arborist/docs/zk-wire-protocol.md
russell@unturf.com 0154191bd4
#000016 parked: ZK frontier-proof bench plan + wire protocol
Two arborist-side artifacts close the v7 § 16.1 hand-wave without
violating the language constraint (arborist stays pure-Python; ZK
toolchain lives in sibling repo arborist-zk-bench).

docs/zk-frontier-bench.md — bench plan + parked verdict
========================================================

Bench plan covers Plonky3 (or equivalent) circuit at three sizes
(256/1024/4096 affine), measurement targets (prover ms, proof
bytes, verify ms, peak memory), platform pair (Apple M3 + Linux
x86), acceptance thresholds (≤30s prover at 4096; ≤100KB proof;
≤100ms verify; ≤16GB peak).

Preliminary projection from published Plonky3/Halo2 numbers:

  size 256:   0.5–2 s prove  / 30–80 KB / 5–30 ms verify
  size 1024:  5–30 s prove   / 50–120 KB / 10–50 ms verify
  size 4096:  60–600 s prove / 100–300 KB / 30–200 ms verify
                              ↑ likely UNAFFORDABLE at frontier scale

Most-likely outcome: VIABLE at small scales (256/1024 — useful for
distillation models); UNAFFORDABLE at LLaMA-class hidden width
(4096).

Decision tree from real measurement:
  VIABLE       → commission ZK as v8 capability;
                 follow-up tickets for arborist [zk] extra,
                 governance_policy.frontier_proof_mode field,
                 hermes-side activation export.
  UNAFFORDABLE → park ZK; v7-Local stays v7's terminal
                 contribution; answer→model binding stays
                 out of scope.

docs/zk-wire-protocol.md — consumer-side schema
================================================

Specifies the wire format arborist consumes WHEN/IF ZK proofs
become available. The sibling repo emits proofs; arborist
validates schema + signature + delegates ZK verification to a
sibling-repo verifier binary. arborist itself NEVER runs the
prover or verifier — that's the language-constraint contract.

  - Artifact: arborist-zk-proof-v1 JSON with schema_version,
    proof_system, circuit_id, frontier_node, public_inputs,
    public_outputs, proof_bytes_b64, verifier_setup_id, issued_at,
    issuer_pubkey_ed25519, issuer_signature_ed25519.
  - Trigger: governance_policy.frontier_proof_mode ∈
    {"reveal" (default), "zk"}.
  - Binding: model_weights_zk_root + frontier_proof_circuit_id
    columns added to providence_cache (schema migration deferred
    to integration ticket).
  - Validation pipeline: schema check → recompute commitment
    sanity → Ed25519 signature check → delegate to sibling
    verifier → bind to audit chain.
  - Threat model: compromised prover, wire-format injection,
    proof replay, issuer-key compromise — each with its own
    mitigation.

Status closure
==============

Ticket flipped from "open · awaiting go/no-go" to "parked ·
bench-plan + wire-protocol landed 2026-05-09 (sibling-repo
measurement pending)." This is honest: arborist's side is done.
Sibling repo `arborist-zk-bench` (not yet built) produces the
real numbers; doc updates roll back here when measurements
arrive.

Result: the v7 § 16.1 hand-wave is replaced with explicit
thresholds + preliminary projection + ready consumer schema.
The hand-wave is closed even though the bench-question is
parked.
2026-05-09 15:05:08 -04:00

