Research · Technical Notes · freely readable

Technical Notes.

Dated, citable, MIT-licensed reference notes. Every claim is independently checkable with the tools we publish.

Verifiable Selective Disclosure and Buyer-Visible Trust Seals for AI Agents

DCS Labs Technical Note · TN-2026-01 Date: June 2026 · Status: Open · MIT-licensed reference implementation Authors: DCS Labs (DCS AI Technologies L.L.C)


Abstract

We describe two layers of the R-Series, an open standard for cryptographically verifiable AI-agent actions: R+4, a zero-knowledge layer that lets an agent prove a statement about a private value (e.g. "balance ≥ threshold", "age ≥ 18", "amount within a range") without revealing the value; and the Trust SKU, a buyer-visible meta-seal that lets a non-technical buyer confirm, in one check, that a vendor runs the full R+1–R+4 stack. We present a working reference implementation of R+4 that requires no trusted setup, using Pedersen commitments with per-bit Chaum-Pedersen OR-proofs under Fiat-Shamir, and a Trust SKU verification scheme that is offline-checkable and resistant to single-key compromise via a required co-signer. Both are released openly; all claims in this note are independently verifiable with the tools we publish.


1. Introduction

Verifiable provenance for AI-agent actions is becoming a procurement requirement: buyers and auditors increasingly need to confirm what an agent did without trusting the vendor's word. The lower R-Series layers (R+1 timestamp, R+2 signed provenance, R+3 audit export) establish a tamper-evident record. This note addresses two harder problems on top of that record:

  1. Selective disclosure (R+4). Some compliance questions require proving a fact about sensitive data — that a balance clears a threshold, that an age qualifies, that a transaction was within a permitted band — without exposing the underlying value. This is the natural domain of zero-knowledge proofs.
  2. Buyer-comprehensible trust (Trust SKU). The cryptographic layers are meaningful to engineers, not to the procurement officer deciding whether to onboard a vendor. The Trust SKU compresses "does this vendor run verifiable AI, provably, right now?" into a single check and a visible badge.

To our knowledge, the second is novel in this category: existing efforts standardize receipt formats, but none provide a buyer-visible seal layered above the format.


2. Methods

2.1 R+4 zero-knowledge range/threshold proofs

The reference implementation proves that a committed value v lies in [0, 2^32) without revealing it, and composes this into threshold and range predicates.

  • Commitments. Pedersen commitments C = vG + rH over the ed25519 prime-order group, with H derived by a nothing-up-my-sleeve procedure so that log_G(H) is not exploitable for forgery.
  • Bit decomposition. v is decomposed into 32 bits; each bit b_i is committed as C_i = b_i G + r_i H.
  • Per-bit OR-proof. For each C_i we run a Chaum-Pedersen OR-proof showing C_i commits to 0 or 1, made non-interactive via Fiat-Shamir. The proof reveals nothing about which value the bit holds.
  • Threshold / range. "value ≥ threshold" is proven by showing value − threshold is in range; "value in [low, high]" by showing value − low is in range. No trusted setup is required for this construction.

The production deployment additionally targets Groth16/Plonk circuits (compiled with circom/snarkjs, finalized by a multi-party trusted-setup ceremony) for the five standard predicates; the verification interface is identical, so the reference and production provers are interchangeable behind one verifier dispatch.

2.2 Trust SKU verification

A Trust SKU seal is a signed JSON document asserting a vendor runs R+1–R+4, with:

  • a 30-day validity window (forcing demonstrated ongoing compliance and bounding key-compromise blast radius),
  • a required co-signature from the DCS Standards Authority (so a stolen vendor key alone cannot forge a valid seal),
  • an optional customer-elected auditor co-signature for a higher trust tier.

Verification is offline: given the seal and the Authority's public key, a buyer confirms expiry, co-signature(s) over the RFC 8785 canonical seal body, and that each standard is active — with no dependency on a live DCS endpoint.


