title: Briefing — ZKP exploration programme: findings and runnable evidence
date: 2026-07-18
audience: DTG ZKP Task Force (with seams noted for the Credentials, Trust Task, Registry/Governance, and Human Experience workstreams)
from: Mitch (co-chair, Soulbis)
status: DRAFT — local working document, prepared for co-chair review before any circulation; nothing in it has been submitted upstream

Briefing — ZKP exploration programme: findings and runnable evidence

1. What this reports

Since the Credentials Core Specification (WD v1.0) and the Predicate & Assurance-Boundary Decision
Document (v0.1.0-draft) landed, a local research programme has been run against them: nineteen design
explorations, an ecosystem cross-check against Semaphore v4, and sixteen runnable reference suites
totalling 162 property tests, all passing — including the programme's first real circuits (groth16,
end-to-end) and a second-language consumer that proves the conformance fixtures are
implementation-independent. Every
reference is zero-dependency, offline, and deterministic. Nothing here is a product or a proposed
normative text; it is evidence prepared for the group's decisions — each item traceable to a section
of the decision document, built in the order that document requires: statement first, construction
after (§2.1, §25).

The framing claim the programme tests: the Credentials Core Spec deliberately deferred its ZK protocol
details and registry-ZK interactions; DTG ZKP V1.0 can be written as the conformant binding for the
spec's two named constructions — the missing layer, not a parallel scheme.

2. Evidence highlights

Predicates (the strawman §4 seed). A nullifier reference (9/9 properties) whose decisive test is
the user-as-adversary self-Sybil case, and a trust-graph-formation reference (11/11) contributing a
candidate edge-admissibility rule for the deferred registry-ZK interaction. Both use the narrow
language throughout: a nullifier establishes scoped reuse detection, never "one unique human."
The formation model now also carries its witness seat (15/15): a personhood-anchored third party
attests that a collision occurred — transcript-bound, never seeing the encounter's content, and
structurally unable to mint or unlock an edge (the separation property the VWC pattern requires).

Canonical encodings. The §6.2 context descriptor and §15.2 canonical transcript exist as one shared
runnable encoding (11/11), where "a bare nonce is insufficient" is literally a failing test, unknown
fields are rejected (no covert extension channels), and the requested-predicate set is digest-relevant.

Conformance fixtures (§26). The largest single result: a fixture format — canonical vectors in
three classes (accept / reject / lint), a versioned rejection-reason register (now v2: 66 exact
codes + 29 parameterized families, append-only, every runtime-emitted code absorbed byte-exactly and
triggered live from its emitting module), extraction that freezes the reference suites as data, and a
consumer that re-runs every vector demanding the same outcome and the same named reason. §26.1
coverage is 11 of 11 bullets, no gaps (two bullets via minimal semantic models, honestly flagged).
The lint class is worth the group's attention: the register's "verifier must not infer" column cannot
be tested cryptographically, but it can be tested as claim-language linting of verifier outputs —
§26.1 already mandates one such rejection, so the class has a normative foothold.

The fixtures are now proven implementation-independent. A second consumer, written in Python with
standard library only and zero shared code (it parses the reason register out of the JavaScript as
text, never executing it), consumes all 39 vectors with matching outcomes and byte-identical reason
codes, and re-derives all 29 embedded canonical digests byte-exactly across languages. Zero
divergences found. This is §26's "independent implementations instantiate the decisions consistently"
demonstrated, and the strongest exhibit for adopting the fixture format.

Context legibility (§6.8). The six-question human-legibility test is runnable as a context card
generated from the canonical descriptor itself — the §26 evidence requirement ("user-facing
description matches the cryptographic configuration") becomes a derivation, not an audit; silent
context expansion (§6.6) becomes structurally impossible to hide because the card version is the
descriptor digest. (13/13; seam to the Human Experience workstream, §27.4.)