281 lines
11 KiB
Markdown

# ZK wire protocol — arborist consumer schema
**Companion to:** `docs/zk-frontier-bench.md` (the bench plan +
parked verdict for #000016).
**Status:** draft v0, 2026-05-09.
This document specifies the **wire format** arborist consumes
when ZK proofs become available. The toolchain producing those
proofs lives in a sibling repo (`arborist-zk-bench`); arborist
itself stays pure-Python and never gains a Rust dependency. The
language constraint from #000016 §1.21 is hard — this protocol
is the bridge.
---
## 1. Scope
**What this protocol covers:**
- Canonical-byte JSON schema for one frontier-proof artifact.
- The minimum binding info arborist needs to fold a proof into
the v9.8 audit chain.
- Versioning + forward-compatibility rules.
- The trigger condition (`governance_policy.frontier_proof_mode`)
for when arborist asks for a proof.
**What this protocol does NOT cover:**
- The prover or verifier itself (sibling repo).
- Plonky3 / Halo2 / Groth16 internals.
- Circuit construction.
- ZK-side privacy analysis.
## 2. Trigger condition
ZK proofs are opt-in via a governance-policy flag:
```python
governance_policy["frontier_proof_mode"] = "zk" # opt-in
governance_policy["frontier_proof_mode"] = "reveal" # default
```
The flag folds into `governance_policy_hash` so the v9.8
8-dim cache_key reflects which mode produced any given record.
Switching modes invalidates prior records on lookup.
When `mode == "zk"`, arborist's QA path (or any caller producing
a frontier-proof binding) calls out to the sibling repo, gets
back a wire-protocol artifact, validates it, and folds the proof
root into the providence record.
When `mode == "reveal"` (default), no ZK call happens; activations
are revealed locally per v7-Local.
## 3. Artifact shape (`proof_v1.json`)
```json
{
"schema_version": "arborist-zk-proof-v1",
"proof_system": "plonky3 | halo2 | groth16 | stark",
"proof_system_version": "<vendor-specific git sha or release tag>",
"circuit_id": "<sha256 of canonical circuit definition>",
"circuit_kind": "affine_preact" ,
"frontier_node": {
"node_kind": "affine_preact",
"size": 4096,
"epsilon_num": 1,
"epsilon_den": 100
},
"public_inputs": {
"weights_commitment_sha256": "<64-hex>",
"activations_commitment_sha256": "<64-hex>",
"subset_mask_sha256": "<64-hex>",
"epsilon_pair": [1, 100]
},
"public_outputs": {
"epsilon_inequality_holds": true
},
"proof_bytes_b64": "<standard-base64 of the raw proof bytes>",
"verifier_setup_id": "<sha256 of the verifier's KZG/CRS or FRI setup>",
"issued_at": "<ISO-8601 UTC>",
"issuer_pubkey_ed25519": "<32-byte hex>",
"issuer_signature_ed25519": "<64-byte hex over canonical-bytes(this object)>"
}
```
### 3.1 Field-by-field
- **`schema_version`** — fixed string `"arborist-zk-proof-v1"`. Arborist
rejects any artifact with an unknown schema_version; v2 onward bumps
the validator.
- **`proof_system`** — vendor name. Determines which deserializer
(in the sibling repo) parses `proof_bytes_b64`.
- **`circuit_id`** — sha256 of the canonical circuit definition. Two
artifacts produced by the same circuit have the same `circuit_id`;
this is what arborist binds to in the audit chain rather than the
raw proof bytes (which vary per witness).
- **`frontier_node`** — describes WHICH node was proved. arborist
matches this against the model snapshot's frontier-node enumeration
to confirm the proof's scope.
- **`public_inputs`** — the inputs the verifier checks. arborist
recomputes `weights_commitment_sha256` from its own copy of the
weight bytes, `activations_commitment_sha256` from the published
activation chain, and confirms equality before consuming the
proof.
- **`public_outputs.epsilon_inequality_holds`** — the boolean the
prover claims is true, which the verifier checks.
- **`proof_bytes_b64`** — the actual proof. Arborist does NOT parse
this; it forwards to the sibling-repo verifier (or a separate
trusted binary).
- **`verifier_setup_id`** — sha256 of the verifier's setup (KZG
CRS / FRI parameters). Pinned per-circuit; rotation is a circuit
bump, which changes `circuit_id`.
- **`issued_at`** — wall-clock UTC, sanity-check only.
- **`issuer_pubkey_ed25519` + `issuer_signature_ed25519`** — the
prover signs the canonical-byte form of this artifact (with the
signature field omitted) so arborist can verify provenance of
the proof artifact independently of the ZK proof itself.
### 3.2 Canonical-byte form
For SHA-256 commitment + signature purposes, the canonical bytes
of this artifact are: JSON with sorted keys, separators=(",",":"),
no whitespace, UTF-8 encoded, EXCLUDING the
`issuer_signature_ed25519` field. Standard JSON-canonicalization
the rest of arborist uses (`_canonical_json` in `arborist.qa.dag`).
## 4. Binding into the v9.8 audit chain
When arborist consumes a valid proof, it folds the proof's
identity into the providence record:
```sql
ALTER TABLE providence_cache ADD COLUMN model_weights_zk_root TEXT;
ALTER TABLE providence_cache ADD COLUMN frontier_proof_circuit_id TEXT;
```
(Schema migration would land at the integration ticket, not
this doc. Sketched here so the binding shape is explicit.)
The new fields:
- `model_weights_zk_root` — SHA-256 of the artifact's canonical
bytes (excluding signature). Becomes part of the audit-event body
