{ "round": "two-frontiers-cove-r1", "gate": "G-TOOLS", "verdict": "PASS", "detail": "the tool channel is closed on every arm, read off each call's own record", "started_utc": "2026-09-06T13:43:59Z", "finished_utc": "2026-09-06T13:43:59Z", "evidence": { "calls_of_its_own": 0, "why_no_call": "the CLI transport asserts the whole G-TOOLS set on EVERY call and stops the arm on a hit, and an HTTP transport's reading is about the body it sent — so §5's gate budget spends nothing here and this gate reads the G-EFFORT records.", "hooks": "CLOSED BY A LAUNCH FLAG AND RECEIPTED BY NOTHING: the vendor's init record exposes no hooks key, so this is stated on the row rather than inferred from the other assertions.", "session_persistence": "closed by the same launch flag; the sealed child runs in an empty scratch cwd, and the seal is a CHANNEL seal rather than a filesystem one — the child inherits a real home directory because the vendor CLI's only credential lives there.", "problems": [] }, "arms": { "cli-claude-fable-5-1": { "verdict": "PASS", "detail": "init.model; init.tools == []; init.mcp_servers == []; init.slash_commands == []; init.apiKeySource == 'none' (the OAuth/subscription value); no content block of ANY tool-shaped type, anywhere in the stream (census); no unknown content block type (fail-closed); result.usage.server_tool_use counters all zero; result.permission_denials == []; the scratch cwd is still empty after the call", "transport_class": "agent-harness-cli", "law_record": { "seal": "agent-harness-cli", "argv_carries_the_prompt": false, "prompt_channel": "stdin (bytes)", "prompt_bytes": 75, "system_bytes": 135, "system_bytes_limit": 16384, "argv_element_wall_bytes": 131072, "system_channel": "argv (--system-prompt). Refused over 16,384 bytes: a single argv element dies at 131,072 on this box, measured, and execve raises before a process exists.", "effort": "high", "effort_channel": "argv only (--effort); no CLAUDE_* env var reaches the child", "sampler": "not settable on this transport (disclosed, never approximated)", "output_cap": "none settable", "g_id": { "asserted": [ "no `fallback` content block in the stream", "no `system/model_refusal_fallback` record in the stream", "every assistant `message.model` is the arm's registered model" ], "registered_model": "claude-fable-5-1", "served_by_model": [ "claude-fable-5-1" ], "unregistered_models": [], "fallback": null, "fallback_blocks": [], "model_refusal_fallbacks": [], "findings": [], "verdict": "PASS", "why": "identity is receipted per CALL, and a failed reading files THIS cell without stopping the arm (PREREG §3, A7): one swapped cell does not un-vouch its neighbours." }, "g_tools": { "asserted": [ "init.model", "init.tools == []", "init.mcp_servers == []", "init.slash_commands == []", "init.apiKeySource == 'none' (the OAuth/subscription value)", "no content block of ANY tool-shaped type, anywhere in the stream (census)", "no unknown content block type (fail-closed)", "result.usage.server_tool_use counters all zero", "result.permission_denials == []", "the scratch cwd is still empty after the call" ], "breaches": [], "unverifiable": [], "failures": [], "verdict": "PASS" }, "stream": { "records": 7, "unparsed_lines": 0, "block_type_census": { "thinking": 1, "text": 1 }, "known_block_types": [ "text", "thinking", "redacted_thinking", "fallback" ], "tool_shaped_block_types": [], "unknown_block_types": [], "string_content_records": 0, "unshaped_blocks": 0, "assistant_model_census": { "claude-fable-5-1": 2 }, "fallback_blocks": 0, "model_refusal_fallback_records": 0, "tool_use_blocks": 0, "server_tool_use": { "web_search_requests": 0, "web_fetch_requests": 0 }, "result_subtype": "success", "result_is_error": false, "stop_reason": "end_turn", "num_turns": 1, "api_key_source": "none", "rate_limit_status": "allowed_warning", "rate_limit_status_recognised": true, "rate_limit_statuses_known": [ "allowed", "allowed_warning", "rejected" ] } } }, "openai-gpt-6-astra": { "verdict": "PASS", "detail": "no tool field of any kind was on the body sent, read off the body rather than remembered", "transport_class": "openai-api", "law_record": { "think": "n/a (no such field on this API; the model's own default posture governs)", "response_format_sent": false, "schema_mode": "prompted", "num_ctx": "server", "sampler": "server defaults (no sampler field is sent; read off the