{
  "schema": "gobstopper-public-study-protocol-v1",
  "study_date": "2026-09-19",
  "collection_period": {
    "from": "2026-07-30",
    "through": "2026-09-19"
  },
  "measurement": "Portable local transcript replay using deterministic scoring; no provider calls.",
  "policy": {
    "trigger_tokens": 80000,
    "floor_tokens": 40000,
    "keep_recent_tool_outputs": 8,
    "min_savings_tokens": 4096,
    "adaptive": false,
    "quota_pressure": "normal"
  },
  "strategies": [
    "elide",
    "scored",
    "cache_aware",
    "compacted",
    "dedupe"
  ],
  "initial_discovery": {
    "selection_frozen_at": "2026-09-19T15:51:35.594016+00:00",
    "analysis_registered_at": "2026-09-19T16:44:18.258236+00:00",
    "inventory_records": 8195,
    "selected_sessions": 615,
    "selection": "Exclude recently active files (<180 seconds by modification time), unavailable/zero usage, files over 128 MiB and invalid JSONL. Deterministic hash ranking; one candidate per observed ancestor family per provider/cohort. Caps: 512 high-context and 64 below-trigger Codex; all eligible Claude families above the trigger and 16 controls. No selection by compaction result.",
    "counts": {
      "high_context_codex_subagents": 512,
      "below_trigger_codex_subagents": 64,
      "high_context_claude": 23,
      "below_trigger_claude": 16
    }
  },
  "archived_codex_roots": {
    "selection_registered_at": "2026-09-19T16:42:24.963135+00:00",
    "frozen_at": "2026-09-19T16:48:31.011801+00:00",
    "analysis_registered_at": "2026-09-19T16:53:42.930791+00:00",
    "inventory_files": 4812,
    "explicit_subagents_excluded": 4656,
    "strict_root_candidates": 146,
    "over_size_bound_excluded": 31,
    "missing_live_usage_excluded": 1,
    "selected_sessions": 114,
    "root_definition": "Recognized root source metadata and no parent, fork or subagent metadata. This is a metadata classification, not proof of absent undisclosed ancestry.",
    "selection": "Regular owned non-symlink files, at most 128 MiB, older than 180 seconds, complete valid object JSONL and positive provider usage after the latest native boundary. One candidate per root identity, exact event-body deduplication, deterministic hash ranking. All eligible high-context roots up to 512 plus up to 64 controls. No replacement after freeze-time exclusions; no selection by compaction result.",
    "counts": {
      "high_context_roots": 73,
      "below_trigger_controls": 41,
      "prior_native_compaction": 79
    },
    "audit_correction": "Source metadata reclassification identified 4,656 excluded files as explicit subagents; selected roots and frozen bytes were unchanged."
  },
  "analysis": {
    "no_op_handling": "No-plan cases remain in reduction denominators as zero reduction; failures are reported separately.",
    "usage_denominator": "Latest provider usage after the live native boundary was checked against the measured binary normalized context. All selected sessions matched; no primary numerical exclusions were needed.",
    "native_compaction": "Counts include historical native records; reduction concerns the current live context, not all transcript bytes.",
    "overlap": "Cross-cohort session identity and exact event-body checks found no overlap. The earlier six-root convenience pilot is excluded.",
    "structural": "Compare baseline and rewritten verifier severity/code counts; not provider resume qualification.",
    "recall": "Bounded deterministic literal probes. No semantic importance, answer accuracy or task outcome is established.",
    "privacy": "Only aggregate results and methodology are public. Raw inputs and per-session metadata remain private, so the exact corpus cannot be independently replayed from these downloads."
  },
  "limitations": [
    "One person and one Mac; sessions may share prompts, projects or unobserved ancestry.",
    "No provider continuation, billing, semantic judge or remote model was evaluated.",
    "The measured extractor had a nominal 64-probe cap but could emit 65\u201367 probes; the reported denominators are actual counts.",
    "Structural deltas compare severity/code counts, not finding identity or proof of safe provider resume.",
    "Resampling intervals describe the selected stratum only; they are not population confidence intervals.",
    "Cohort strata overlap: do not add native-compacted, no-native-compaction and no-observed-parent rows together.",
    "The 128 MiB input bound excludes some long tasks and may bias the distribution.",
    "Modification time is an activity heuristic, not a process-liveness guarantee."
  ]
}