# Break LA4KB — evidence check for agent answers, signed verdicts (V3) > "Machines may participate. Rules keep the authority." — Las máquinas pueden participar. Las reglas conservan la autoridad. > Your agent proposes an answer; LA4KB checks whether the declared evidence supports it. Send one claim plus up to 10 evidence items; get an Ed25519-signed verdict (SUPPORTED, CONTRADICTED, CONFLICTED, INSUFFICIENT_EVIDENCE, NOT_COMPARABLE…). Deterministic rules, explicit scope, no model in the decision path. LA4KB judges the relationship between the claim and the evidence it receives; whoever supplies a source defines its quality. The signed result is bound to the declared payload hash. The full V3 verifier can recompute that hash from the original admitted request; receipt metadata remains unsigned. Live engine build v3t-1013d7ecf712 (commit caa1b44). Try it, or try to break it. ## Start here - [Full guide for LLMs: input contract, system prompt, curl](https://break-la4kb.eidrax.com/llms-full.txt) - [OpenAPI spec for /api/v3/](https://break-la4kb.eidrax.com/openapi.json) - [Minimal valid example → SUPPORTED](https://break-la4kb.eidrax.com/examples/llm-guide-valid.json) - [Abstention example → INSUFFICIENT_EVIDENCE](https://break-la4kb.eidrax.com/examples/llm-guide-abstain.json) - [AB04 conformance example](https://break-la4kb.eidrax.com/examples/ab04.pretty.json) - [Full V3 request and receipt verifier](https://break-la4kb.eidrax.com/examples/verify_v3_receipt.py) - [V3 canonicalization and five signed vectors](https://break-la4kb.eidrax.com/reference/v3-payload-hash-verification.md) - [Basic signature-only verifier](https://break-la4kb.eidrax.com/examples/verify_receipt.py) - [Breaks & Repairs registry](https://break-la4kb.eidrax.com/data/breaks.json) - [Complete V3 typed contract](https://break-la4kb.eidrax.com/reference/la4kb-v3-typed-contract.txt) ## API (V3) - Base: https://break-la4kb.eidrax.com/api/v3/ (no auth, JSON) - GET health | GET api-info | GET pubkey | POST challenge | GET challenge/{job_id} - POST → 202 {job_id, status: QUEUED}; 422 INVALID_PAYLOAD with errors[].loc; 503 WORKER_UNAVAILABLE if degraded. - Poll GET challenge/{job_id} until status == COMPLETED. - Limits: 5 requests/second per IP, body <= 64 KiB, <= 10 evidence items. ## How an LLM must build the input (summary) 1. Typed values: subject and every scope value are objects, e.g. {"type":"label","value":"svc"}. Never bare strings. 2. value is a JSON number. No unit field: the unit goes in the predicate name (latency_ms), identical in claim and evidence. 3. Citations, one rule. The schema requires claim.source_id + operator_provenance {"type":"operator_declared","operator":,"source_id":} (ASCII [A-Za-z0-9][A-Za-z0-9._-]{0,127}); missing → 422. Recommended for traceability: a source_id on every evidence item with the exact table/page/line (optional in the schema). 4. Use the cited source's authority_tier; never upgrade it. 5. Never invent. No evidence → "evidence_items": [] → INSUFFICIENT_EVIDENCE. If the claim cannot be expressed, do not submit it. 6. Unknown fields reject 422. A 422 is a contract error, never a verdict. 7. An empty `{"dimensions":{}}` without `"is_explicit_global":true` is unknown scope, not global coverage. Matching empty claim/evidence scopes do not by themselves establish support. Declare typed coordinates for a known subset; use explicit global only when the cited source actually covers the global population. One INSTANCE cannot prove a UNIVERSAL claim. 8. Different `measurement_identity` values denote different observations. For an `UNQUANTIFIED` claim, equal-authority, comparable instances that support and refute the same claim can still produce `CONFLICTED`; different identities alone do not cause conflict. A `STATISTICAL` claim is decided by aggregates, not individual instances. 9. An equal-endpoint claim RANGE with an excluded endpoint is an admitted empty claim and yields `INSUFFICIENT_EVIDENCE` against comparable nonempty evidence. Reversed bounds reject 422. ## For reliable loads - The system supplies subjects, predicate names (unit inside the name, e.g. latency_ms) and scope keys; the LLM does not invent them. - The LLM only extracts values, each with the source_id of the exact passage. - When a mapping does not exist, abstain: send nothing, or send "evidence_items": [] (→ INSUFFICIENT_EVIDENCE). - The kernel accepts any ASCII predicate name and compares names exactly. Scope dimensions for la4kb.generic.v1: cluster_size, deployment, enabled, hardware, network_topology, os, region, run, storage, storage_technology, threads, tier, zone. ## Decision example (real, engine v3t-1013d7ecf712) LLM: "all feed lots meet DON <= 0.9 mg/kg" (quantifier UNIVERSAL). Evidence: one lab report for run=lote-1 (0.42). Verdict: INSUFFICIENT_EVIDENCE → the agent abstains from "all lots" and answers for lot 1 only. The claim's empty `dimensions` do not assert global coverage; the single-lot INSTANCE also cannot prove a UNIVERSAL claim. Input: https://break-la4kb.eidrax.com/examples/decision-lotes-input.json · Receipt: https://break-la4kb.eidrax.com/examples/decision-lotes-receipt.json ## Signatures (Ed25519, key from GET pubkey) - result_hash = sha256(canonical JSON {client_reference or "", engine_build, job_id, verdict}) - signature signs "{engine_build}:{job_id}:{result_hash}" - binding_signature signs b"LA4KB-PAYLOAD-RESULT-BINDING-V1" + 0x1F + canonical JSON {version:"la4kb-payload-result-binding-v1", payload_hash, result_hash, engine_build, job_id} - canonical JSON = sorted keys, separators (",",":"), ASCII. The V3 request hash uses the admitted typed request with top-level and provenance defaults; see the build-specific canonicalization and signed vectors. Integer and float JSON values remain distinct. ## Declare yourself Add "provenance": {"declared_origin":"AI","agent_name":"...","model_declared":"..."} and a client_reference. ## What is a break Wrong verdict under the contract; distinct inputs treated as equal (or the reverse); non-determinism; altered signed field that still verifies or genuine signed core that fails; 5xx/hang from input; access to V1/V2 data via V3. Not a break: 422 on ambiguous input; correct verdict on false evidence you declared; documented limits; rate-limit errors. ## Previous versions (read-only archive) V1 and V2 are a read-only archive. Submit/challenge POSTs return HTTP 410. Reads still work.