{"criteria_id":"error-recoverability/v1","criteria_revision":"1.0.0","family":"error-recoverability","op":"probe","ops_not_offered":[],"corpus":{"id":"error-recoverability-corpus/v1","sha256":"08101caf7d5d32513db5c17acd51e448142c2e011ea1cf3e013bf3bbcd8199f4"},"corpus_pin_definition":"The corpus pin is the SHA-256 of the corpus's canonical serialization (the corpus object excluding its sha256 field), computed and verifiable by the published harness (harness.mjs). Expected outputs inside corpus cases are derived from the criteria text by an independent implementation in gen-corpus.mjs (never the engine's code path), and each case labels its provenance (local stub-server fixture).","price":{"usdc_units_per_call":50000,"label":"$0.05","basis":"per battery run","status":"RULED"},"free_tier":{"per_day":100,"min_interval_ms":5000,"on_exhaustion":"402"},"input_contract":{"body":"One JSON request body: { target_url, tool_name, required_parameter, expected_type, valid_arguments, attestation, caller? (free pre-check only) }.","caps":{"body_bytes":262144,"valid_arguments_chars":8192},"over_cap":"A request body over 262,144 bytes is a free pre-gate refusal naming the cap.","validation":"The whole submission (URL policy, tool declaration, valid arguments, attestation) is validated FREE, before any payment: a malformed or policy-violating submission fails free and never consumes a payment. The battery itself runs only after payment verification."},"target_safety":"Target-URL policy (A1's phase-1 SSRF discipline, extended for the fleet's first active probe — extended, never weakened): the submitted URL must use http or https; must carry no embedded credentials (no userinfo); loopback, private, and link-local targets are refused (localhost; 127.0.0.0/8, 10.0.0.0/8, 172.16.0.0/12, 192.168.0.0/16, 169.254.0.0/16, 0.0.0.0/8, ::1, fe80::/10). Every battery call is a POST to the one attested target URL exactly as submitted — no other URL is constructed, discovered, or followed; redirects are NOT followed (a 3xx response is an abort condition). Per-call timeout 5,000 ms; per-response body cap 65,536 bytes. The battery is published verbatim in BATTERY.md before it is ever run. Malformed URLs, policy violations, and attestation failures are free refusals before any payment. The battery runs only after payment verification (or on a verified-marker resume with no second charge) — unpaid callers cannot trigger a single target call.","attestation_rule":"Ownership attestation (required, checked FREE before any payment; a failure means NO probe runs): the submission's attestation object must carry operator_wallet (a valid Solana pubkey), statement (the published attestation statement, verbatim), target_url (exactly equal to the submission's target_url), battery_id (exactly error-recoverability-battery/v1), consent (exactly true), and attested_at (an ISO-8601 timestamp within the past 7 days and not more than 1 hour in the future). A missing attestation is the free refusal attestation_missing; any field missing or mismatched is the free refusal attestation_malformed naming the field. Phase-1 honesty: the attestation is a statement form inside the submission (the record's own-material pattern), not a cryptographic proof of server control; it is never represented as more than that.","attestation_statement":"I attest that I own or am authorized to operate the MCP server at the target URL named in this submission, and I authorize the published error-recoverability battery (error-recoverability-battery/v1) to be run against it once.","battery_rule":"The battery (error-recoverability-battery/v1, published verbatim in BATTERY.md): exactly 4 JSON-RPC tools/call POSTs to the one attested target URL, sequentially (B1, B2, B3, B4), at least 1,000 ms apart, per-call timeout 5,000 ms, per-response cap 65,536 bytes, redirects never followed. B1 unknown tool: name __rp_no_such_tool__, arguments {}. B2 missing required parameter: the declared tool with valid_arguments minus exactly the required_parameter. B3 wrong type: the declared tool with valid_arguments where exactly the required_parameter is set to the published wrong-type value (the number 12345 when expected_type is string; the string __rp_wrong_type__ otherwise). B4 corrected retry: the declared tool with valid_arguments exactly as submitted. Abort conditions (the battery halts; the item at which the condition was observed and all later items grade NOT-RUN; earlier items keep their grades; the halt is stated in the report): HTTP 429 or 503 from the target; any 3xx response; a response body over the cap; two consecutive transport failures. A single transport failure grades its item NOT-RECOVERABLE (no_response) and