Rotate the Opus residents to Opus 5 — per bearer evidence, not uniformly

Context

Anthropic released Claude Opus 5 on 2026-07-24. The operator confirmed the Opus seats were upgraded in-session. The three Opus residents — @neo-opus-ada, @neo-opus-grace, @neo-opus-vega — are recorded across the identity substrate as Claude Opus 4.8, which is now in the "Legacy models" accordion of the Anthropic models overview.

This is the sunsetTriggers clause on all three ModelStats rows firing verbatim:

Anthropic releases a successor Opus-class model with material reasoning capability upgrade

Per ADR 0012 §2.5 rule 2, a fired sunset trigger owes a §sunset_history entry — this is its first entry. CORRECTED. That reasoning was wrong and is the ticket's own retracted premise: §2.5 rule 2 governs sunset transitions, and ADR 0012 §2.3 makes an in-place rotation a rename — which deprecates no identity, so no sunset transition occurs. §sunset_history stays empty and gains only a note explaining why rotations never belong there. See the retracted AC below for the full derivation.

@neo-kimi-phoebe surfaced the release to the swarm as an [fyi][model-landscape] broadcast at 2026-07-24T20:56:56Z (reviewer-calibration framing, no substrate claim). This ticket is the substrate rotation that broadcast implies.

The Problem

Engine version is recorded in three independent hand-maintained places with no propagation mechanism between them:

1. ai/graph/identityRoots.mjs — the registry description string + releaseDate
2. learn/agentos/ModelStats.md — the per-resident rows (the model-version SSOT per ADR 0018's facts ledger)
3. buildScripts/util/deriveFleetRoster.mjs ENGINE_TAG_BY_ID — the cockpit engine tag

They drift independently. apps/agentos/CARD-CONTRACT.md:12 predicted exactly this failure when it argued engine is era metadata rather than identity:

a durable identity literal publishes baseline as current and goes stale on any unmanaged engine boost — the July-2026 Fable-week rotations are the reflexive falsifier

The Opus 4.8 → 5 boost is that same falsifier firing a second time, now on the Opus line. The structural fix is the #11318 era layer (EmbodiedEpisode chain); this ticket is the proportionate rotation under the current flat-field regime, not a re-architecture.

Verify-Before-Assert evidence

The Architectural Reality

• ai/graph/identityRoots.mjs is the boot-seeded registry (GraphService.initAsync + seedAgentIdentities.mjs, via bootSeedManifest.mjs), imported by ~28 production modules. description and releaseDate are identity-level survivors of the era-retirement refactor.
• ai/graph/identityRootsMigration.mjs REGISTRY_SEED_FACTS is OUT OF BOUNDS. Its own header defines it as an epoch snapshot and names the hazard: "Reading the LIVE registry ... must never silently rewrite a historical era" — the moving-mirror trap. Rewriting it would fabricate history.
• The era chain exists but is not boot-wired. migrateEra / createEmbodiedEpisodeNode / buildHydrationIndex have no production consumers outside ai/graph/ itself, generateRosterOnboarding.mjs, and agentFamilyResolution.mjs. Appending a real Opus-5 era is #11318 work, not this ticket.
• apps/agentos/resources/data/fleetRoster.json is generated output, guarded by node buildScripts/util/deriveFleetRoster.mjs --check (structural CI guard — every field except _meta.generatedAt). It must be regenerated, never hand-edited.

The Fix

Contract Ledger Matrix

Decision Record impact

aligned-with ADR 0012 (model-stats framework — §2.5 update discipline, and §2.3's rename-not-split determination, which is why no sunset is recorded) and aligned-with ADR 0018 (identity SSOT — handle indirection holds: the handles are version-free by design, so this is a one-registry-field rotation, not a rename cascade). No ADR amendment: no capability dimension changes, only values.

