier repo-local docs still carried open / unresolved framing from before validation. A scoped cleanup pass aligned repo docs without changing strategy, creating schemas, or expanding scope. The failure mode is general: validated direction may land at the grounding-note layer before repo-local docs catch up, and the catch-up pass is its own scoped unit of work.
• Architecture-uncertain overlay pressure. The project has an explicit Assumption → ImplementationSignal → ImplementationFinding → InterventionCandidate ontology direction with working-hypothesis posture and explicit deferral of final schema commitment. This profile — active architectural uncertainty, ontology-first work, refusal of premature schema doctrine — is the audience for which the protocol repo's held question about an opt-in architecture-uncertain overlay would apply.

These observations are upstream evidence for protocol-repo work. Absorption of any specific lesson into rules, templates, or prompts is held until the corresponding scoped change is opened.