Composition (§2.6, §27.2). A show — several predicates in one presentation — is modelled as one
atomic canonical transcript: partial acceptance, member transplant across transcripts, and
à-la-carte predicate requests are all named rejections (10/10). Two findings for the group: a show's
joint disclosure is strictly wider than the union of per-predicate boundaries (intersection
narrowing, request-pattern leakage, bundle-shape correlation), and the constructive remedy is
governed bundle profiles — named, versioned predicate bundles, as cipher suites replaced
à-la-carte negotiation. (Seam to the Trust Task workstream; the credential/artifact wall holds: a show
never yields completion evidence.)

Observable events (§20). The ruling "unlinkability, not undetectability" made testable: a
per-observer leakage budget (all §19 observers, in §2.4 three-parameter form), a "quiet
presentation" conformance tier (status-schedule independence from presentations, one uniform
external error surface with reason codes holder-routed, profile-constant proof shape, on-grid
retries), and a log-field register with a prohibited list implementing §6.6's telemetry rule
(8/8). Honest deferrals stated: distribution-level indistinguishability needs real traffic, not a
reference model.

Recovery and rotation (§29). The open research item "biometric-derived secret rotation" is
answered as a decidable profile choice (10/10): routine epoch rotation is holder-side key
derivation and never touches enrolment; only catastrophic loss re-enters enrolment, via issuer-blind
replacement with a structural audit that no old→new pair survives in issuer state, rate-limited by a
recovery-domain nullifier (§13.4's input, given semantics). The trust-graph consequence is worked:
edge re-formation vs bridged continuity, with the bridge visible only inside the edge's own context
(§6.5/§6.6).

Lifecycle (§22). The ten §22.1 clocks sort into two families: certification clocks (an
authority sets and can renew) and erosion clocks (nobody renews; only monitoring plus re-base). An
erosion-aware residual-risk record extends Appendix B so "for how long" can be stated as a rate
rather than a cliff — with overclaiming structurally banned (a claimed measured erosion horizon is a
validation failure; only declared/estimated are admissible) (8/8). Offered strictly as an
informative lens, never normative.

Mediated proving (§21). The §21.2 control checklist as mechanism (9/9): a three-level mediator
taxonomy where the prohibited honeypot is unbuildable at the constructor, session-isolation and
non-retention as state-diff property tests against a frozen audit-field list, per-context
mediator-facing pseudonyms closing the §6.6 correlation-hub risk, and the §21.3 downgrade rule as a
four-exit state machine in which a silent transition cannot be constructed — even a de-listed
mediator fails closed visibly.

Multi-issuer aggregation (§12.5). The non-normative multiplicative-confidence mechanism made
executable (10/10): the ∏ε bound with its independence assumption enforced through a registry of
declared dependency classes — issuers sharing a vendor, pipeline, parent, or jurisdiction collapse
to one effective factor, which always weakens the bound; the verified statement reports the collapsed
effective k, never the raw count; k itself is governed by a small tier vocabulary because k is a
disclosure. Honest scoping: aggregation multiplies assurance for personhood, not for liveness.

Guardian recovery (§29, social leg). The recovery pattern that never re-touches the biometric and
survives issuer death (12/12): t-of-n attestations over a committed, personhood-gated guardian set
(a fabricated guardian cannot join — the Sybil-self-guardian attack dies at commit time), seat
nullifiers per recovery context and epoch, guardianship that expires and is re-affirmed (§22.2), a
contested recovery that freezes re-issuance with issuer state provably untouched, and the rule that
attestations are never completion evidence — the issuer's re-issuance is the outcome artifact. The
correlator trade is stated in full: resistance to issuer and verifier is bought at the price of t
colluding guardians and a social-graph disclosure, both named.

Circuits (the deferred layer, in constraints). The nullifier-membership gadget is real: Poseidon
commitment, depth-20 Merkle inclusion, domain-tagged nullifier, groth16 over BN254, proofs verifying
end-to-end. 11,523 constraints · ~630 ms proving on a consumer desktop · ~8 ms verification · 725-byte
proof (including the one-constraint in-circuit transcript binding added after the cross-check —
proofs are now cryptographically single-presentation while the nullifier's per-context stability is
preserved and tested) — the first concrete datapoints for the performance envelope (§25) and the strawman's
claim-ceiling question. The hash-pinning requirement is made concrete (domain tag =
sha256('dtg-zkp/nullifier/v0') mod p, derivation documented), and the circuit's context input is the
§6.2 canonical descriptor digest — the same bytes the legibility card renders and the fixtures
serialize. A sibling circuit realizes the k=2 multi-issuer tier: two memberships under one root plus
in-circuit nullifier distinctness at a cost of exactly 2× + 1 constraint, where a duplicate issuer
cannot even produce a witness. Setup caveat stated plainly: the trusted setup is a local, lab-only
ceremony, not production.