for `providence_query` rows produced under `frontier_proof_mode
= "zk"`.
- `frontier_proof_circuit_id` — the artifact's `circuit_id`, exposed
for cross-record queries ("show me all rows that used circuit X").
`governance_policy_hash` folds in the `frontier_proof_mode` field
so cache_key changes when the mode flips. Records produced under
`reveal` and `zk` for the same question land at distinct cache_keys.
## 5. Validation pipeline (arborist side)
When arborist receives an artifact:
```python
def validate_zk_proof(artifact: dict) -> None:
"""Pure-Python validation. NEVER runs the ZK verifier;
delegates that to the sibling-repo verifier binary."""
# 1. Schema check.
if artifact.get("schema_version") != "arborist-zk-proof-v1":
raise PiStarError("unknown schema_version")
for required in (
"proof_system", "circuit_id", "circuit_kind",
"frontier_node", "public_inputs", "public_outputs",
"proof_bytes_b64", "verifier_setup_id", "issued_at",
"issuer_pubkey_ed25519", "issuer_signature_ed25519",
):
if required not in artifact:
raise PiStarError(f"missing field: {required!r}")
# 2. Recompute commitment sanity.
pub = artifact["public_inputs"]
if not _is_hex_64(pub["weights_commitment_sha256"]):
raise PiStarError("bad weights commitment")
# … similar for activations + subset_mask
# 3. Issuer signature check (Ed25519).
canonical = _canonical_json_no_signature(artifact)
if not _ed25519_verify(
artifact["issuer_pubkey_ed25519"],
canonical,
artifact["issuer_signature_ed25519"],
):
raise PiStarError("issuer signature failed")
# 4. ZK verification — DELEGATED.
# This is where arborist hands off to the sibling-repo verifier
# binary or a trusted external process. Arborist itself does
# NOT run the ZK verifier.
if not _delegate_zk_verify(artifact):
raise PiStarError("zk verifier rejected proof")
# 5. Bind into audit chain.
# Caller adds model_weights_zk_root + frontier_proof_circuit_id
# to the providence row.
```
`_delegate_zk_verify` is the wire-protocol's other side: arborist
either spawns a subprocess running the sibling verifier OR posts
to a known endpoint. Either way, arborist sees only the boolean
verdict.
## 6. Versioning + forward compatibility
- Schema bumps require a new `schema_version` string. v1 → v2
drops a field, adds one, or changes semantics.
- arborist supports a fixed set of schema versions at any time;
unknown versions are rejected loudly.
- The sibling repo's git sha pins which prover/verifier produced
the proof. arborist surfaces this in the audit body so a
future replay knows which sibling-repo version to invoke.
- `circuit_id` is content-addressed; circuit changes always
produce a new `circuit_id`, never reuse.
## 7. Threat model
- **Compromised sibling-repo prover.** A compromised prover
can forge `epsilon_inequality_holds = true` only if it can
forge the proof bytes themselves AND the issuer signature.
Arborist's defenses: Ed25519 signature check on the artifact
(step 3), delegated verifier invocation (step 4). The verifier
is the trust anchor; if it's compromised, the chain breaks
upstream of arborist.
- **Wire-format injection.** Adversary submits malformed JSON
hoping to bypass schema validation. Mitigation: every required
field checked; type checks on commitment hex; signature check
before delegating.
- **Proof replay.** Adversary replays a valid proof for a
different question. Mitigation: `circuit_id` and `frontier_node`
are tied to the model snapshot's frontier enumeration;
arborist binds the proof to a specific (cache_key,
frontier_node) pair before storage. Replay against a different
cache_key fails the binding check.
- **Issuer-key compromise.** Adversary obtains the prover's
signing key. Mitigation: keys rotate per major-version;
arborist maintains a published list of trusted issuer keys,
rejects artifacts signed under revoked keys.
## 8. What arborist gains over time
If ZK turns out VIABLE per the bench (`docs/zk-frontier-bench.md`),
arborist gains:
- An `[zk]` extra in `pyproject.toml` pulling JSON-schema
validator + Ed25519 verification (NOT the prover/verifier
binary).
- An `arborist/zk/wire.py` module implementing
`validate_zk_proof()` plus `_delegate_zk_verify()` calling
the sibling-repo verifier.
- A schema migration adding `model_weights_zk_root` and
`frontier_proof_circuit_id` columns to `providence_cache`,
plus a CHECK constraint update.
- A new policy field `frontier_proof_mode` folded into
`governance_policy_hash`.
- New audit-event type `providence_zk_proof` for the binding
event.
If ZK stays UNAFFORDABLE, this protocol stays as a documented
artifact — when ZK matures (hardware, libraries, theory), the
arborist side is ready.
## 9. Cross-references
- `docs/zk-frontier-bench.md` — bench plan + parked verdict.
- #000016 — the ticket this closes the arborist-side
hand-wave for.
- #000018 — soft-hash covert-channel analysis; the M3 mitigation
there is the only existing defense if ZK proves UNAFFORDABLE.
- #000013 — v7-W spatial-temporal substrate; world-state
observations have the same privacy concerns and inherit
this protocol.
- v7 § 16.1 — the original "swap SHA-256 → Poseidon" hand-wave.
## 10. Status
Draft v0. Becomes the consumer-side schema spec when the sibling
repo lands measurements + the validator. Until then: this is the
contract arborist commits to honor on the receive side.