0075: Preserve LLM parameter authorization through composition¶
2026-10-25. Issue #108, progressing #19. Baseline:
baeee45f9564f67cfcb03ed35217684e2255c3a9.
Hypothesis and scope¶
The composed workflow can preserve the existing signed LLM parameter enrollment, reject invalid authorization before paid preparation, and check validity at the final join. No new registry or signing format is needed. The standalone verifier's authorization checks become shared script helpers; its exact case, source identity, three verification-time checks and signature semantics remain unchanged.
Mutable paths are the two controllers, composition tests, documentation and numeric evidence. The model, quantization, prompt/answer, arithmetic, proof relations, classifier data/splits/thresholds and separate verifiers stay frozen. The new source inventory needs its own authorization; old signatures are not reused to authorize changed controller bytes.
Boundary¶
The composed controller accepts the same --enrollment / --trusted-key arguments and
LLM_ENROLLMENT / LLM_TRUST_KEY environment variables as the native verifier.
Either input requires the other. The caller independently supplies the public
key; no key is obtained from the signed record or proof bundle.
Before any fresh LLM or classifier preparation, validate the signature, exact
case/source pins and validity interval. Snapshot the signed bytes into the new
run directory and pass that copy to native verification. Require its reported
authority to match the caller's authorization, excluding check time/stage only.
After classifier proving and source checks, validate the same authorization again.
Expiry at that final join leaves a failure, never a composed success. Results
report parameter_authority; unsigned research runs report checkout-case.
The signature authorizes the LLM case and verifier source inventory. It does not authorize classifier policy, setup keys or operational release. This extension provides no persistent revocation/rollback mechanism, key custody, independent security/hiding qualification or hardware provenance. The v1 registry remains separate and its entry meanings are unchanged.
Acceptance and bounded experiment¶
First reproduce missing enrollment support, then require invalid/partial records, wrong keys, changed parameters and expiry to reject before preparation. Fixture tests must also reject changed native authority and expiry during classifier proving, and preserve unsigned behavior and a frozen caller snapshot.
Run one actual signed make e2e-composed campaign using the already retained
LLM/classifier preparation from decision 0074. Rebuild and freshly verify all
23 LLM segments, requiring two honest checks and all 36 mutations. Create three
new classifier setups/proofs over the same opened request, require all eleven
binding rejections and 399 arithmetic outcomes, and retain the synthetic CNN's
block. Compare all model/circuit and LLM proof identities to the frozen baseline.
Standing local-research/storage authority permits this foreground experiment: one campaign, 7,200 seconds, four worker threads, 8 GiB sampled descendant RSS stop, 12 GiB new allocated storage and a 32 GiB free-space floor. Native workers retain their tighter 4 GiB stop. New cloud jobs/spend are zero. Dependency caches remain populated; locked builds may fetch software dependencies. A disposable issuer key stays only in the local issuer process; only the public key, signed record and public test evidence are retained. No contact, private input or key provisioning occurs. Preserve failures in new run directories.
Local repository, required-training and installed core/training package gates remain required, followed by exact-head and exact-merge hosted checks.
Outcome¶
The actual signed Make command passes in 866.49 seconds. Fresh native verification takes 211.64 seconds, accepting two honest checks and rejecting all 36 mutations. Three new classifier setups/proofs and their checks take 653.43 seconds; all eleven binding mutations and 399 separate-arithmetic outcomes pass. The final authorization check succeeds after those stages. All five checks bind the same record/public key, and all 69 original LLM files, both model identities and all three circuits match the baseline. All new setup keys differ from the prior run.
Allocated growth is 3,345,911,808 bytes; sampled summed descendant RSS peaks at 2,490,826,752 bytes. The sample interval is one second. These are one diagnostic observation with other host activity uncontrolled, not latency distributions. New cloud jobs/spend are zero. Ten real CLI preflight cases reject; the first harness incorrectly expected WireError rather than the correct SIGNATURE_INVALID HallmarkError for a tampered signature. Its failed harness, inputs and logs remain retained; no full campaign had started. Source behavior was not changed to pass it.
Numeric evidence precedes this interpretation. The signer remains a disposable integration key. No operational custody, revocation/rollback, classifier-policy adoption, independent security/hiding review or hardware claim follows.