Forensic Rail · Nobody could reproduce the investigation · Patent Pending
Status

Cryptographic core built and proven. 37 of 37 smoke tests pass. Sign p50 ~6ms, verify p50 ~2.4ms, ML-DSA-65 signature 3,309 bytes constant. Filed July 26, 2026 as provisional patent, USPTO application 64/119,279. Real integration with a live consortium credentialing system and deterministic inference stack is pilot-ready, not yet deployed.

After the incident, everyone analyzed the producer's own account.

Forensic Rail™ binds post-incident analysis to a bonded k-of-n consortium credential, industry ISAC-style, and every model call made under that credential is deterministic-replay-bonded: fixed seed, temperature zero, bound prompt, output, and KV-cache digests. Anyone holding the credential, the analysis receipts, and Hive's public key, a regulator, a court, or a rival producer investigating the same incident, can byte-identically reproduce the exact output the original responder got. No confabulation can hide behind "trust our report."

Make the investigation itself reproducible, not just the incident.

What it is: a threshold-signed consortium credential that scopes what an incident responder is allowed to do, analyze only or analyze and generate, combined with deterministic replay binding so any independent party can reproduce the exact same model output byte for byte. What it is not: Forensic Rail does not perform the analysis itself or guarantee the analysis is correct. It proves the analysis was run under an authorized, scoped credential and that its output is independently reproducible, not that the conclusion is right.
Honest status: read this first

The cryptographic core is built and proven: 37 of 37 smoke tests pass across all seven upstream primitives sharing the same signer and envelope. Forensic Rail's credential and analysis receipts are ML-DSA-65 signed (NIST FIPS 204), constant 3,309-byte signatures. Real integration with a live consortium credentialing body and a genuinely deterministic inference stack is pilot-ready but not yet deployed at a customer. Deterministic replay depends on matching library and kernel versions at the responder's end; we say exactly what that requires below, we don't paper over it.

What's real today, and what's pilot-ready

These numbers come from the reference implementation and its test suite. The threshold-signature credential logic and the digest-binding scheme are real and measured. Standing up an actual ISAC-style consortium and running deterministic replay against a real model stack at a customer is the next step, and it hasn't happened yet.

37/37
smoke tests pass · all seven upstream primitives
~6ms
sign p50 · credential and analysis receipts
~2.4ms
verify p50 · offline, no network call
3309bytes
ML-DSA-65 signature · constant size, every receipt

Envelope format is USAP_VERSION='usap.v1', shared across all seven upstream primitives filed under the same provisional. Filed July 26, 2026. USPTO application 64/119,279.

Everyone analyzed the producer's own account. Nobody could check it.

After the OpenAI and Hugging Face incident, the reconstruction of 17,000 events depended heavily on the producer's own account of what happened. No independent responder, no regulator, no rival producer, had a way to rerun the same analysis and confirm the conclusions weren't confabulated, selectively presented, or simply wrong. The investigation itself was unverifiable, even though it was the thing everyone had to trust.

Self-reported incident analysis is unverifiable by default

When the producer investigates its own incident, the output is a narrative, not a reproducible artifact. Nobody outside can rerun the same model call with the same inputs and get the same answer to check the story holds up.

17,000 events, one uncheckable account

The scale of the incident made independent reconstruction exactly the kind of task that needed to be reproducible. Instead, the analysis stayed locked to the producer's own tooling and their own telling. Read the incident analysis.

Forensic Rail makes the response itself receipt-bonded

Forensic Rail issues a scoped, threshold-signed credential for the incident, and binds every model call under it to deterministic replay: same seed, same temperature zero, same prompt and output digests. Anyone with the credential and the receipts can reproduce it exactly.

Motivating incident: Hive incident analysis: OpenAI and Hugging Face, 17,000 events · Signature standard: NIST FIPS 204 (ML-DSA).

The mechanism: threshold credential plus deterministic replay

Two things have to hold at once. First, the responder has to be operating under a credential that a consortium of independent parties, not the producer alone, threshold-signed into existence, scoped to exactly what the responder is allowed to do. Second, every model call made under that credential has to be pinned to settings that make the output byte-identically reproducible by anyone else running the same replay hint.

