R12 Post-Commit Packet Transport V3 Preregistration

Status: frozen repair specification only. V2 is a post-commit NO-GO. No neural fit, Shohin adapter, SFT, H100 job, workspace claim, or reasoning claim is authorized unless v3 passes a clean post-commit run and an independent adversarial audit.

1. Frozen question

Can an explicit four-cell packet be transported through source-free one-event processes and answer late linear consumers after fresh output recoding, while a rank-two motor packet fails at exactly chance on hidden coordinates, under an OS-enforced interface and independently recomputed evidence?

The algebra, five public cells, 15 decisive cells, depths, state/motor packet definitions, exact score thresholds, stale/shuffled/horizon controls, and 337-role cardinality per core are unchanged from v2. V3 changes custody only. The v2 numerical result may not tune any cell, threshold, or control.

2. Claim-blocking v2 defects

V3 exists only to close these frozen defects:

1. a self-hash-consistent attacker could alter invocation evidence and make verify_report accept the forgery;
2. child processes could traverse the shared temporary parent and repository;
3. the challenge was deterministic from committed bytes before phase one;
4. the completed artifact/result hash had not been committed before audit.

One failure to close any item is a v3 NO-GO.

3. OS-enforced role boundary

Every writer, updater, reader, raw reader, and oracle must execute through /usr/bin/sandbox-exec; absence or failure of that launcher is fatal.

The generated profile must:

• start from (deny default) and deny network access, so no finite path blacklist can be bypassed by another writable mount or directory;
• import neither system.sb nor dyld-support.sb, because both grant broad inherited reads; inline only the fixed non-filesystem dyld bootstrap operations and explicitly allow content reads from the root-owned Command Line Tools Python runtime, one exact committed role file, one fresh invocation directory, and the exact /, /dev/urandom, and /dev/null bootstrap objects required to launch that runtime;
• allow writes only below that invocation directory;
• use exact metadata grants only for path traversal to those allowlisted objects, never a metadata subtree over the run root or repository;
• execute the root-owned Command Line Tools Python binary rather than a user-writable framework, Homebrew, virtualenv, or repository interpreter;
• re-allow read access to exactly the committed seed-free role executable;
• re-allow only the invocation directory and its declared packet files;
• give stdin/stdout-only roles a fresh empty invocation directory;
• remove PYTHONPATH, set HOME and TMPDIR to the invocation directory, disable user site packages and bytecode writes, launch Python with -I -S, and use an allowlisted environment; and
• record a path-normalized policy hash in every invocation.

Before phase one, an executed sentinel probe under the identical policy must prove that the child can read its allowed file but cannot read a parent-run sentinel, cannot read AGENT_RUNBOOK.md, cannot read or write /tmp, cannot write /Applications, cannot open a network socket, and cannot write outside its invocation directory. Parent-run and repository directory enumeration must also fail. Separate sentinels must also prove that writable /Library/Caches, /private/var/db/DiagnosticsReporter, /Library/OSAnalytics/Diagnostics, its data-volume firmlink, /Library/Logs/DiagnosticReports, and /Library/Application Support/Blackmagic Design content is neither readable nor writable. /etc/passwd content and metadata and /System directory listing must also fail, specifically detecting inherited baseline-profile grants. The report must bind that probe. Directory listings remain necessary but are no longer treated as sufficient isolation.

4. Post-phase unpredictable challenge

The canonical parent harness must sample the primary 256-bit nonce with secrets.token_bytes(32) only after both phase-one writers exit and the phase-one manifest is fsynced and mode 0444. The nonce is then fsynced to a mode-0444 commitment file. Challenge seeds are derived from:

SHA256(domain || phase_one_manifest_sha256 || challenge_nonce)

Here domain and the lowercase manifest digest are ASCII bytes, the nonce is the raw 32-byte value, and the deterministic challenge generator is seeded by the unsigned big-endian integer represented by the digest's first eight bytes.

No nonce, seed, event, consumer, codebook, or challenge path may be visible to the writer. Each updater sees exactly one current event and its one input packet, but no source, prior/future event history, consumer, output recoding, nonce, seed, codebook, or challenge path. The artifact records the nonce and commitment hash so an auditor can replay the exact challenge.

The artifact proves commitment, ordering consistency, and deterministic seed derivation, but it cannot cryptographically prove that the recorded bytes came from an honest entropy source. The final external reviewer must witness the canonical CLI path; no external nonce or entropy-attestation claim is permitted.

The second complete core reuses the first core's committed nonce solely for byte-identical replay. Its writers still run before receiving any challenge input and remain OS-confined from the nonce and phase-two files.

5. Independent verifier

pipeline/audit_post_commit_packet_transport_v3.py must not import the orchestrator or role executable. It must independently implement canonical JSON, F_17 algebra, packet serialization, challenge parsing, transport, oracle, reader, collision, horizon, stale, shuffled, and score calculations.

For each core it must reconstruct the exact expected invocation sequence and validate every role record, including command/arguments, stdin/stdout hashes, file bytes/modes/hashes, exit code, directory inventories, sandbox launcher, normalized policy hash, and scientific source identity. It must reject self-hash-consistent mutations of any evidence field and independently verify both complete cores and final payload hashes.

This is evidence reconstruction under an honestly executed committed local harness, not cryptographic proof of historical execution against a malicious machine owner. The independent verifier can prove exact consistency and replay all committed finite calculations; a later external reviewer witnesses the canonical run and inspects the source. V3 must not claim hardware-rooted or tamper-proof attestation.

Regression tests must mutate at least exit codes, extra visible files, packet hashes, output permutations, missing records, sandbox fields, nonce binding, both-core self-hashes, and publisher receipts. Every mutation must fail.

6. Canonical execution and custody

The scientific commit must include this preregistration, orchestrator, seed-free role executable, independent verifier, and both test suites. The canonical path is:

artifacts/r12/post_commit_packet_transport_v3.json

The command must fail closed if any scientific path differs from HEAD. It must build two full byte-identical cores before writing, run both the internal verifier, and hand a read-only candidate to the independent verifier process. Only that independent process may create the canonical artifact. It must first reconstruct all evidence with Git verification enabled, then atomically create, fsync, and chmod the canonical artifact 0444, and finally create a separate read-only receipt at artifacts/r12/post_commit_packet_transport_v3.receipt.json. The receipt must bind the final artifact path, bytes, mode, MD5, file SHA-256, payload SHA-256, scientific commit, verifier source SHA-256, committed verifier-path SHA-256, and Git-verification mode. The parent process must contain no canonical writer and must not accept a caller-supplied dictionary as proof of verification.

The completed result document must record commit, bytes, mode, MD5, file SHA-256, payload SHA-256, exact scores, both invocation ledgers, and independent audit verdict. The artifact and result record must be committed before the final claim audit.

7. Pass boundary

A v3 pass validates only the exact symbolic packet algebra and reproducible process-isolation behavior under the committed harness. It does not establish tamper-proof historical attestation, learning, reasoning, language ability, or novelty. Only after an independent audit passes may a separate learned rank-thinned recurrent packet-lane preregistration be written.