consensus-gate

name	consensus-gate
description	Multi-vendor validation gate with direct synthetic.new integration. Routes code and logic through three frontier models (GLM-5.1, Kimi-K2.6, MiniMax-M2.5) in parallel; flags disagreement before merge or evolution. Currently active in production as the Phase 1 gate for zouroboros procedure evolution.
compatibility	Created for Zo Computer
metadata	{"author":"marlandoj.zo.computer","version":"2.1.0","last_updated":"2026-05-11T00:00:00.000Z","api":"synthetic.new (OpenAI-compatible)","models":["hf:zai-org/GLM-5.1","hf:moonshotai/Kimi-K2.6","hf:MiniMaxAI/MiniMax-M2.5"]}

Consensus Gate

Multi-vendor code validation — currently live in zouroboros procedure evolution. Before any evolved procedure step commits to shared-facts.db, three frontier models review it in parallel. Unanimous failure blocks the commit.

v2.1 changes: diff-scoped prompts, structured objection classification (PATCH_SPECIFIC / PRE_EXISTING / OUT_OF_SCOPE), VENDOR_ERROR bucket excluded from signal/noise math, and noise-stats for tracking signal-vs-noise over time. Backward-compatible — verdicts[].issues still emits formatted strings.

---

Quick start

cd /home/workspace/Skills/consensus-gate

# Validate full file/snippet
bun scripts/consensus-gate.ts validate \
  --code "function addUser(name, email) { ... }" \
  --criteria "security,correctness,perf" \
  --label "new-user-step"

# Validate a diff (auto-detected from `diff --git` / `@@` markers)
bun scripts/consensus-gate.ts validate \
  --file /tmp/patch.diff \
  --label "PR #123: ledger fix"

# Force scope explicitly
bun scripts/consensus-gate.ts validate --file foo.ts --scope full --label "foo"
bun scripts/consensus-gate.ts validate --file bar.diff --scope diff --label "bar"

# Noise/signal report (last 7 days)
bun scripts/consensus-gate.ts noise-stats --since 7

# Custom model trio
bun scripts/consensus-gate.ts validate \
  --models "hf:zai-org/GLM-5.1,hf:moonshotai/Kimi-K2.6,hf:deepseek-ai/DeepSeek-V3.2" \
  --code "..." --criteria "security"

# Retrieve a result
bun scripts/consensus-gate.ts result cg-1778362278914-x16fxm

# List recent validations
bun scripts/consensus-gate.ts list --limit 10

---

How it works

1. Input — code (or diff) to validate + criteria (e.g., correctness,security)
2. Scope selection — --scope auto|diff|full (default auto). Auto-detects unified-diff hunk markers (diff --git, @@ -.. +.. @@) or the === DIFF === header used by pre-merge-gate. In diff mode the prompt instructs reviewers to ignore unchanged context and to mark any pre-existing concern as PRE_EXISTING.
3. Parallel fan-out — three models fire simultaneously to api.synthetic.new/openai/v1/chat/completions
4. Collect verdicts — each returns { pass, objections[], confidence } (see schema below). Legacy { issues: [string] } responses still parse.
5. Per-verdict effective pass — a verdict is pass: false ONLY if at least one objection is PATCH_SPECIFIC AND severity: blocker. Verdicts that returned pass: false but contained only PRE_EXISTING / OUT_OF_SCOPE objections are overridden to PASS and tagged overridden: true.
6. Consensus rule (computed from effective pass):
   • All pass → ✅ passed, proceed
   • All fail → ❌ rejected, reject with merged issue list
   • Mix → ⚠️ split, log, proceed (soft gate)
7. Output — JSON result to stdout. Each verdict carries both the structured objections array and the backward-compat issues: string[] (formatted as "[scope/severity] text").

---

Model lineup

Role	Model	Source	Quirk
Coherence + impl quality	hf:zai-org/GLM-5.1	Zhipu/API	Outputs in reasoning_content
Edge cases + security	hf:moonshotai/Kimi-K2.6	Moonshot	Deep context, thorough
Architectural consistency	hf:MiniMaxAI/MiniMax-M2.5	MiniMax	Occasional false caution on trivial code

All three require SYNTHETIC_NEW_API_KEY in Settings > Advanced. No per-call charges — the key is credited.

---

Integration: zouroboros (LIVE)

Wired at zouroboros/packages/memory/src/standalone/memory.ts:1103 in evolveProcedure():

const stepCode = JSON.stringify(validSteps, null, 2);
const validationOutput = execSync(
  `bun "${consensusGatePath}" validate --code ${JSON.stringify(stepCode)} --criteria "correctness,consistency,security" --label "procedure-${procedureName}-v${current.version + 1}"`,
  { encoding: "utf-8", stdio: ["pipe","pipe","pipe"] }
);
const validationResult = JSON.parse(validationOutput);

if (validationResult.consensus.pass === false) {
  const issues = validationResult.verdicts
    .flatMap((v: any) => v.issues || [])
    .join("; ");
  throw new Error(`Consensus gate rejected: ${issues}`);
}

The gate is wrapped in try/catch — script failure (missing, crash, API timeout) logs a warning and proceeds. This ensures zouroboros never hard-blocks on a gate outage.

