Every model call gives you one answer. That answer comes from one context, one corpus version, signed by one key. Hive DELTA adds a fifth dimension: cryptographic proof of what the model would have said under a different condition, plus a signed record of the gap between the two answers.
Inference has always had more angles to it than most people use. D5 isn't just a new feature. It changes what a signed inference receipt can actually hold.
One prompt, one response. Linear. No memory of the space traversed.
UniversalCCACW: Compile, Compress, Amplify, Certify, Writeback. It happens in stages, but it's still one line of work.
Live · HiveECHO + MEMBRANE + QUORUM + ATTEST. Parallel primitives on the same input, different lenses on one moment.
Live · XCALIBURSMSH lineage chain. The model's inference knows its own signed history: earlier receipts, earlier corpus commits. T₂ is a function of T₁ and T₀.
Live · SMSHHAHS-Δ. Signed proof of what the model would have produced under a different corpus, rule version, or identity context, plus the delta between the two branches.
Preliminary · DELTAHAHS-Δ runs two inference branches from the same corpus root commitment. Each branch gets its own ML-DSA-65 signed receipt. The delta, meaning the difference between branch A and branch B, is itself a signed record tied back to both parent receipts.
The actual inference. The answer under the committed corpus and current rule version. Ed25519 + ML-DSA-65 signed. HAHS envelope, cert_id anchored to chain.
The counterfactual. Same corpus root, different rule version. What would the model have said yesterday? ML-DSA-65 signed. counterfactual_ref points to Branch A cert_id.
A third artifact is produced: the signed delta between Branch A and Branch B. It carries both parent cert_ids, a SpectralZK proof that both branches diverged from the same corpus root, and the semantic divergence score. The delta is what the regulator, auditor, or CFO holds.
HAHS-Δ builds on the existing HAHS receipt envelope. The schema adds one new field, counterfactual_ref. Every other field stays the same, so any tool that already reads HAHS receipts keeps working.
Fields shown in amber are new to HAHS-Δ. Every other field comes straight from HAHS v1, so existing verifiers can read it with no changes. See the HAHS canon spec.
Every industry running regulated inference has some version of the same question: what would the model have said under different conditions? HAHS-Δ is the first primitive that gives you a signed answer to that question.
What would this contract clause produce under California law vs. Delaware law? Signed delta. Both branches admissible.
What would minimum weight be if the torque sensor reading were X instead of Y? The receipt shows the decision and its counterfactual. The dispute resolves at the verifier.
What would this policy decision produce under EU-AI Act v1 vs. v2? Signed delta for auditors. No spreadsheet. A receipt.
What would this agent payment route produce under rate regime A vs. B? Signed oracle delta for treasury ops and CFO sign-off.
What would the clinical decision support have recommended under the prior protocol version? Signed delta for incident review and HIPAA audit trails.
What would the structural compliance assessment produce under IBC 2021 vs. IBC 2024? Signed delta for permit submission and dispute resolution.
Produces one answer per call. Logs it. Maybe stores it. If you need to know what it would have said under different conditions, you run it again manually, with no cryptographic proof the two runs started from the same state. No signed artifact. No auditable delta.
Produces two signed branches from the same corpus root, plus a third record: the signed delta. SpectralZK proves both branches came from the same committed corpus without revealing what's in it. ML-DSA-65 signs everything. The result holds up in court, reads clearly for a regulator, and is something a CFO can actually keep on file. It's not a log. It's a receipt.
The HAHS-Δ schema is specified. The counterfactual_ref and semantic_divergence fields are defined and forward-compatible with the live HAHS v1 envelope. The ML-DSA-65 signing path (NIST FIPS-204) is live on hivemorph today and applies directly to HAHS-Δ receipts.
The SpectralZK delta proof, the cryptographic proof that two branches came from the same corpus root without revealing what's in it, is still a research problem. It needs a ZK circuit built over the Merkle commitment scheme used in HAHS-Δ. We're actively working on it, but we don't claim it's live yet. When it ships, it moves to Hive Verified with a published benchmark and Cohen’s d computed against a null baseline.
The semantic divergence score (semantic_divergence) is implementable today using the SMSH compression kernel as an embedding layer. The ZK proof that the score was computed correctly over the signed corpus is the hard part.