{
 "measurement": "Whether the skills curve behind this site's context calculator reproduces in a fresh session, at both ends",
 "ran_on": "2026-08-18",
 "claude_code_version": "2.1.234",
 "claude_code_version_recorded_at_runtime": "2.1.234 (Claude Code)",
 "prompt": "Reply with only: OK",
 "the_prediction": {
  "registered_before_the_run": "That the 10-skill cell, published at 400 tokens, would move because it is the same order as the 198 and 521 intermittent components already published here, and that the 1000-skill cell, published at 11711, would hold relative to its size.",
  "outcome": "CONFIRMED in direction",
  "how": "The 10-skill cell's spread is 281% of its own mode; the 1000-skill cell's is 8%. Relative unreliability differs by 36x, in the predicted direction. The prediction said the large cell would hold to 'a few percent at most' and it came in at 8%, so the direction is right and that bound was optimistic."
 },
 "method": {
  "shape": "Two cells, 10 and 1000 skills, 8 rounds each, in a session months of releases after the published curve. The two ENDS were chosen deliberately: a middle cell would move under either explanation and so could not distinguish them.",
  "fixture": "The original harness's generator is imported and called, not re-implemented, and the emitted SKILL.md is asserted at 336 bytes before any run. Re-implementing it produced 290 tokens against a published 400 and would have been published as a 27% drift in the calculator.",
  "no_permission_flags": "Deliberately matches the flags of the harness that produced the published curve, so this is a re-measurement rather than a new condition.",
  "controls": {
   "paired": "Floor re-measured every round; every delta is against that round's floor.",
   "arrival": "Every reply must be exactly OK; asserted before assembling.",
   "isolation": "--setting-sources project asserted onto argv.",
   "fixture_asserted": "SKILL.md byte size checked against the original every round."
  }
 },
 "results": {
  "by_cell": {
   "10": {
    "rounds": 8,
    "all_rounds": [
     192,
     192,
     390,
     390,
     390,
     390,
     588,
     1287
    ],
    "mode": 390,
    "mode_seen": 4,
    "min": 192,
    "max": 1287,
    "spread": 1095,
    "spread_as_share_of_mode": 2.808,
    "distinct_pairwise_gaps": [
     198,
     396,
     699,
     897,
     1095
    ],
    "published_2_1_233": 400,
    "mode_minus_published": -10,
    "estimator": "REFUSED: the spread is 281% of the mode, so no central value is publishable for this cell."
   },
   "1000": {
    "rounds": 8,
    "all_rounds": [
     11701,
     11701,
     11701,
     11701,
     11701,
     11899,
     11899,
     12598
    ],
    "mode": 11701,
    "mode_seen": 5,
    "min": 11701,
    "max": 12598,
    "spread": 897,
    "spread_as_share_of_mode": 0.077,
    "distinct_pairwise_gaps": [
     198,
     699,
     897
    ],
    "published_2_1_233": 11711,
    "mode_minus_published": -10,
    "estimator": "Mode, with the spread published beside it."
