DCL Trust Oracle — AI/LLM Output Audit (x402 MCP)

STRICT Pre-Action Audit ($0.05). Runs the response against a broader, higher-bar "strict" policy: the union of all forbidden phrases from the default, anti-jailbreak, and safety policies (8 phrases total), with a 0.85 minimum-confidence threshold instead of the default policy's 0.7. Each matched phrase costs 0.4 confidence; if confidence falls below 0.85, or any phrase matches, the verdict is NO_COMMIT with `reason` listing every match found. Use this instead of dcl_evaluate_fast when the cost of a false COMMIT is high — e.g. before an irreversible or high-stakes agent action — since it catches jailbreak- and safety-adjacent phrasing that the plain default policy would miss.

POST-ACTION Basic Audit ($0.10). Retrieves a record from the tamper-evident chain by tx_hash.

POST-ACTION Deep Forensic Audit ($0.50). Extended output with drift_context and full chain integrity verification.

FINAL-STEP Leibniz Layer Crypto Commit ($0.01). Writes a trading/agent decision to the append-only Leibniz Layer audit chain and returns a Merkle-proof-style receipt: `tx_hash` (proof of this specific commit), `chain_hash` (the previous commit's hash, linking this one into the chain), and `chain_depth` (this commit's position in the chain). Unlike the evaluate_* tools, this call has no pass/fail verdict of its own — it always succeeds and simply seals the decision. Passing `prior_checks` is optional but recommended: it records which earlier pipeline steps (firewall/wallet/trade/MEV) this specific commit is downstream of, in one auditable record. Always run this LAST, after every other crypto-suite check has passed.

PRE-ACTION Bulk Processing ($0.10). Evaluates a list of items in one call; each item is a dict shaped {"response": str, "policy"?: str}, where policy defaults to "default" if omitted and may be any built-in policy name (default, strict, anti_jailbreak, safety, content_quality). Each item gets its own independent COMMIT/NO_COMMIT verdict via the same logic as the matching single-item evaluate_* tool; results are returned in input order under `results`, plus a shared `batch_id`. Capped at 200 items per call — oversized batches are rejected. Use this instead of multiple single-item evaluate_* calls when checking several responses — optionally against different policies — in one priced call rather than paying per item separately.

FAST Pre-Action Audit ($0.01). Runs the response through the server's "default" policy: a substring check against 3 forbidden phrases ("ignore previous instructions", "jailbreak", "bypass safety") with a 0.7 minimum-confidence threshold. Each forbidden match found costs 0.4 confidence; if confidence falls below 0.7, or any match is found, the verdict is NO_COMMIT and `reason` lists which phrase triggered it. Otherwise COMMIT. Use this as the default low-cost first-pass gate before a risky agent action; switch to dcl_evaluate_strict for a broader, higher-bar check, or to dcl_evaluate_jailbreak / dcl_evaluate_safety / dcl_evaluate_quality for a narrower, single-topic check instead of the general-purpose default policy.

PRE-ACTION Instruction Adherence Check ($0.02). Runs the "anti_jailbreak" policy: a substring check against 6 forbidden phrases ("ignore previous instructions", "jailbreak", "bypass safety", "pretend you are", "act as if", "DAN") with a 0.8 minimum-confidence threshold — each match costs 0.4 confidence. Returns COMMIT if no phrase matches and confidence stays at or above 0.8, otherwise NO_COMMIT with `reason` listing the matched phrase(s). Use this as a targeted, cheaper check when the concern is specifically prompt-injection / persona-hijack risk; use dcl_evaluate_strict instead when you also want safety- and default-policy phrases covered in the same call.

PRE-ACTION Crypto Jailbreak & Injection Detection ($0.02). Crypto-specialized instruction-override/jailbreak/injection screen: standard role-switch and instruction-override patterns, plus crypto-specific drain-wallet injection (e.g. "transfer all funds to...", fake "test transaction" requesting full balance) and unlimited-approval injection (e.g. type(uint256).max, "approve unlimited allowance", skip-slippage-confirmation framing). Any match returns NO_COMMIT with `reason` and `findings` naming the matched category/categories; run this FIRST in the DCL crypto pipeline, before wallet/trade/MEV checks, since it screens the input itself rather than a decision built on top of it.

POST-ACTION MEV & Market-Abuse Compliance Screen ($0.03). Text-level screen (not a mempool/transaction analyzer) for front-running/sandwich-attack language, wash trading/layering/spoofing, KYC/AML red flags (mixers, structuring, obscuring fund origin), and pump-and-dump/rug-pull language. Any critical-severity finding, or two or more major-severity findings, returns NO_COMMIT; a single major-severity finding is also returned as NO_COMMIT but with a distinctly higher `confidence` (~0.55 vs ~0.05-0.2 for harder violations) so downstream callers can tell a soft single flag apart from a hard multi-finding block. Each finding includes an illustrative `regulatory_reference` tag (MiFID II, FCA, or an EU AI Act article).