body sent)", "sampler_fields_sent": [], "tool_fields_sent": [], "reasoning_effort": "high", "reasoning_effort_sent": true, "stream": false, "endpoint": "/v1/chat/completions", "deltas": { "num_ctx": "server", "sampler": "server defaults (no sampler field is sent; read off the body sent)", "sampler_fields_sent": [], "reasoning_effort": "high", "reasoning_effort_note": "the vendor's own effort ladder, pinned and recorded -- not a sampler, and not an equivalence with any other vendor's ladder. PLAN §3 registers it on both hosted arms.", "latency_class": "client-wall only: this endpoint reports no eval_duration, so there is no decomposition to print beside the wall clock, and tok_s_eval is EMPTY -- never zero." } } }, "cloud-glm-5-3": { "verdict": "PASS", "detail": "no tool field of any kind was on the body sent, read off the body rather than remembered", "transport_class": "ollama-cloud", "law_record": { "think": false, "format_sent": false, "keep_alive_sent": false, "num_ctx": "server", "stream": false, "options": {}, "dropped_by_cloud_law": {}, "schema_mode": "prompted", "tool_fields_sent": [], "asymmetry_note": "the shelf drops `format` and `num_ctx` (cloud law 2), so this arm answers in prompted mode while this leg's local arm answers under an ENFORCED schema. A registered delta on the row, not a footnote." } }, "local-gemma4-26b": { "verdict": "PASS", "detail": "no tool field of any kind was on the body sent, read off the body rather than remembered", "transport_class": "local-ollama", "law_record": { "seat": "local-ollama", "think": false, "format_sent": true, "schema_mode": "enforced", "sampler": "num_ctx=32768", "options_sent": { "num_ctx": 32768 }, "tool_fields_sent": [], "num_ctx": 32768, "output_cap": null, "output_cap_note": "the local arm is capped and both hosted arms are not; PLAN §3 registers that asymmetry as its own delta, and G6b prints 'n/a - no cap settable' for the hosted rows.", "stream": false, "endpoint": "/api/chat", "production_port": true, "production_port_amendment": "A2", "latency_published": false, "latency_note": "EMPTY under amendment A2: a timing taken off a contended production instance is not a serving figure, and this round publishes no latency it cannot stand behind.", "credential": "none: this transport has no key channel at all", "format_value": "json", "schema_registered": "json", "schema_authority": "COVE-A1", "schema_law": "the local arm answers under the schema its ROSTER registered (D-20260906-30 R3: the open call's exact house posture, format:'json' ACTUALLY ENFORCED). This is a CONTRACT ADVANTAGE over the three prompted-mode arms of this leg -- August disclosed it and this leg re-discloses it verbatim. two-frontiers-r1's own law 6 refuses a format field outright, for its own round's reason; that file is frozen by its freeze index and is not edited from here.", "posture_registered": { "think": false, "format": "json", "num_ctx": 32768 }, "house_law": "harness/cove/local_arm.py -- this leg's own local law: two-frontiers-r1's law 6 refuses a format field, and its file is frozen by that round's freeze index." } } }, "records": [], "withheld": { "rule": "a receipt publishes in the kit, so these fields are removed here rather than at kit time — and named, because a field quietly absent from a receipt is indistinguishable from a gate that never read it." }, "note": "written by harness/cove/gates_cove.py. Every generation call rides the same send() the scored run uses; a gate that hand-rolled its own call would receipt a field the run never sent.", "supersedes": { "earlier_runs": [ { "receipt": "g-tools-cove.json", "verdict": "PASS", "finished_utc": "2026-09-06T13:30:58Z", "detail": "the tool channel is closed on every arm, read off each call's own record" }, { "receipt": "g-tools-cove.2.json", "verdict": "FAIL", "finished_utc": "2026-09-06T13:41:39Z", "detail": "cloud-glm-5-3: the ollama-cloud law record carries no `tool_fields_sent` key, so this gate cann" }, { "receipt": "g-tools-cove.3.json", "verdict": "PASS", "finished_utc": "2026-09-06T13:43:01Z", "detail": "the tool channel is closed on every arm, read off each call's own record" } ], "rule": "this gate ran before and its earlier receipt(s) are KEPT in place. This file is the later reading; an earlier FAIL is not erased by it, and a reader can see both. The freeze index pins whichever receipt is the gate of record." } }