---

Output schema

{
  id: string;                    // cg-{timestamp}-{random}
  timestamp: string;             // ISO 8601
  label: string;                 // "procedure-x-v3"
  code: string;                  // original input
  criteria: string;              // "correctness,security"
  verdicts: [{
    model: string;               // "hf:zai-org/GLM-5.1"
    pass: boolean;               // EFFECTIVE pass (post noise-suppression)
    modelPass: boolean;          // raw verdict from model
    overridden: boolean;         // true if effective != modelPass
    issues: string[];            // ["[PATCH_SPECIFIC/blocker] missing null-check"]
    objections: [{
      text: string;
      scope: "PATCH_SPECIFIC" | "PRE_EXISTING" | "OUT_OF_SCOPE"
           | "UNCLASSIFIED" | "VENDOR_ERROR";
      severity: "blocker" | "warning" | "info";
    }],
    confidence: number;          // 0.0–1.0 (vendor errors = 0)
    latencyMs: number;           // 4500
    api: "synthetic.new" | "zo-proxy";
  }],
  consensus: {
    unanimous: boolean;
    pass: boolean | null;        // true / false / null (split)
    confidence: number;
  },
  metrics: {
    scope: "diff" | "full";
    detected_diff: boolean;
    total_objections: number;    // includes vendor_errors
    classified_objections: number; // total - vendor_errors; denominator for ratios
    patch_specific: number;
    pre_existing: number;
    out_of_scope: number;
    unclassified: number;        // legacy responses (no scope field)
    vendor_errors: number;       // HTTP/timeout/parse failures — infra, not quality
    blockers_patch_specific: number;
    signal_ratio: number;        // patch_specific / classified_objections
    noise_ratio: number;         // (pre_existing + out_of_scope) / classified_objections
    overrides: number;           // verdicts where effective != modelPass
    would_have_failed: boolean;  // would consensus have rejected without override?
  },
  status: "passed" | "rejected" | "split"
}

Reviewer prompt schema (model output)

{
  "pass": true,
  "objections": [
    {
      "text": "missing null-check on user.email before lookup",
      "scope": "PATCH_SPECIFIC",
      "severity": "blocker"
    }
  ],
  "confidence": 0.85
}

scope is REQUIRED on every objection. pass: false is reserved for blocking PATCH_SPECIFIC objections; PRE_EXISTING and OUT_OF_SCOPE concerns are surfaced but do not block.

VENDOR_ERROR is reserved for the gate itself — emitted when a model call returns HTTP error, times out, or returns unparseable JSON. It is excluded from signal_ratio and noise_ratio denominators so a provider outage cannot trip the noise alarm. Models should never emit VENDOR_ERROR themselves; any model-emitted occurrence is silently downgraded to UNCLASSIFIED.

---

Logging

Database: ~/.zouroboros/consensus-gate.json (JSON array, all results — includes structured metrics)

Log: ~/.zouroboros/consensus-gate.log (JSON Lines, append-only — includes per-run metrics.signal_ratio, noise_ratio, overrides, would_have_failed)

Query from bash:

grep '"status":"rejected"' ~/.zouroboros/consensus-gate.log | tail -5

Monitoring noise vs signal

bun scripts/consensus-gate.ts noise-stats --since 7

Reports:

• signal_ratio (PATCH_SPECIFIC / classified objections) and noise_ratio ((PRE_EXISTING + OUT_OF_SCOPE) / classified objections) over the window. Classified = total minus VENDOR_ERROR, so a provider outage cannot move the needle.
• vendor_error_rate — fraction of verdicts that failed for infrastructure reasons (HTTP, timeout, parse). Watch separately; this is a vendor-health signal, not a gate-quality signal.
• override_rate — fraction of verdicts where the noise-suppression rule flipped a model's pass: false to PASS
• would_have_failed_rate — fraction of runs that would have been REJECTED if we trusted raw model pass instead of the effective rule
• Top-5 noisiest and top-5 highest-signal labels (n≥3 classified objections)
• Trend: noise ratio in the recent half of the window vs the prior half

If noise climbs (e.g. noise_ratio > 45% sustained, or would_have_failed_rate > 20%) the gate is generating more friction than value — that's the signal to wire deferred item #3 (scope precheck + 2-round cap for refactors).

Automated noise watch

bun scripts/noise-watch.ts --since 7

Computes HEALTHY / WARN / CRITICAL / DATA_GAP from the noise-stats payload and emits a one-line verdict. Default thresholds: WARN at noise≥45% or would-fail≥20%, CRITICAL at noise≥65% or would-fail≥40%. Exit codes: 0 healthy, 1 data gap, 2 warn, 3 critical — usable as the command of a scheduled agent that pages on regression.

To wire as a weekly Zo agent (suggested, not auto-created):

schedule: weekly Mon 09:00 America/Phoenix
command: source /root/.zo_secrets && bun /home/workspace/Skills/consensus-gate/scripts/noise-watch.ts --since 7
delivery: email if exit ≥ 2

---

Cost

Path	Cost
Synthetic.new direct (default)	$0 (credited key)
Zo /zo/ask fallback	Consumes Zo credits

Direct path is preferred: lower latency, no credit burn.

---

Failure modes

Scenario	Behavior
API 429 / 500	Individual model FAILs with "HTTP 429", consensus computed from remaining
SYNTHETIC_NEW_API_KEY missing	Falls back to Zo /zo/ask proxy (if ZO_TOKEN available)
Missing script (ENOENT)	Warn + skip, no block
Model returns empty	Verdict pass: false, issues: ["Empty response"]
Model returns malformed JSON	Verdict pass: false, issues: ["JSON parse failed: ..."]
All 3 models fail	Consensus = rejected (unanimous failure to validate)

---

Related

• INTEGRATION.md — zouroboros wiring details
• agent-doctor — diagnose scheduled agents
• Three-stage-eval — pre-execute swarm validation