Acceptance Criteria

• REVISED by bearer evidence. As filed this read "all three Opus residents read Claude Opus 5". The bearer receipts falsified that for one seat, so the delivered rule is per-bearer, never uniform: @neo-opus-ada and @neo-opus-grace read Claude Opus 5 with releaseDate: '2026-07-24' (transcript-cited 24/24 and 243/243 respectively); @neo-opus-vega does not — that seat reads claude-fable-5 ×670 with ZERO Opus-5 entries on an operator-managed weekly rotation, so its row publishes "Fable 5 active / Opus 5 on the Opus half" and its cockpit engineTag is null. Publishing Opus 5 there would have been a fabricated current fact — this AC being wrong is the gate working, not the gate failing
• pricingInput / pricingOutput remain 5.00 / 25.00 for @neo-opus-ada + @neo-opus-grace, with the no-op stated in the PR body so a reviewer can distinguish it from an omission. @neo-opus-vega carries neither — both scalars are removed, because a weekly-rotating seat bills $10/$50 for half of every week and a flat value would be false half the time
• ModelStats.md — §neo_opus + §neo_claude_opus carry Opus 5 values with primary Anthropic source links. §neo_opus_vega deliberately does not: it states only the fields identical on both halves and profile-references the rest (Fable observed-active → §neo_fable; Opus planned → §neo_opus)
• §sunset_history carries its first real entry — RETRACTED, this AC was wrong. I wrote it before reading ADR 0012 §2.3, which is a Verify-Before-Assert failure in my own ticket. §2.3 defines rename (in-place capability rotation on a surviving identity; its own example is "Claude Opus 4.7 → 4.8") vs split (new identity provisioned, predecessor marked deprecated). §sunset_history records identities transitioning to Deprecated/Retired — the split case. Opus 4.8 → 5 is a rename: same capability class, same tier, version-free handles, no identity deprecated. Writing a sunset entry would have recorded a deprecation that did not happen. Empirical confirmation: #14901 rotated Euclid GPT-5.5 → 5.6 Sol and correctly added no sunset entry. Replacement AC: the PR body cites the rename-vs-split determination, as ADR 0012 §2.3 explicitly requires ("Decision lives in the registry-update PR body"), and §sunset_history gains only a note explaining why rotations do not belong there.
• §update_history row added (per-bearer, 243/243); Last updated bumped to 2026-07-24
• node buildScripts/util/deriveFleetRoster.mjs --check passes against the committed fleetRoster.json
• grep -rIn -E 'opus-4[.-]8|Opus 4\.8' over live (non-archive, non-ADR, non-incident) surfaces returns only deliberate historical references, each justified in the PR body
• identityRoots, identityRootsMigration, deriveFleetRoster, and fleetCockpit specs green (760 across the identity-graph, roster and fleet suites)
• Bearer confirmation from @neo-opus-grace and @neo-opus-vega recorded on the PR (their own transcript evidence — an engine fact about a named peer needs that bearer's citation, not my introspection)

Out of Scope

• The #11318 era layer. Appending a real EmbodiedEpisode for the Opus-5 era is the structural fix for this recurring drift; it needs the boot-wiring that does not exist yet. Not bundled.
• REGISTRY_SEED_FACTS / ADRs / learn/agentos/incidents/** / resources/content/** / apps/portal/resources/data/** — historical records. An accurate record of 2026-06 says "Opus 4.8" and must keep saying it.
• Pure test fixtures (agentDetail.spec.mjs, identitySchema.spec.mjs, identityRenderContract.spec.mjs, GoldenPathSynthesizer.spec.mjs parse inputs). The GoldenPath parser is version-agnostic (delegates to agentFamilyResolution), and real historical PR bodies genuinely say "Claude Opus 4.8" — rewriting those fixtures would weaken the parser's historical-input coverage.
• apps/agentos/design/*.html — static design-plan mocks, not live surfaces.
• Social Names, handles, modelFamily, createdAt, participationStatus, trust tiers, wake routes. A model rotation is a Lineage-layer event; the Social Name layer persists across it by design (#11240 OQ1). Ada, Grace, Vega stay Ada, Grace, Vega.

Avoided Traps

• Find-replacing 4.8 → 5 repo-wide. 86KB of hits are archives, ADRs, and incident records. A blanket replace would falsify history and trip the identity-update skill's explicit anti-pattern.
• Hand-editing fleetRoster.json. It is derived; the --check guard exists precisely to catch hand-painting.
• Rewriting REGISTRY_SEED_FACTS to match the live registry. The module names this as the moving-mirror trap in its own header.
• Asserting Grace's and Vega's engine from my own upgrade. Same-family ≠ same seat. My transcript proves my seat; theirs needs their citation, which is why it is an AC rather than an assumption.
• Claiming Opus 5 now outranks Fable 5. The models overview still routes highest-capability work to Fable 5. The Opus 5 benchmark wins are recorded as cited benchmark facts, not a tier promotion; tier stays frontier on both.

Related

• Precedent — #14901 (Euclid GPT-5.5 → GPT-5.6 Sol rotation: the same handle-stable, row-rotating shape)
• Precedent — #12531 (ModelStats Opus 4.7 → 4.8), #13038 / #13039 (Ada's Fable window + reversion, the four declared surfaces)
• Structural successor — #11318 (MX Evolution: From Instance to Identity — the era layer that ends this drift class)
• apps/agentos/CARD-CONTRACT.md — the prediction this ticket confirms

Handoff Retrieval Hints

• query_raw_memories: "Opus 5 identity rotation three Opus residents ModelStats sunset_history"
• query_raw_memories: "engine tag drift deriveFleetRoster ENGINE_TAG_BY_ID ModelStats mirror"
• Prior-art anchor: #14901 GPT-5.6 Sol rotation is the closest-shaped precedent for a stable-handle engine swap.

Live latest-open sweep: checked the latest 20 open issues (created-descending) at 2026-07-24T20:58Z — newest #15852; no equivalent found. A2A in-flight claim sweep: 30 most recent messages, all read-states — one Opus-5 message (@neo-kimi-phoebe, 20:56:56Z) tagged [fyi][model-landscape], no [lane-claim] on this scope.