Ecosystem position (Semaphore v4). A source-verified cross-check finds the construction
structurally conformant with Semaphore v4 — same skeleton, same narrow nullifier semantics (their
documentation and ours both describe scoped reuse detection) — but byte-incompatible at every
preimage, each divergence tagged deliberate (the §13.4 version tag, commitment blinding, governed
context vs free scope) or flagged. The one accidental gap it found — no in-circuit transcript binding,
which Semaphore achieves for a single constraint — has been closed in the lab reference (a public
transcript-digest input; the input layout is offered for the group's ratification). Two strategic
questions fall out for the group: whether to define an optional Semaphore-compatibility profile
(their audited circuit and production ceremony, with scope set to our governed descriptor digest), and
whether EdDSA-based identity is worth adopting for holder-binding synergy.

3. Candidate contributions, in discussion order

1. E1–E7 — seven coherence edits interlocking the liveness strawman with the Credentials Core Spec
   (normative reference, per-predicate anchor table, "accredited set" → trust registry, the two
   constructions named, PHC/IDVC framing, taskContext freshness, concrete Credentials-TF alignment).
2. A §26.2 fixture format — the vector schema, the reason register (append-only, versioned, with
   §13.6 retry-ambiguity marked per code), and the lint class, offered for ratification alongside §28.
3. The PHC-realization note — how a VTC's one-membership-per-person governance rule becomes
   enforceable in-context via the nullifier (governance dedupes the enrolment; the proof makes the
   dedup binding), stated in PR-UNQ's narrow language.
4. The recovery profile choice — the routine/catastrophic split plus the recovery-domain nullifier,
   turning §29's open item into options with stated trades.
5. Bundle profiles — joint work with Trust Tasks (§27.2).
6. The context card — joint work with Human Experience (§27.4).
7. The Semaphore position — ratify the deliberate divergences, decide the transcript-binding input
   layout, and choose whether a Semaphore-compatibility profile ships as an optional second profile
   (audited circuits + production ceremony for deployments that want them; the native gadget stays the
   reference for the full §13.4 binding).

4. Open questions surfaced for the group

Beyond the §28 ratification list (which stands), the programme surfaced these:

• Log retention's clock family has two faces: as a set bound it is a certification parameter; as
  correlation accumulation it is erosion. The record format may need both; unresolved by design.
• Bundle-registry authority: the context authority (§6.7), the Trust Task workstream, or a shared
  body? The registries interlock.
• Fixture witness opacity: semantic vectors (construction-neutral) vs construction vectors
  (per proving system) — proposed two-layer answer needs appetite.
• Assurance-class vocabulary for the IDVC boundary object: whose enumeration, with cardinality
  minimised (§18.2)?
• Witness membership scope (VWC seat): same-VTC witnesses only, or any registry-anchored witness?
• Transcript-binding input layout: the lab reference now binds the §15.2 canonical transcript
  digest into the proof (closing the cross-check's one accidental-gap finding); the public-signal
  layout needs ratification before any circuit text is proposed.
• EdDSA identity: Semaphore's signing-capable enrolment secret enables holder-binding and
  off-circuit ownership proofs ours cannot; adopting it (with our blinding retained) is a real
  capability question, not a cosmetic one.

5. Disposition

All work is local. The tracked working tree of the TF clone carries only one pending two-line README
edit; reference code lives in an excluded lab; design documents live in a separate research root.
Anything above moves upstream only as individual PRs/issues after group discussion — beginning, if the
group agrees, with E1–E7 and the fixture-format proposal.