DJZS Deterministic Pre-Execution Audit for Agents

DJZS audits the reasoning behind an autonomous agent's move before capital is committed, and returns one of three verdicts PASS, WAIT, or FAIL with the specific reasoning defects flagged and a reproducible hash of the decision.

The wedge: transaction-security tools answer "is this transaction safe to sign?". DJZS answers a different question "should this position be taken at all?". A perfectly safe transaction can still rest on broken reasoning; DJZS is the layer that catches the bad thesis.

Use it now — the live MCP gate

verify_pm_trade is deployed as a Model Context Protocol tool (streamable HTTP) on a Cloudflare Worker. Add it to an MCP-capable agent in one line:

claude mcp add --transport http djzs-trust https://mcp.djzs.ai/mcp

Then call verify_pm_trade with a free-text prediction-market trade thesis. It extracts the reasoning, audits it against the calibrated DJZS-M taxonomy, and returns PASS → PROCEED / FAIL / WAIT → HALT, with the flagged defects and a verdict_hash.

First external audit on record. An outside agent audited a benchmark thesis (pm-block-008) through the deployed tool and received:

verdict:       FAIL
flags:         DJZS-M03 (PROBABILITY_UNSOURCED), DJZS-M04 (CONSENSUS_NO_EDGE)
risk_score:    40
disagreements: []
verdict_hash:  0x85918814b3dffa31b00d6892c2e00b2001efd35f7e0044b4cd3789fe1df14937

Behavioral parity against the offline batteries (verdict + flags + extracted input) is green. Hash parity is a pending re-mint the calibration key lapsed during deployment so the live hash above stands as the record until it is re-run.

Pay it over HTTP x402 - the ecosystem rail

Beyond the MCP transport above, verify_pm_trade is now payable over the HTTP x402 transport, live on Base mainnet:

POST https://mcp.djzs.ai/x402/verify

An unpaid request returns an HTTP 402 challenge (exact scheme, eip155:8453, USDC, 2.00 USDC to the treasury); the client signs an EIP-3009 authorization and retries. This is the same rail Base MCP, MetaMask Agent Wallet, and x402-axios speak, so any HTTP x402 client can pay the gate directly - no MCP integration required.

Free-refusal is preserved on both transports: an out-of-scope intent is answered 402 and never settled, so a refused audit costs nothing. Verified end to end on Base Sepolia (paid settlement plus an on-chain zero-delta refusal) before the mainnet deploy.

Architecture C — extraction reports, the engine decides

DJZS separates the model-bound step from the trusted step:

free-text intent
      │
      ▼
EXTRACTION  (LLM, N=3 consensus)          reports observable FACTS, never a verdict
  · per-field state unanimity across the 3 samples; any disagreement → unknown
  · quote gates: a claimed "absent" must quote the intent verbatim, or it demotes to unknown
  · evidence-unanimity: a critical-driving "absent" needs the same quote across all samples
      │
      ▼
AuditInput  (each fact is tri-state: present / absent / unknown)
      │
      ▼
DETERMINISTIC ENGINE  (pure code — no model, no network, no clock)
  rules fire → weighted score → verdict
      │
      ▼
PASS · WAIT · FAIL   +   verdict_hash

Why this shape: the verdict rules are model-independent. The LLM only turns messy text into a structured struct; the audit rules that decide the verdict are frozen code. Swap the model and the extraction quality changes the rules, weights, and hash do not.

Determinism by construction. The same AuditInput always yields the same verdict and the same hash:

verdict_hash = sha256(canonicalize({ verdict, risk_score, flags: codes.sorted(), unknown_fields }))

WAIT is never silent. An under-specified thesis — one where a decision-critical fact is unknown — resolves to WAIT, mapped to HALT for the caller, with unknown_fields returned so the agent knows exactly what to clarify. Abstention is a first-class outcome, never a guessed PASS or FAIL.

Taxonomies (frozen)

DJZS-M — Prediction Market (DJZS-PM-v1.0) the calibrated, live path

Weights sum to 100; FAIL threshold 25. All four codes are implemented and calibrated.

M04 is advisory-grade. On its own it does not block a lone M04 rides a PASS with the flag on the certificate (weight 15 is below the FAIL threshold). It contributes to a block only by stacking with another finding (e.g. M03 + M04 = 40 → FAIL, as in the first audit above).

verify_pm_trade is PM-only: if the intent does not extract as a prediction-market thesis (including the undetermined case), it returns in_scope: false rather than silently running a perpetuals audit.

DJZS-LF — Perpetuals / general reasoning (DJZS-LF-v1.1)

Weights sum to 200. The taxonomy is frozen at 11 codes; 3 are wired live in the engine today (marked ●). The rest are defined and weighted but not yet firing — coverage is stated honestly rather than implied.

Both weight tables are hash-locked (WEIGHTS_HASH / TAXONOMY_HASH, and the PM equivalents, are exported constants); changing a weight is a deliberate re-derive, not a hot patch.

MCP tools

The deployed Worker exposes three tools over streamable HTTP at /mcp:

Honest v1 posture

What the tool does not do yet — recorded deliberately, to be re-ruled before the gate is broadly promoted:

• No ProofOfLogic write on the tool. verify_pm_trade returns a verdict_hash but does not yet anchor a certificate. Existing Irys certificates are prior-architecture lineage, not output of this tool.
• Taxonomy hashes not in the response. The four hash constants are exported from the frozen tables but are not yet included in the tool's JSON response.

The response contract today is: verdict, action, risk_score, flags, unknown_fields, disagreements (the per-field sample-agreement telemetry), verdict_hash, extraction_failsafe, in_scope, and taxonomy versions.

Known divergence: the subgraph's raw failCount counts WAIT as FAIL

query_agent_trust excludes WAIT verdicts from both sides of the fail rate and reports them as wait_count. Anyone reading the subgraph directly, rather than through the tool, gets a different answer. djzs-subgraph/src/trust-score.ts handleScoreUpdated increments failCount on an else branch (verdict == "PASS" / else), so a WAIT — an abstention, not a failure — lands in failCount and inflates the rate for any agent that ever received one.

The Worker recomputes from per-audit verdicts, so the tool is right. The indexed counters are not, and they stay that way until the mapping is corrected and re-indexed. Both numbers are returned side by side (subgraph_counters, labelled) so the discrepancy is visible rather than discovered. Tracked as a follow-up; deliberately not folded into the read-layer fix, because a mapping change needs a subgraph redeploy and a re-index.

Legacy HTTP API (backward-compatible)

Before the MCP gate, DJZS ran as a paid HTTP audit service. Those endpoints remain for backward compatibility the MCP gate above is the current surface. The tiers are metered in USDC on Base Mainnet via an x-payment-proof header:

POST /api/audit aliases the Micro tier, and an escrow-settled variant (POST /api/audit/escrow) takes on-chain escrow in place of a payment header. These tiers run the older detection path, not the Architecture C engine — treat their output as legacy.

On-chain artifacts

The A2A manifest (agent.json) attests four DJZS contracts on Base Mainnet, each marked verified. Source for each lives in contracts/: