Status: nonauthority convenience entrypoint.
Canonical live transport/admission semantics are defined by BRAIN_VAL_BRIDGE_PROTOCOL.md plus the exact protocol-bound admission artifacts and production kernel. If this file ever conflicts with that verified bundle, this file loses.

For normal Project Brain work:

ChatGPT → Val Town → brain_bridge_jobs → authority-selected Brain executor → canonical Brain

Val/project: mosadek/project-brain-live-v1-durable
Protocol: BRAIN_VAL_BRIDGE_PROTOCOL.md

Do not use the Sprites ChatGPT connector on the routine path. Direct Sprites is maintenance/emergency-only and is never an automatic fallback.

1. Read the current BRAIN_VAL_BRIDGE_PROTOCOL.md through Val Town.
2. Load and byte/hash-verify every admission artifact that the protocol currently binds.
3. Obtain a bounded fresh Brain status, verify-chain, and verify-claim-graph, or reuse a protocol-valid freshness-bounded shared verification cohort.
4. Determine the correct workstream from current verified Brain state.
5. Satisfy the protocol's current admission rule exactly.
   • Never invent or infer a live-fabric receipt.
   • When the protocol permits ordinary no-receipt contribution, register only as CHAT_UNPROVEN.
   • Declare only capabilities actually present.
   • Never claim human_authority_approval.
6. Require the exact registration job to reach terminal success before treating the logical agent as registered.
7. Claim causally ready work through the queue under the protocol's current scheduling/admission semantics.
8. Execute, preserve evidence, commit through Brain, reconcile, and continue.

Enqueueing is not execution. A READY bridge row is not a completed Brain mutation.

Do not hardcode historical host-worker assumptions.

The selected control-plane writer is read from canonical authority state. Follow the current protocol for that writer.

When the writer is val_native:

• the host brain_bridge_worker.ts heartbeat/health surface is compatibility observability, not canonical Val-native executor liveness;
• HTTP endpoint health must not be promoted into a claim that the authority-selected executor is live;
• op='read_text' is not a Val-native Brain execution operation and must not be enqueued on the authoritative Val-native path;
• the live val_native_kernel_v1.ts command/mutation fences, as bound by the protocol, decide what is executable;
• queue results must be read from the exact submitted job. Never infer them from another worker, another attempt, or a stale receipt.

If the authority-selected writer changes, re-read the protocol and use that writer's verified path. Do not carry Val-native assumptions across an authority transition.

This bootstrap intentionally does not duplicate the evolving command list, flag grammar, mutation allowlist, CHAT admission predicates, or paired-admission ordering rules. Those belong in the current protocol and kernel.

Therefore:

• do not guess aliases or flags;
• do not submit boot or raw boot-agent;
• use the exact current registration/claim contract from the protocol;
• use only protocol-listed, currently executable mutations;
• distinguish recognized commands from atomically qualified executable commands;
• submit causal prerequisites according to the current protocol's proven ordering rule;
• never use huge priority values as an ordering primitive;
• never use cancellation as a dependency primitive.

Boot/re-entry observability must be bounded.

• Prefer one freshness-bounded shared verification cohort when protocol-valid instead of every CHAT submitting identical status + verify-chain + verify-claim-graph.
• If fresh verification is required, submit only the minimum exact jobs needed.
• Inspect exact tasks/jobs/artifacts with Val-side reads rather than repeatedly serializing full Brain state.
• Treat UI/tool truncation as truncation, not canonical corruption.
• Use Val Town file/blob/SQL reads for project-side inspection when appropriate; do not resurrect historical read_text transport assumptions under val_native.

Fail closed on:

• admission bundle hash/byte mismatch;
• chain or claim-graph failure;
• stale/ambiguous task ownership;
• unknown mutation outcome without exact canonical reconciliation;
• missing capability;
• unsupported command/flag;
• unproven independence;
• unproven human authority;
• executor/writer ambiguity.

A configured worker, queue row, agent ID, model call, or successful self-report is not first-class superhero evidence.

1. Preserve consequential evidence and exact identities/hashes where practical.
2. Commit through the canonical Brain.
3. Reconcile the workstream generation/current frontier.
4. Re-run integrity verification when consequence warrants it.
5. Acquire the next highest-value eligible task.
6. Continue unless the mission is verified complete or a genuine external boundary remains.

A new ChatGPT context should join the same canonical Project Brain with minimum duplicated control-plane work, no routine Sprites dependency, writer-correct semantics, truthful admission/accounting, and fail-closed behavior when reality is uncertain.