3. Results

The reference implementation is tested and the test suite is public. Observed results:

  • R+4: range and threshold proofs verify; false statements (value below threshold) cannot be proven; tampered proofs are rejected; two proofs of the same value differ (confirming the proofs are randomized and zero-knowledge). 8/8 tests pass.
  • Trust SKU: minted seals verify offline; a tampered seal body, a missing Authority co-signature, an inactive standard, or an expired window all yield the correct negative verdict; the Tier-2 path correctly requires an auditor co-signer. 7/7 tests pass.
  • Cross-layer: R+2 receipts produced by the issuing service verify under both a TypeScript verifier and an independent pure-stdlib Python verifier — evidence of genuine cross-implementation interoperability rather than a single-codebase artifact.

The full R-Series reference suite reports 40 passing tests across the R+2 core, post-quantum signing, conformance vectors, R+3 export, R+4, and Trust SKU.


4. Discussion

The design choices reflect a consistent principle: verifiability should not require trusting us. R+4 proofs are checkable by anyone with the verifier; Trust SKU seals are checkable offline with a public key; R+2 receipts are checkable by an independent Python implementation. This is deliberate — independent verifiability is the credibility mechanism, not our assertion of correctness.

Two honest limitations. First, the no-trusted-setup reference proof is a bit-decomposition range proof; the production Groth16/Plonk path is gated on a multi-party trusted-setup ceremony, which is in progress and whose integrity depends on independent contributors. Second, the Trust SKU's strength rests on the security of the Authority co-signing key; the 30-day rotation bounds, but does not eliminate, the consequences of its compromise — which is why production custody is HSM-based.

For sensitive domains (healthcare, government, finance, and other safety- or rights-affecting verticals), the R-Series provides auditability of actions, and DCS Labs holds those applications in research-mode per its published Research Policy. The standard records and verifies what occurred; it does not make clinical, financial, or other domain decisions.


5. Availability

The specifications are open and MIT-licensed. The reference implementation, conformance vectors, and the independent Python verifier are published so that every claim in this note can be checked. A live verification demo lets anyone validate a receipt in their browser.

Cite as: DCS Labs, "Verifiable Selective Disclosure and Buyer-Visible Trust Seals for AI Agents," DCS Labs Technical Note TN-2026-01, June 2026.


Hash-Chained Action Provenance and a Tamper-Evident Post-Quantum Migration Path

DCS Labs Technical Note · TN-2026-02 Date: June 2026 · Status: Open · MIT-licensed reference implementation Authors: DCS Labs (DCS AI Technologies L.L.C)


Abstract

We describe R+2, the provenance layer of the R-Series: a scheme for producing signed, hash-chained, independently verifiable records of AI-agent actions, and the crypto-agility and post-quantum migration design that lets those records remain verifiable across a change of signature algorithm without breaking the chain or weakening its guarantees. R+2 uses Ed25519 signatures over RFC 8785 (JCS) canonical JSON, with each receipt referencing its predecessor by content identifier to form a per-agent chain. We add an explicit, hybrid-signed algorithm-transition receipt that records a migration from a classical to a post-quantum scheme (ML-DSA-65, FIPS 204) such that the migration is itself an auditable, tamper-evident, downgrade-resistant event. The reference implementation is published and independently verified by a second implementation in a different language.


1. Introduction

An action record is only useful if a third party can confirm it without trusting the recorder. R+2 targets two properties: authenticity (the record was produced by the claimed agent and is unmodified) and continuity (records form an ordered chain that cannot be silently reordered, inserted into, or truncated). A third concern is longevity: a record signed today may need to remain verifiable for decades, which raises the prospect of the signature scheme itself becoming obsolete — most acutely, the eventual arrival of cryptographically relevant quantum computers.

This note covers the R+2 construction and then the migration design that addresses longevity: how an agent moves from Ed25519 to a post-quantum scheme while keeping its entire history verifiable and its migration auditable.