the battery continues.","battery":{"id":"error-recoverability-battery/v1","max_calls":4,"min_interval_ms":1000,"per_call_timeout_ms":5000,"per_response_cap_bytes":65536,"redirects":"never followed (a 3xx response is an abort condition)","method":"POST (JSON-RPC 2.0, method tools/call)","items":[{"id":"B1","kind":"unknown_tool","name":"__rp_no_such_tool__","arguments":{}},{"id":"B2","kind":"missing_required","rule":"valid_arguments with exactly required_parameter removed"},{"id":"B3","kind":"wrong_type","rule":"valid_arguments with exactly required_parameter set to the published wrong-type value","wrong_type_value_for_string_expected":12345,"wrong_type_value_otherwise":"__rp_wrong_type__"},{"id":"B4","kind":"corrected_retry","rule":"valid_arguments exactly as submitted"}],"abort_conditions":["http_429","http_503","redirect_3xx","body_over_cap","two_consecutive_transport_failures"]},"rubric_rule":"Recoverability rubric (per battery item; factors printed; NO composite score): for the malformed items B1-B3 the factors are responded, channel (tool_result = JSON-RPC result with isError true; protocol_error = JSON-RPC error object; http_error; success; none), error_surfaced, names_offending (the error text names: B1 the unknown tool name; B2/B3 the required parameter), states_expected (B1: the tool is stated unknown or not found; B2: the parameter is stated required; B3: the expected type is stated), and a bounded excerpt (at most 500 characters) of the target's own response. Naming checks search the item's response body text, case-insensitively. Grades, first match wins (B1-B3): (1) NOT-RUN — aborted before or at this item. (2) NOT-RECOVERABLE, no_response. (3) NOT-RECOVERABLE, no_error_surfaced — the malformed call was silently accepted. (4) RECOVERABLE — tool_result channel AND names_offending AND states_expected. (5) PARTIALLY-RECOVERABLE, wrong_channel_but_specific — names AND states in another channel. (6) PARTIALLY-RECOVERABLE, correct_channel_but_underspecified — tool_result channel but naming or expected detail missing. (7) NOT-RECOVERABLE, wrong_channel_and_underspecified. For B4 the factors are responded, channel, retry_succeeded, excerpt; grades: NOT-RUN if aborted; NOT-RECOVERABLE no_response; RECOVERABLE corrected_retry_succeeded (a non-error JSON-RPC result); PARTIALLY-RECOVERABLE success_unconfirmed (HTTP 200 with no JSON-RPC result in the body); NOT-RECOVERABLE corrected_retry_failed.","grades":["RECOVERABLE","PARTIALLY-RECOVERABLE","NOT-RECOVERABLE","NOT-RUN"],"excerpt":{"max_chars":500,"surface":"the delivered report only"},"verdict_grammar":{"line":"ERROR RECOVERABILITY REPORT — recoverable: <R>, partially-recoverable: <P>, not-recoverable: <N>, not-run: <U> — <run-end UTC>","vocabulary":"The line carries per-grade counts (printed factors). There is no composite score, grade, or band anywhere in this op."},"free_precheck":"The free pre-check (POST /precheck) validates the whole submission (URL policy, tool declaration, attestation) and returns assessability ONLY: the common signed fields, whether the submission is in-policy (or the free refusal naming the violation), and the disclosure/honesty/retention lines. It runs NO probe, performs NO target call, and returns no report. Free allowance: 100 per day per caller wallet, minimum interval 5,000 ms, 402 on exhaustion.","custody":{"construction":"In-memory processing within a single request lifetime; target responses persist nowhere except the bounded excerpts inside the delivered report. No hash of a target response appears anywhere.","ledger_closed_fields":["payer wallet","timestamp/day","price (USDC units) with the CAD fair-market-value block (rate, rate date, basis, cad_pending, refund)","payment tx id + Solana fee (lamports)","criteria version","verdict line","per-grade counts (recoverable, partially-recoverable, not-recoverable, not-run)","aborted flag","house flag"],"ledger_note":"The ledger list above is closed: adding a field is a criteria/handling version change. The ledger NEVER carries the target URL, any target response, any excerpt, or any hash of a response.","errors":"Error handling is closed-form: static strings plus numeric facts only (lengths against caps, rule ids, field names from the published submission contract). No error path interpolates any part of a target response."