{
  "protocol_id": "aihackers-sol-cli-adapter-observational-20261030-v2",
  "supersedes_for_execution_only": "protocol.json v1 remains frozen and unexecuted",
  "status": "registered; independent methodological implementation review GO; ready after freeze",
  "authorization": "Parent explicitly authorized transparent observational adapter and continuation through the same 12 cells after fixture qualification and independent review.",
  "claim": "Exploratory matched observations of this separately captured CLI orchestration adapter on two frozen synthesis/arithmetic tasks. Not a native-subagent benchmark, equal-compute experiment, open research test or general model verdict.",
  "unchanged": [
    "Same two task prompts, customer, source packets, source snapshot date and primary URLs",
    "Same randomized 12-cell schedule with three lanes and two repeats",
    "Same requested gpt-6.1-sol/high and forced ChatGPT authentication for every session",
    "Same progressive skill contents, source access and decimal task-tool implementation",
    "Same final answer cap600 words and fixed acceptance rubric/critical-error rules",
    "First attempts preserved; failed/deviating runs retained; no winner-seeking retries",
    "Independent scoring after all first attempts; scorer receives only anonymized final answers and frozen evaluation materials"
  ],
  "treatment": {
    "solo": "Fresh coordinator completes alone; action final only. No native or adapter delegation.",
    "team": "Fresh coordinator makes a structured dispatch request for exactly evidence_checker and arithmetic_checker. Outer adapter concurrently starts two fresh CLI sessions, captures all actions, returns factual notes, then resumes the original coordinator session to synthesize.",
    "hybrid": "Same optional specialist roles, zero, one or two once each. Zero delegation valid. Delegation decision and returned notes logged.",
    "specialists": "Same fixed role prompts from v1 plus common task/tool restrictions. No coordinator draft/history or other-run answer supplied. Same complete relevant task packet accessible. No further delegation."
  },
  "mechanism": {
    "runtime": "Installed codex-cli0.165.0; Python standard-library adapter; no added package/model service",
    "native_permissions": "Native read-only sandbox and same per-invocation task-tool configuration for all roles. No global configuration, managed security policy, credential or permission changes. Available ambient tools are procedurally restricted and actual use audited; no hard isolation claim.",
    "coordinator_context": "New session per run. If delegation occurs, resume only its owned session ID; fixture demonstrated continuity. Separate process overhead remains in end-to-end latency.",
    "source_tools": "adapter/toolbox.py called via native shell: source(ID), skill(ID), aggregate(), Decimal calc(expr). Identical tool implementation and input bytes across lanes. Multiple JSON-array elements allow solo tool batching.",
    "dispatch": "Structured coordinator output requests allowed role names. External adapter logs dispatch reason, exact fixed child prompts, child-action streams and returned notes. This is adapter orchestration, not native spawn_agent behavior.",
    "native_agent_tools": "Unavailable by the same requested per-invocation configuration in every participant; no native collaboration events observed in qualification. Shell remains at native read-only permissions; study prohibits external CLI invocation or other commands by participants and audits use."
  },
  "work_proxies": {
    "aggregate_task_calls": 24,
    "final_answer_words": 600,
    "factual_note_words_shared": 500,
    "each_specialist_note_words": 250,
    "aggregate_visible_artifact_words": 1100,
    "artifact_definition": "Whitespace units in final_answer + coordinator verification_notes + all specialist final notes + dispatch_reason. JSON keys, duplicated prompts, tool results and progress commentary excluded from artifact cap but retained/quantified separately. Source/tool calls and all sessions are counted.",
    "word_cap_enforcement": "Post-hoc compliance audit; a violation is retained and fails operational acceptance. No hard generation cap.",
    "wall_clock_target_seconds": 480,
    "timing": "Initial coordinator process start request through answer/termination, including CLI startup, dispatch, child waiting and coordinator resumption. Outer processes share one deadline and receive termination on timeout; immediate remote inference termination remains unverified.",
    "tokens": "Keep every observed invocation counter and per-response event when available. Resumed-session counter semantics require reconciliation; unverified aggregates stay unknown. No token equality or dollar conversion.",
    "retries": "No task/agent retries or repairs. Provider internal transport retry details unknown unless surfaced; record observed errors and retries. No claim of seed/temperature control.",
    "each_specialist_task_call_cap": 5,
    "allocation_note": "Each bounded specialist is instructed to use at most five task calls; coordinator and children still share the aggregate24-call allowance. This role allocation is part of this adapter treatment and is reported."
  },
  "measurement": {
    "required": [
      "Attributable session starts/prompts, command-action results, dispatch, returns, final output, errors and termination",
      "Matching reported selectable model/effort/auth tags where present",
      "Task-tool calls, artifact words, time and protocol violations for every scheduled first attempt"
    ],
    "optional": [
      "Exact backend revision",
      "Complete token aggregation",
      "Subscription allowance consumption",
      "Dollar-equivalent compute"
    ],
    "cost": "No API-key request, credit purchase or top-up. Incremental paid API spend $0; subscription/credit burn and economic cost unknown.",
    "telemetry_privacy": "Whitelist event fields and model/auth/usage tags before storage. Drop reasoning items. Never save or publish raw stderr, credentials, account IDs/emails or native private rollouts."
  },
  "procedural_isolation": {
    "participants": "Fresh history, explicit non-access instructions and audited shell/tool use. Shared filesystem access is not technically denied. No other-run outputs/scorer/key reads permitted.",
    "scorer": "Fresh independent contexts receive inline anonymized final answers, task facts, rubric and ground truth only. No tools or file access requested; key and logs withheld. All first attempts finish before scoring. Record style leakage and any breach.",
    "limitations": "Procedural isolation and blindness cannot prove absence of hidden platform effects or inference from style. Findings are conditional on observed compliance."
  },
  "qualification": {
    "fixture": "setup-evidence/adapter-fixture-01",
    "result": "passed attributable action/continuity/model-tag checks on non-scored fixture; 3 participants,4 CLI invocations,11 task-tool calls,38.077539 seconds,no recorded protocol violation",
    "usage_caveat": "Raw invocation sum from fixture is preserved but not asserted as true aggregate because resumed-session counter overlap was not resolved."
  },
  "failure_handling": "Each scheduled first attempt is immutable. Capture malformed schema, timeout, wrong model, missing receipts, forbidden actions, extra children and caps as failures/deviations. Do not replace or silently exclude. Stop later execution if a systemic parity/action-coverage flaw is found; otherwise retain individual task failures and continue fixed schedule.",
  "scoring": "Use unchanged scorer-only/rubric.json and ground-truth.json. Scorer scores content blind. Root later combines content acceptance with operational compliance (no violations) and records each separately. Two independent scorers; adjudicate critical-error disputes or >=10 total-point disagreement. No repair or guide merging before scores lock.",
  "analysis": "Report four observations per lane grouped by two task clusters, individual variation, actual failures and concrete benefits if observed. No p-values or universal winner. Model review effort is not human cleanup. Missing human diagnosis/cleanup remains unknown.",
  "freeze": "Separate V2-FREEZE.json hashes amendment, adapter code, qualified fixture evidence, inherited v1 input manifest and scoring/schedule. Freeze after review corrections, before first task.",
  "independent_review": "setup-evidence/adapter-implementation-review.md"
}