2. Methods

2.1 R+2 receipts and chaining

  • Canonical form. Each receipt is serialized via RFC 8785 (JCS) so that signing and verification operate on identical bytes regardless of serializer. The signature covers the canonical form of the receipt excluding signature fields.
  • Signature. Ed25519 by default (the ed25519-sha256 suite). The receipt carries the agent's public key and is verified against it.
  • Content identifier (CID). A receipt's CID is the SHA-256 of its full canonical form. The CID is stable and collision-resistant.
  • Chaining. Each receipt references the previous receipt's CID. A verifier walks the chain confirming each link, so tampering, reordering, insertion, or truncation breaks verification.

2.2 Crypto-agility

Receipts carry an explicit algorithm identifier (alg) resolved against a registry that marks each suite's status (active / experimental / reserved) and whether it is production-approved. Verification fails closed: unknown or reserved suites are rejected, and an experimental suite is accepted only when explicitly enabled. This makes the signature scheme a first-class, negotiable property rather than a hard-coded assumption.

2.3 Post-quantum migration

For longevity we support ML-DSA-65 (FIPS 204) and a hybrid Ed25519+ML-DSA suite. The migration itself is the careful part:

  • An agent's chain may begin under Ed25519 and later move to a post-quantum scheme. The switch is recorded as an explicit algorithm-transition receipt.
  • The transition receipt is hybrid-signed by both the outgoing and incoming keys: the old key proves the legitimate holder authorized the switch; the new key proves control of the new key by the same party.
  • The receipt records the from/to algorithms and key fingerprints. A verifier confirms both signatures and that the fingerprints are consistent.
  • Downgrade resistance: once a transition to the new scheme is recorded, later receipts reverting to the old scheme without a fresh transition are rejected. An attacker holding only the old (potentially future-broken) key cannot forge a valid transition or a valid post-transition receipt.

Large post-quantum public keys are handled by anchoring a hash of the key on-chain and serving the full key off-chain; a verifier confirms the full key matches its anchor.


3. Results

The reference implementation is public and tested:

  • R+2 conformance: 30/30 vectors pass — legacy and explicit-suite receipts, all signature suites, multi-receipt chains, and seven negative cases (tampered payload, wrong key, unknown algorithm, reserved algorithm, version mismatch, a bad hybrid leg, and a broken chain).
  • Post-quantum identity + transition: 6/6 — key resolution, on-chain hash-anchor verification, a valid classical→PQ transition, rejection of a tampered transition, rejection of a forged new-key leg, and rejection of a mismatched target key.
  • Cross-implementation: receipts produced by the TypeScript implementation verify under an independent pure-stdlib Python verifier — demonstrating the construction is specification-level, not codebase-specific.

The first production R+2 receipts are anchored on a public chain, and an issuing service produces chained receipts that verify end-to-end with both the reference verifier and the in-browser demo.


4. Discussion

The migration design reflects a specific worry: a naive algorithm switch — just changing the signature scheme mid-chain — is indistinguishable from an attack. By making the transition an explicit, doubly-signed, fingerprint-bound event, the migration becomes more auditable than the steady state, not less, and the downgrade rule closes the obvious attack of replaying the old scheme after a break.

Honest limitations. The hybrid suite's guarantee is "valid if either scheme holds," which is the conservative choice during the transition era; a pure-PQ chain is available where a buyer mandates it (e.g. CNSA 2.0), at the cost of larger signatures. Hash-function agility (moving beyond SHA-256) is reserved in the registry but not yet exercised. And as with all such systems, verification establishes provenance — that an action occurred and is unmodified — not the correctness or wisdom of the action itself; sensitive-domain applications remain research-mode per our Research Policy.


5. Availability

The R+2 specification, the algorithm registry, the conformance vectors, the post-quantum reference code, and the independent Python verifier are published and MIT-licensed. Every result above is reproducible with the published tools.

Cite as: DCS Labs, "Hash-Chained Action Provenance and a Tamper-Evident Post-Quantum Migration Path," DCS Labs Technical Note TN-2026-02, June 2026.