FINAL-GATE Output Sanitizer ($0.02). Post-processing checkpoint that strips secrets/credentials, PII, crypto material (seed phrases, private keys, wallet addresses), internal network details (private IPs, MAC addresses, .internal/.local/.corp hostnames), and unsafe shell/SQL/path-traversal fragments from a raw model response — plus a narrow, high-precision safety net for direct self-harm-instruction-seeking and targeted-harassment phrasing (not a general toxicity classifier). Returns a single `sanitized_output` with every match replaced by `[REDACTED]`; use that instead of the original whenever verdict is NO_COMMIT. Run this as the LAST gate before a response reaches its destination — after `dcl_evaluate_jailbreak_crypto`/other input-side checks have already run, and immediately before `dcl_commit` seals the final decision. Internally re-uses the same detection tables as `dcl_evaluate_secrets`/`dcl_evaluate_pii` for the secrets/PII categories, so results stay consistent with those tools.

POST-ACTION PII Detection Scan ($0.02). Regex-based scan across 8 personal-data categories, with a Luhn checksum on card numbers to reduce false positives. Any finding results in NO_COMMIT.

PRE-ACTION Content Quality & Drift Check ($0.03). Runs the "content_quality" policy: flags 12 absolutist or unverifiable-claim phrases (e.g. "guaranteed returns", "100% accurate", "studies show", "without a doubt") with a 0.85 minimum-confidence threshold — the highest bar of any single-policy tool. Returns NO_COMMIT if any phrase matches or confidence falls below 0.85, with `reason` listing the matched phrase(s). Use this to catch overconfident or unsubstantiated claims in generated content — a different concern from jailbreak or safety phrasing — e.g. before publishing agent-written copy or reports.

PRE-ACTION Baseline Safety Check ($0.01). Runs the "safety" policy: flags 2 forbidden disclaimers ("I cannot be held responsible", "no guarantees") and additionally REQUIRES the substring "AI" to appear somewhere in the response — missing it costs 0.2 confidence even with no forbidden phrase present. Minimum confidence is 0.75. Returns NO_COMMIT if confidence drops below 0.75, with `reason` naming the forbidden phrase found or the missing required pattern. Use this when you specifically need to confirm an AI-disclosure marker is present and the two disclaimer phrases are absent — not as a general-purpose safety net; for broader coverage use dcl_evaluate_fast or dcl_evaluate_strict instead.

POST-ACTION Secret & Credential Leak Scan ($0.02). Regex-based scan across 8 categories (API keys, cloud credentials, tokens/JWTs, private keys, DB URLs, connection strings, env assignments, webhook secrets, internal endpoints with auth). Any finding results in NO_COMMIT.

POST-ACTION Market Signal Fabrication Screen ($0.03). Pattern-based heuristic on the output text alone (no source price feed) — flags guaranteed-price-prediction language ("will definitely hit $X"), absolute-certainty claims ("100% certain", "cannot go down"), a fabricated-price flag when a specific dollar figure co-occurs with a guaranteed-outcome claim, and an invented-token flag when a "$TICKER" cashtag doesn't match a small set of well-known symbols (false positives are possible for legitimate lesser-known tickers — this is a heuristic pre-check, not ground truth). For a full claim-by-claim check against an actual price-feed snapshot, use the local grounding workflow instead of this live tool. Verdict/confidence collapsing follows the same rule as dcl_evaluate_mev: any critical finding or 2+ major findings is a hard NO_COMMIT; exactly one major finding is a softer NO_COMMIT at ~0.55 confidence.

PRE-ACTION Trade Decision Verifier ($0.02). Screens trade-decision language for guaranteed-return claims, zero-risk/"can't lose" framing, and unqualified "buy/sell X now" directives — any match is NO_COMMIT. If no unsafe language is found, COMMIT additionally requires the word "risk" to appear anywhere in the text as a minimum disclosure marker; its absence alone triggers NO_COMMIT with `reason` noting the missing disclosure. Produces an immutable `trade_receipt` (tx_hash/chain_hash/chain_depth) distinct from the top-level audit hash, for downstream systems that specifically need a trade-shaped receipt object.

POST-ACTION Wallet Secret Guardian ($0.02). Scans for BIP-39 seed phrases (12 or 24 consecutive wordlist words), raw hex or WIF-format private keys, Ethereum/Bitcoin wallet addresses, and API keys/bearer tokens appearing near wallet/custody/signing terminology. Any finding results in NO_COMMIT — wallet secrets have no safe threshold, unlike other DCL evaluators. Returns a `sanitized_output` with all matches redacted (null if nothing was found) and a masked `redacted_sample` per finding — the real value is never returned or stored server-side.

SESSION Management ($0.05). Generates a new `pipeline_id` and returns session metadata (scope, expiry, initial drift_mode) for organizing a series of related checks under one identifier. Note: this call does not currently link the returned pipeline_id to later evaluate_* calls — there is no server-side session state that ties subsequent audits back to it; it is an identifier/timestamp issuer, not an active tracking session. Use this to obtain a shared reference ID for your own client-side grouping of a multi-step audit sequence; do not rely on it to automatically aggregate drift across calls.