{
  "schema": "gobstopper-public-retention-ablation-protocol-v1",
  "study_date": "2026-09-19",
  "registered_at": "2026-09-19T17:21:44.099556+00:00",
  "corpus": "All 114 previously frozen archived Codex roots: 73 high-context sessions and 41 below-trigger controls. Selection is unchanged from the archived-root cohort of the 729-session study.",
  "selection_protocol": "protocol.json",
  "strategy": "scored",
  "variants": [
    "baseline_disabled",
    "candidate_disabled",
    "cutoff_0.35",
    "cutoff_0.5",
    "cutoff_0.65"
  ],
  "policy": {
    "trigger_tokens": 80000,
    "floor_tokens": 40000,
    "keep_recent_tool_outputs": 8,
    "min_savings_tokens": 4096,
    "adaptive": false
  },
  "scorer": "Deterministic offline heuristic; no local or remote model calls.",
  "probe_cap": 64,
  "hypothesis": "An optional keep-score cutoff may preserve more sampled source evidence at the cost of less reduction. Reaching the context target is not required if the cutoff protects remaining candidates.",
  "analysis": {
    "disabled_equivalence": "Compare baseline and candidate with no cutoff on all 114 inputs; ignore timings and rationale, but require matching plans and probe results.",
    "paired_retention": "Compare each fixed cutoff with the disabled baseline on the same probes. Keep all high-context sessions in the denominator. When a cutoff produces no plan, leaving the source unchanged retains the baseline probes by construction; label these cases as derived.",
    "controls": "All 41 below-trigger controls must remain no-ops under every variant. Controls have no reported literal-probe result.",
    "failures": "Retain failures and no-plan cases in reporting; no numerical exclusions.",
    "integrity": "Check each frozen input hash before and after evaluation and compare verifier severity/code counts.",
    "default": "The cutoff is opt-in for scored only. Existing defaults remain unchanged."
  },
  "limitations": [
    "The same 114 archived root sessions from one person and one Mac were replayed; this is not a new independent sample.",
    "This development comparison reuses an already studied corpus. The original baseline was already known when the three cutoffs were registered before this run; this is not held-out or external validation.",
    "All three cutoffs were registered before outcomes, but none is established as a tuned optimum on held-out tasks.",
    "Literal matching is not semantic importance, correct continuation, task success or automatic recovery.",
    "No provider continuation, billing, local model or remote model was evaluated; this is not a fast-jev-compaction comparison.",
    "No-plan cases remain in reduction denominators as zero. For two high-context no-plan cases at cutoff 0.35, retention is derived from leaving the source unchanged, not measured on a rewrite.",
    "The baseline and candidate use the corrected 64-probe cap. Probe totals differ from the original 729-session study; do not compare their retention ratios as a policy effect.",
    "Structural severity/code counts did not increase; that does not establish safe provider resume.",
    "Timing is descriptive under shared host load, not a latency benchmark.",
    "Private transcripts, session identifiers, paths, per-session hashes, corpus manifests and per-session results are excluded; the exact private corpus cannot be independently replayed from this download."
  ]
}