# Atlas / Understanding-Ecosystem — External Review (ChatGPT) + Dispatch Adjudication **Snapshot:** 2026-07-16. Input = a ChatGPT critique of the end-of-round understanding/assurance ecosystem (internal brain dump → portable brief → workflow review → system review → System Atlas). Owner directive: "As for the portable dump: go with recs." The rest is treated as HYPOTHESES to adjudicate (per the ecosystem's own rule + ChatGPT's own #5/#6), not auto-adopted. This doc is the design input for when the deferred deliverables are built (after the passes + brain-dump re-baseline). **⛔ Nothing here is implemented. These are spec refinements to UNBUILT deferred deliverables, pending owner adoption. Governance unchanged: map reality only; proposals stay "candidate — not adopted."** ## ChatGPT's points → Dispatch verdict (ADOPT / PARTIAL / REJECT · confidence) 1. **Evidence — not the brain dump — is the epistemic ROOT.** The brain dump is the canonical *reconstruction/synthesis*, not the root; every claim points DOWN to an evidence type: owner-approved intent · repository reality (at a commit) · deployed reality (served rev, SW/cache version) · runtime reality (real-browser/offline/migration/storage) · verification evidence (tests/gates/reviews + their limits). Corrected chain: **evidence baseline → canonical internal reconstruction → portable projections + Atlas views.** → **ADOPT. High confidence.** This resolves the real contradiction (a doc that can be corrected by the repo isn't the root). Fixes how we frame the brain dump forever. 2. **Add ONE tiny artifact: a Round Baseline Manifest** (mostly machine-readable) to prevent snapshot skew (Atlas=commit A, dump=B, tests=C, deployed=D). Records: round id · commit · branch · timestamp · brain-dump version · deployed rev · SW+cache version · schema version · game-data version · test/gate run used · post-deploy result · known exclusions. Every artifact pins to the SAME baseline. → **ADOPT. High confidence.** It's the one genuinely-worth-it new artifact; it changes a real decision ("are two claims about the same RobCo?"). Small, mostly generated. 3. **Separate LIFECYCLE status from EVIDENCE status** (two independent axes per node/claim). Lifecycle: current/incomplete/queued/proposed/deprecated/superseded/intentionally-absent. Evidence: verified/inferred/unverified/contradicted/stale/unknown. → **ADOPT. High confidence.** Kills the dangerous conflations (planned≠unverified, current≠working, documented≠implemented, tested≠verified-in-deploy, not-found≠doesn't-exist). This is the same "green is scoped evidence" DNA. 4. **Add 3 Atlas VIEWS (not artifacts):** (a) provenance/freshness (what supports this, when last verified, which baseline, contradicted?), (b) change-since-last-round (temporal drift — needs STABLE IDs for features/subsystems/protocols/docs/roadmap), (c) repository→deployment→client version-plane view (where each version lives + where verification stops; justified by the historical silent-SW-failure). → **ADOPT. High confidence.** All three are epistemic gaps the current 8 views miss. Stable IDs are the enabling prerequisite. 5. **Reframe the system review as an "External System MODEL Review"** — it reviews your *representation* of RobCo, not RobCo. It CAN find contradictions/weak reasoning/blind spots; it CANNOT prove a subsystem exists/behaves/deploys as described. Findings format: **Concern → why it matters → evidence needed → what would falsify it → confidence.** Agreement raises investigation PRIORITY, not truth. → **ADOPT. High confidence.** Consistent with the whole "AI agreement isn't truth" spine. 6. **Don't fold review answers into the Atlas without ADJUDICATION.** Every external finding gets a disposition: confirmed/partially/unsupported/contradicted/duplicate/intentional/outdated/candidate-not-adopted/no-action. Chain: external finding → evidence investigation → verdict → owner decision → optional candidate. Atlas keeps the observation + its adjudicated status (so a disproven-but-dramatic comment can't stay prominent). → **ADOPT. High confidence.** This is exactly THIS doc's own pattern. 7. **One shared entity/relationship MODEL — but don't force everything into graph form.** Graph for relationships (ownership/deps/nav/reads-writes/protection/shared-vs-game/consumers); linked evidence for depth (rationale/incidents/uncertainty/long test evidence/temporal). Atlas = an INDEX into evidence, not a swallow-everything store. Node links to the brain-dump section / repo evidence / test / protocol / deploy result / external finding. → **ADOPT. High confidence.** Prevents the Atlas from becoming an unmaintainable everything-blob. 8. **Control granularity aggressively** — primary layer (major screens/capabilities/subsystems/state-owners/persistence/AI-boundaries/protocols/release/roadmap) + drill-down layer (files/functions/tests/schemas). Default view is NOT every file/function. → **ADOPT. High confidence.** Directly protects the "generated artifact, not a maintenance sink" constraint. 9. **Add an explicit MAP-COVERAGE BOUNDARY.** "No material gap found" means only "within the mapped+verified scope" — never "proof of absence." Expose: mapped / partially-mapped / unmapped / unsupported-claims / failed-to-parse / excluded-this-round / runtime-not-exercised / negative-conclusions-with-limited-scope. → **ADOPT. High confidence.** Critical: an AI-generated Atlas can omit a subsystem then confidently say nothing connects to it. 10. **Portable brain dump — justified, ephemeral, one constraint.** Not redundant (different audience, ephemeral). Keep generated fresh; include the baseline manifest + intended review question + known omissions + which claims are owner-intent vs repo-fact + where it summarizes vs full evidence. Do NOT maintain model-specific FACTUAL versions — per-AI customization affects structure/terminology/framing/emphasis only; ONE canonical content source generates all variants. → **ADOPT (owner: "go with recs"). High confidence.** 11. **Don't auto-run every review every round** — risk-triggered, not ceremonial. Re-baseline the reconstruction when the round materially changes the system; regenerate the Atlas when the graph materially changes; workflow review when the workflow changed / a workflow-caused failure / blind-spot evidence; system review after a substantial architectural/gameplay/data/AI-boundary/platform change — not after a contained bug-fix round. → **ADOPT. High confidence.** Matches RobCo's risk-triggered-audit philosophy exactly. 12. **Explicitly do NOT add** (ChatGPT's reject list): a standing protocol catalog · test-coverage matrix · second architecture diagram · separate feature inventory · permanent AI-opinion DB · per-AI maintained brain dumps · a broad ADR system · a permanent review-derived ideas backlog · more reviewers just to raise agreement. Candidate-additions stays ATTACHED TO ITS BASELINE — promoted by owner decision or left a historical observation; never a silent backlog. → **ADOPT. High confidence.** Aligns with our governance-restraint + net-protocol-count discipline. ## Recommended final structure (ChatGPT's, endorsed) EVIDENCE BASELINE (owner intent · repo reality · deployed/runtime reality · scoped verification) → CANONICAL INTERNAL RECONSTRUCTION (evidence-linked) → GENERATED PROJECTIONS (portable brief + Atlas views) → INDEPENDENT CRITIQUES (blind workflow review + portable-brief system-MODEL review) → ADJUDICATION (evidence check → verdict → owner decision → candidate-not-adopted). ## Dispatch's net read The critique is strong and I adjudicate **all 12 as ADOPT** for the deferred-deliverable design — but they are DESIGN refinements to unbuilt artifacts, not app/protocol changes, so adopting them costs nothing now and improves the eventual build. The single most important is #1 (evidence-as-root). None of this is built until after the passes + brain-dump re-baseline. Owner confirmation requested for folding the non-portable-dump items into the deferred spec (portable dump already confirmed).