{
  "schema": "gobstopper-public-apple-retention-pilot-v1",
  "study_date": "2026-09-19",
  "started_at": "2026-09-19T17:43:42.727681+00:00",
  "completed_at": "2026-09-19T17:44:18.902579+00:00",
  "scope": "Three frozen convenience-selected Codex root-task inputs from one Mac. Bounded Apple overlay versus the offline heuristic, both using scored with keep_score_threshold 0.5. Mechanism qualification only; separate from the 729-session study and 114-root development comparison.",
  "provenance": {
    "binary_sha256": "0d924d0b8322bd58c8ab4cee7318c41177c809433c61d12aa0ddb7cd26f5b88a",
    "bridge_sha256": "788f4a8aaae8860028f66634e6b5c46046deeb3bccc47fdcc0327917e6739796",
    "binary_and_bridge_unchanged": true,
    "inputs_unchanged": true
  },
  "summary": {
    "total_evaluations": 6,
    "failed_evaluations": 0,
    "wall_seconds": 36.181,
    "remote_model_calls": 0,
    "mechanism_qualified": true
  },
  "aggregates": [
    {
      "variant": "heuristic",
      "selected_inputs": 3,
      "successful_evaluations": 3,
      "failed_evaluations": 0,
      "plans": 3,
      "no_plan": 0,
      "sum_context_before": 644539,
      "sum_projected_reclaimed": 323532,
      "median_projected_reduction_percent_including_noops": 50.34716684211086,
      "min_projected_reduction_percent": 46.54106204133757,
      "max_projected_reduction_percent": 53.73961218836565,
      "literal_samples": 3,
      "literal_total": 192,
      "literal_retained": 148,
      "literal_unchanged_source_derived_samples": 0,
      "tail_samples_with_probes": 2,
      "tail_total": 3,
      "tail_retained": 3,
      "new_errors": 0,
      "new_warnings": 0,
      "selected_outputs": 107,
      "score_protected_outputs_reported_for_plans": 4,
      "median_wall_ms": 90,
      "sum_wall_ms": 252,
      "sum_apply_ms": 77
    },
    {
      "variant": "apple",
      "selected_inputs": 3,
      "successful_evaluations": 3,
      "failed_evaluations": 0,
      "plans": 1,
      "no_plan": 2,
      "sum_context_before": 644539,
      "sum_projected_reclaimed": 5574,
      "median_projected_reduction_percent_including_noops": 0.0,
      "min_projected_reduction_percent": 0.0,
      "max_projected_reduction_percent": 2.4469799070200313,
      "literal_samples": 3,
      "literal_total": 192,
      "literal_retained": 192,
      "literal_unchanged_source_derived_samples": 2,
      "tail_samples_with_probes": 2,
      "tail_total": 3,
      "tail_retained": 3,
      "new_errors": 0,
      "new_warnings": 0,
      "selected_outputs": 14,
      "score_protected_outputs_reported_for_plans": 33,
      "median_wall_ms": 11912,
      "sum_wall_ms": 35866,
      "sum_apply_ms": 27,
      "apple_summary_available_samples": 3,
      "apple_overlay_successful_samples": 3,
      "fallback_diagnostic_samples": 0,
      "apple_totals": {
        "selected_candidates": 92,
        "unique_lines": 92,
        "cached_batches": 0,
        "model_calls": 12,
        "items_overlaid": 92,
        "failed_batches": 0,
        "scoring_duration_ms": 35430
      }
    }
  ],
  "paired": {
    "successful_pairs": 3,
    "changed_selected_outputs": 3,
    "changed_projected_savings": 3,
    "literal_comparable_pairs": 3,
    "apple_literal_retention_higher": 3,
    "apple_literal_retention_lower": 0
  },
  "interpretation": "The Apple overlay executed successfully and was more conservative on these three inputs, with longer observed wall time. More literal retention came with two no-op cases and much less projected reduction; this establishes no task-quality, cost, latency or backend-superiority claim.",
  "retention_accounting": "All variants use 192 baseline probes. For the Apple variant, 128 of 192 retained probes are derived from its two unchanged no-plan inputs; only the remaining 64 probes were checked against its rewritten output.",
  "coverage": "The Apple model overlays at most 32 selected candidates per input. Successful overlays covered 92 of 92 selected candidates; candidates outside that bound retain heuristic scores. That expected heuristic remainder is distinct from failed-backend fallback.",
  "limitations": [
    "Three convenience inputs from one Mac; this qualifies the Apple-plus-cutoff execution mechanism only.",
    "Both variants use the preregistered 0.5 threshold; this does not tune or establish an optimal threshold.",
    "Apple overlays at most 32 selected candidates; unscored candidates retain heuristic scores. Heuristic remainder is expected coverage, distinct from backend failure fallback.",
    "Literal string survival is not semantic understanding, task correctness, or resume qualification. Protected-tail probes can be absent.",
    "Token changes are offline estimates, not measured provider-billed savings, live context, cache cost, or productivity.",
    "One pass without repeatability or model cold/warm control; timings are descriptive and model outputs may vary.",
    "Raw inputs, input hashes, selected line positions, prompts, responses and missed probes are never public."
  ],
  "privacy": "Aggregate methodology only. Private inputs, input hashes, selected line positions, prompts, responses, missed probes and per-input records are excluded."
}