},"forbidden":"Forbidden in every surface of this op (report, pre-check, errors, ledger, metrics): any composite score or overall grade; any target other than the one attested URL; any call beyond the 4 published battery calls; any conformance verdict; any target response content beyond the per-item bounded excerpts inside the delivered report.","boundary_statement":"Boundary: this service is NOT third-party testing. The only permissible target is the submitting operator's OWN MCP server, under the ownership attestation in the submission; the barred third-party probing shapes are outside this service entirely. This service grades RECOVERABILITY (Eval #3 E6): given an error, can the caller self-correct from it — does it name the offending parameter, state the expected type/format, and does a corrected retry then succeed. Channel CONFORMANCE (Eval #15's half of the boundary: does the error arrive in the channel the spec prescribes) is not graded as a verdict here; the channel is one rubric factor, because the record shows a model can only self-correct from the tool-result channel. No report claims spec conformance.","disclosure_sentence":"Submitting a target URL to any third party, this service included, discloses that you are probing that server; the battery's calls are visible to the target operator in their own logs, and the attestation in your submission is the authorization those calls rely on.","honesty_line":"The report grades the target's responses exactly as served at probe time to this service's calls; a server can behave differently for other callers or at other times, and a RECOVERABLE grade states that the error for that battery item carried the published recoverability factors at that moment — it is not a guarantee about any other call, tool, or time.","retention":"Target responses are processed in memory only. Bounded excerpts (at most 500 characters per item) live ONLY inside the delivered report. Nothing else about a target response is stored — not in KV, not in logs, not in the payments ledger, not in metrics, and not in error records; no response body, fragment beyond the report excerpts, or hash of a response is recorded anywhere. The payments ledger records only the closed field list published in the criteria. Because nothing is kept, a paid retry re-submits the full submission.","payment_includes":"$0.05 per battery run (price status: RULED — labelled RULED (price ruled 2026-11-03)). The payment includes: one run of the published error-recoverability battery (error-recoverability-battery/v1, exactly 4 calls: unknown tool, missing required parameter, wrong type, corrected retry) against the one target URL you attested you own or are authorized to operate, and one signed report grading each battery item against the published rubric of error-recoverability/v1 (channel used; whether the error names the offending parameter and the expected type/format; whether the corrected retry succeeds), with every factor printed and a bounded excerpt (at most 500 characters) of your own server's response for each item quoted as evidence inside the report. No composite score is computed. A free pre-check validates the submission, the target-URL policy, and the ownership attestation only — it runs NO probe and returns no report. Target responses are processed in memory only and are never stored beyond the excerpts inside your report; the payments ledger records only the payer, the time, the price, the per-grade counts, and the report line.","kill_line":"Kill line (M-50 / Q-31, under the operator's 2026-11-03 build-bank-hold doctrine): build, bank, and hold; judged at the fleet's periodic reviews. There is NO demand kill line for this item. Day-0 = first outside paid settle (accounting only; house proofs never count). Tripwires, all stop-and-review: (1) refunds >5% in any 30-day window; (2) free use >10x paid use; (3) a wrong verdict from undisclosed stale or partial data (the error enters the corpus first as a regression case, corrected rules publish as a new criteria version, the corpus re-passes at 100% before serving resumes; issued reports are never amended); (4) ANY probe-safety incident on a target server = stop-and-review; (5) any battery call outside the published battery, or to any URL other than the one attested target.","never":"This op never probes a third-party server; never runs without a complete, verbatim ownership attestation; never follows a redirect; never sends credentials; never makes a call outside the published battery; never stores a target response beyond the report's bounded excerpts; never computes a composite score.","service":"recoverability-probe","version":"1.0.0","endpoints":{"criteria":"GET /criteria (free)","precheck":"POST /precheck (free, submission + attestation validation and assessability only: NO probe, no target call, no report)","probe":"POST /probe (paid, $0.05 per battery run — RULED: signed per-item error-recoverability report)","metrics":"GET /metrics","not_offered":"Any other POST path answers NOT COVERED — this worker serves op probe only."}}