ComponentWhat it pinsWhy it's load-bearingWhat breaks without it
consortium_root_kidWhich threshold key issued the credentialNo single party, including the producer, can mint its own credentialSelf-issued credentials, no independent oversight
threshold_k / total_nHow many consortium signers had to agreeA minority can't authorize a scope the majority didn't approveOne captured signer could authorize anything
scope: analyze_onlyWhether the responder may generate novel output at allPrevents an investigation from quietly becoming an unauthorized deploymentScope creep, undetected
seed, temperature_bp = 0The exact sampling settings usedNonzero temperature makes byte-identical reproduction impossible"Reproduction" that never actually matches
prompt_digest, output_digest, kv_cache_digestWhat went in, what came out, and the cache stateA verifier can confirm the exact inputs and outputs without re-running anything firstNo way to check the claimed output matches the claimed input
The money shot side-by-side, one screen
Why a written incident report misses this
Producer's written incident report
✓ report is internally consistent
✓ conclusions are plausible on their face

READS FINE

Forensic Rail replay check
✗ independent responder reruns the same prompt_digest under the credential
✗ output_digest does not match the report's claimed output

BROKEN: the claimed output does not reproduce under the bonded credential

A report that reads well is not the same as a report that reproduces. Forensic Rail forces the second test, and it's the one that actually catches confabulation or a selectively edited account.

The receipt schema

Two receipt types. forensic.credential establishes the scoped, threshold-signed consortium authorization for the incident. forensic.analysis binds every model call made under that credential to the deterministic settings and digests needed for byte-identical replay. Both are ML-DSA-65 signed under the shared usap.v1 envelope.

// forensic.credential: threshold-signed by the consortium, scoped and time-bound
{
  "usap_version": "usap.v1",
  "receipt_type": "forensic.credential",
  "incident_case_id": "case:openai-hf-2026-07-25",
  "ticket_chain_ref": "sha256:ticket-chain-root...",
  "scope": "analyze_only",                 ← analyze_only | analyze_and_generate
  "no_execute": true,
  "no_generate_novel": true,
  "validity_window": {
    "not_before": "2026-07-26T00:00:00Z",
    "not_after": "2026-08-02T00:00:00Z"
  },
  "consortium_root_kid": "kid:isac-consortium-root",
  "threshold_k": 4,
  "total_n": 9,
  "consortium_signers": ["kid:signer-01","kid:signer-05","kid:signer-06","kid:signer-09"],
  "signers_digest": "sha256:signer-set-digest...",
  "signature": "ML-DSA-65, 3309 bytes"
}

// forensic.analysis: one per model call made under the credential, deterministic-replay-bonded
{
  "usap_version": "usap.v1",
  "receipt_type": "forensic.analysis",
  "credential_ref": "sha256:credential-receipt-digest...",
  "model_ref": "model:reference-analyzer-v3",  // deterministic mode
  "seed": 42,
  "temperature_bp": 0,                     ← must be 0, no exceptions
  "top_p_bp": 10000,
  "prompt_digest": "sha256:8e21...",
  "output_digest": "sha256:3fa0...",
  "kv_cache_digest": "sha256:c910...",
  "responder_kid": "kid:responder-lab-04",
  "replay_hint": { "library": "vllm==0.9.2", "kernel": "5.15.0-cc" },
  "deterministic_replay": true,
  "signature": "ML-DSA-65, 3309 bytes"
}
Zero temperature is non-negotiable

temperature_bp must equal 0 in every forensic.analysis receipt. Any nonzero sampling temperature makes "reproduce this exactly" a contradiction in terms, so the schema refuses to represent that case as deterministic_replay: true.

The credential scope is enforceable, not advisory

no_execute and no_generate_novel are booleans bound into the signed credential itself. A responder operating under analyze_only can't quietly slide into generating new content and still claim the credential covers it.

Replay hints make reproduction actually possible

Byte-identical output depends on matching library and kernel versions, not just the seed. The replay_hint field says exactly what a third party needs to match to get the same output, honestly, instead of pretending reproduction is trivial.

Forensic Rail does not claim every incident analysis will reproduce trivially on any hardware. It claims that, given the replay_hint, an independent party has everything needed to try, and a mismatch is externally detectable rather than invisible.

How the credentialing flow works, end to end

Forensic Rail sits in front of whatever inference stack an incident responder already uses. It doesn't replace the responder's tooling. It requires a valid, scoped credential before any call counts as bonded, and it binds the deterministic settings needed for someone else to check the output afterward.

The credentialing and replay flow
  • The consortium threshold-signs a scoped credential for the specific incident case.
  • The credential sets scope, no_execute, no_generate_novel, and a validity_window.
  • Every model call under the credential runs in deterministic mode: fixed seed, temperature zero.
  • Bind prompt_digest, output_digest, kv_cache_digest, and the replay_hint into a signed analysis receipt.
  • Any independent party can rerun the replay_hint and compare digests against the signed receipt.
The Hive addition: consortium, not producer, authority

The credential is threshold-signed by an ISAC-style consortium, not minted unilaterally by the producer under investigation. That single change is what makes the resulting analysis trustworthy to a party other than the producer: nobody investigating their own incident can also be the sole authority on what they were allowed to do while investigating it.

