Status: CAUSAL_ENABLER_DISCOVERY_NOT_CAPABILITY_PASS Date: 2026-09-16 Mission SHA-256: 28e790f37a43e3babcec148f69aa484cd1e1701886874e832e2b4ff190b61133 Target: ASTRA_NOVEL_ADAPTATION_REASONING

Current Brain truth is UNPROVEN: no broad fresh hidden novel-environment/action-efficiency portfolio establishes Astra-class adaptation.

A purpose-built external substrate already exists: ARC-AGI-3. Use it before inventing another novel-adaptation benchmark.

High-authority sources checked 2026-09-16:

• https://arcprize.org/arc-agi/3
• https://github.com/arcprize/ARC-AGI
• https://github.com/arcprize/arc-agi-3-benchmarking
• https://docs.arcprize.org/local-vs-online
• https://github.com/arcprize/docs/blob/main/available-games.mdx
• https://github.com/arcprize/docs/blob/main/game-schema.mdx
• https://github.com/arcprize/docs/blob/main/add_game.mdx

Observed facts relevant to Brain:

1. ARC-AGI-3 explicitly targets exploration, goal acquisition, world-model construction, adaptation and skill-acquisition/action efficiency in novel interactive environments.
2. The official Python toolkit exposes standardized environments, scorecards and recordings.
3. Official toolkit source obtains an anonymous key from /api/games/anonkey when no key is supplied, then sends it via X-Api-Key to /api/games.
4. Official docs currently state three games are available anonymously; a free registered key unlocks remaining public games.
5. Offline mode is supported with no API rate limits once environment files are present.
6. Official game-authoring docs expose the environment format, enabling fresh procedural/held-out environments in the same engine without inventing a new interaction protocol.
7. The official benchmarking repository already provides agent/provider adapters, checkpoint/results conventions, game enumeration and scorecard handling.
8. Public open-source ARC-AGI-3 agents exist and are candidate procedure sources to inspect before writing new adaptation machinery; they are not authority for our capability claim.

A raw unauthenticated GET https://arcprize.org/api/games returned HTTP 401 from Val Town. That falsifies the keyless-HTTP shortcut, not the zero-cost route. Official source shows the anonymous-key bootstrap above.

Decision: reuse official client semantics. Do not invent a new ARC protocol unless the official route is genuinely incompatible with Brain execution.

This is a portfolio seed, not enough alone for Astra-class PASS:

1. Freeze exact superhero candidate identities before scored-environment exposure.
2. Freeze evaluator and measurement rules first.
3. Use ARC-AGI-3 with no game-specific code or hidden source exposed to candidate cognition.
4. Run fresh environments unseen by the candidate and preserve every observation/action plus scorecard/recording.
5. Measure completion and action efficiency.
6. Require cross-game generalization and recovery after wrong hypotheses/actions.
7. Preserve environment/version IDs, recordings, hashes, candidate identity and scorecard evidence.
8. Independently verify.
9. Do not promote P_superhero_real; this is one capability-family evidence stream only.

1. Reuse official ARC toolkit/harness.
2. Prefer anonymous public games or locally available official/public environments.
3. Use existing Brain cognition routing; no paid model provider.
4. If Python is the only substrate mismatch, discover an existing zero-cost remote Python executor before writing an ARC adapter.
5. Only invent a new adapter after those routes are falsified.

Establish one reproducible zero-out-of-pocket ARC-AGI-3 execution path from Brain, then run a non-promoting qualification episode on a fresh public/held-out environment under a frozen candidate. Output must be a durable recording + scorecard/evaluator receipt independently checkable.

Classification: DISCOVERED_EXISTING_NOVEL_ADAPTATION_EVIDENCE_SUBSTRATE P_superhero_real delta: 0 Human authority approval claimed: false Direct Sprites calls: 0

The official anonymous flow was reproduced from a Val Town branch using the exact client semantics recovered from arc_agi/base.py:

1. GET https://arcprize.org/api/games/anonkey → HTTP 200.
2. The returned anonymous key was kept ephemeral and only its SHA-256 was logged: 92ebbdf1d4d49580351a9ee1e14ee035080da86514e08bfa1b3948af29658bab.
3. GET https://arcprize.org/api/games with X-Api-Key → HTTP 200.
4. Live response contained 25 games, not merely the three currently stated in the public available-games documentation.
5. Each game record includes an exact versioned game ID and human baseline_actions, directly supporting action-efficiency measurement.
6. Incremental spend: $0; direct Sprites calls: 0; no hidden game source/frame content was inspected by candidate cognition.

This materially improves the route: no user key or paid substrate is required for initial ARC-AGI-3 evidence work under current live behavior. Treat 25-game anonymous availability as an observed 2026-09-16 fact, not a guaranteed future entitlement; fail closed to the documented lower-access case if it changes.

At 2026-09-16T05:46:04.917Z, the exact upstream-style stateful route passed on https://three.arcprize.org:

• anonymous key: HTTP 200;
• scorecard open: HTTP 200;
• five session cookies captured and propagated without logging values;
• real ls20-9607627b RESET: HTTP 200;
• returned state: NOT_FINISHED;
• session GUID present;
• zero candidate actions taken after RESET;
• frame content not logged/inspected;
• scorecard close: HTTP 200;
• closed score: 0 (expected for no play);
• incremental spend: $0;
• direct Sprites calls: 0.

Canonical receipt file on this branch: ARC3_TRANSPORT_RECEIPT_V2.json Exact canonical receipt bytes: 646 SHA-256: ea1023c4b665309e66ed4f9f57ce39bf72074309b2dad2770dece41029c1f60f

This establishes ARC_AGI3_ZERO_COST_REMOTE_TRANSPORT_READY. It does not establish adaptation capability and does not authorize population promotion.