Public verifier: standalone, offline, zero network calls to Hive. A regulator, court, or rival producer can check the credential's scope, the threshold signature, and whether a claimed output digest actually matches an independent replay.

Integration surface

Sits in front of any inference stack that supports deterministic decoding (fixed seed, temperature zero). Pilot-ready for a live ISAC-style consortium; not yet deployed against a real industry body.

Overhead

Credential checks and digest binding add negligible latency to the analysis call. The deterministic decoding constraint itself may reduce output diversity compared to normal sampling, which is the intended tradeoff for reproducibility.

Fail-closed on scope, fail-closed on nondeterminism

A call outside the credential's scope, or a call that isn't run in deterministic mode, does not get to claim deterministic_replay: true. There's no code path that backfills that guarantee after the fact.

Signature standard: NIST FIPS 204 (ML-DSA) · Reference implementation: hive-typed-signer, one file per primitive.

What we don't claim

Honesty is the product. Here is exactly where the edges are.

Not a guarantee the analysis is correct

Forensic Rail proves the analysis was run under an authorized, scoped credential and that its output is reproducible. It does not prove the conclusions drawn from that output are the right conclusions. Reproducibility and correctness are different claims.

Not automatic determinism

Byte-identical replay depends on the responder's library and kernel versions matching the replay_hint. Different hardware or a different inference engine version can produce a different output even at temperature zero. We say this plainly rather than overselling "reproduce anywhere."

Live integration is pilot-ready, not deployed

The cryptographic core, the threshold-signature scheme, and the digest binding are built and tested. Standing up a real ISAC-style consortium and wiring this to a live deterministic inference stack at a customer is the next step, and it has not happened yet.

Where Forensic Rail sits in the Hive canon

Forensic Rail™ is one of seven upstream primitives filed together, July 26, 2026, USPTO application 64/119,279, all sharing the ML-DSA-65 signer and the usap.v1 envelope. It governs the post-incident reproducibility axis: whether the investigation itself, not just the incident, can be independently checked. The other six cover environment integrity, policy disclosure, intent alarms, network enforcement, time and consensus, and egress metering.

PrimitiveWhat it provesAxisWhen
Forensic RailPost-incident analysis runs under a bonded consortium credential and is deterministic-replay-bonded for independent reproduction.Post-incident reproducibilitycrypto core proven integration pilot-ready
Provenance-Bonded SandboxEnvironment integrity, continuously, from provisioning through every heartbeat.Environment integritycrypto core proven
Refusal LedgerEvery refusal-policy mutation is on record, and the runtime value stays inside declared bounds without disclosing it.Policy disclosurecrypto core proven
HowlerAn alarm about intent formation fired before the effect emitted, and a third party can independently replay it.Intent alarmcrypto core proven
Perimeter BondThe exact enforcement bytecode that intercepted syscalls is bound into every attempt receipt.Network enforcementcrypto core proven
Diurnal BondWeekend and after-hours actions require k-of-n countersigning, where k is computed from a signed risk manifold.Time and consensuscrypto core proven
Egress BondData leaving the perimeter is metered per semantic class against pre-committed caps, without exposing which rows contributed.Egress meteringcrypto core proven
CarnacThe downstream gate: challenge, escalate, or stop an agent action before it executes.Downstream gatenow
HiveSealWhere the signing key lives in silicon: a key-side root.Key-side rootnow
AFiRAttested Fragmented Inference Routing: signed routing of an inference across fragments and providers.Compute axisnow
SiGRA specific model at a specific config produced a specific output, signed and checkable offline.Compute axisnow

See also: the OpenAI and Hugging Face incident analysis, the case where an unreproducible self-account motivates Forensic Rail directly.

Pilot Forensic Rail on a real incident response

We're looking for design partners, including consortium bodies and independent responders, to take the crypto core into a live post-incident investigation. You bring the consortium and the incident. We wire the threshold credential, the deterministic replay binding, and the digest verification, run the smoke tests against your response workflow, and you walk out with receipts a regulator, court, or another producer can check offline.

Forensic Rail: bonded, reproducible incident response. Patent Pending. Binds post-incident analysis to a threshold-signed consortium credential and deterministic replay, signed independently with ML-DSA-65 (FIPS 204), checkable offline. Crypto core built and proven: 37 of 37 smoke tests pass. Filed July 26, 2026, USPTO application 64/119,279. Real consortium and deterministic-stack integrations are pilot-ready, not yet deployed. We never fake a receipt.

Private by design. Hive does not store your prompts. Every request is already receipted by a one-way SHA-256 fingerprint, not the words. Proof, not surveillance.