   }
  },
  "relative_unreliability_ratio": 36.5,
  "the_published_values_reproduce": "Both modes land -10 tokens from the published figure, an identical constant offset at both ends rather than a proportional drift. The curve itself is intact on 2.1.234; what does not survive is the impression of precision that 'spread 0 over four rounds' gave it.",
  "what_a_reader_should_take": "At 1000 skills the figure is worth quoting to within about 8%. At 10 skills it is not worth quoting as a single number at all: eight readings ran 192 to 1287. Most people have few skills, so the unreliable end is the end most people are standing at.",
  "runs": [
   {
    "cell": 0,
    "round": 0,
    "status": "ok",
    "skills_on_disk": 0,
    "context": 22978,
    "delta": 0,
    "reply_exact_ok": true
   },
   {
    "cell": 10,
    "round": 0,
    "status": "ok",
    "skills_on_disk": 10,
    "context": 23368,
    "delta": 390,
    "reply_exact_ok": true
   },
   {
    "cell": 1000,
    "round": 0,
    "status": "ok",
    "skills_on_disk": 1000,
    "context": 34679,
    "delta": 11701,
    "reply_exact_ok": true
   },
   {
    "cell": 0,
    "round": 1,
    "status": "ok",
    "skills_on_disk": 0,
    "context": 22978,
    "delta": 0,
    "reply_exact_ok": true
   },
   {
    "cell": 10,
    "round": 1,
    "status": "ok",
    "skills_on_disk": 10,
    "context": 23170,
    "delta": 192,
    "reply_exact_ok": true
   },
   {
    "cell": 1000,
    "round": 1,
    "status": "ok",
    "skills_on_disk": 1000,
    "context": 34679,
    "delta": 11701,
    "reply_exact_ok": true
   },
   {
    "cell": 0,
    "round": 2,
    "status": "ok",
    "skills_on_disk": 0,
    "context": 22780,
    "delta": 0,
    "reply_exact_ok": true
   },
   {
    "cell": 10,
    "round": 2,
    "status": "ok",
    "skills_on_disk": 10,
    "context": 23170,
    "delta": 390,
    "reply_exact_ok": true
   },
   {
    "cell": 1000,
    "round": 2,
    "status": "ok",
    "skills_on_disk": 1000,
    "context": 34679,
    "delta": 11899,
    "reply_exact_ok": true
   },
   {
    "cell": 0,
    "round": 3,
    "status": "ok",
    "skills_on_disk": 0,
    "context": 22978,
    "delta": 0,
    "reply_exact_ok": true
   },
   {
    "cell": 10,
    "round": 3,
    "status": "ok",
    "skills_on_disk": 10,
    "context": 23368,
    "delta": 390,
    "reply_exact_ok": true
   },
   {
    "cell": 1000,
    "round": 3,
    "status": "ok",
    "skills_on_disk": 1000,
    "context": 34679,
    "delta": 11701,
    "reply_exact_ok": true
   },
   {
    "cell": 0,
    "round": 4,
    "status": "ok",
    "skills_on_disk": 0,
    "context": 22780,
    "delta": 0,
    "reply_exact_ok": true
   },
   {
    "cell": 10,
    "round": 4,
    "status": "ok",
    "skills_on_disk": 10,
    "context": 23368,
    "delta": 588,
    "reply_exact_ok": true
   },
   {
    "cell": 1000,
    "round": 4,
    "status": "ok",
    "skills_on_disk": 1000,
    "context": 34679,
    "delta": 11899,
    "reply_exact_ok": true
   },
   {
    "cell": 0,
    "round": 5,
    "status": "ok",
    "skills_on_disk": 0,
    "context": 22081,
    "delta": 0,
    "reply_exact_ok": true
   },
   {
    "cell": 10,
    "round": 5,
    "status": "ok",
    "skills_on_disk": 10,
    "context": 23368,
    "delta": 1287,
    "reply_exact_ok": true
   },
   {
    "cell": 1000,
    "round": 5,
    "status": "ok",
    "skills_on_disk": 1000,
    "context": 34679,
    "delta": 12598,
    "reply_exact_ok": true
   },
   {
    "cell": 0,
    "round": 6,
    "status": "ok",
    "skills_on_disk": 0,
    "context": 22978,
    "delta": 0,
    "reply_exact_ok": true
   },
   {
    "cell": 10,
    "round": 6,
    "status": "ok",
    "skills_on_disk": 10,
    "context": 23368,
    "delta": 390,
    "reply_exact_ok": true
   },
   {
    "cell": 1000,
    "round": 6,
    "status": "ok",
    "skills_on_disk": 1000,
    "context": 34679,
    "delta": 11701,
    "reply_exact_ok": true
   },
   {
    "cell": 0,
    "round": 7,
    "status": "ok",
    "skills_on_disk": 0,
    "context": 22978,
    "delta": 0,
    "reply_exact_ok": true
   },
   {
    "cell": 10,
    "round": 7,
    "status": "ok",
    "skills_on_disk": 10,
    "context": 23170,
    "delta": 192,
    "reply_exact_ok": true
   },
   {
    "cell": 1000,
    "round": 7,
    "status": "ok",
    "skills_on_disk": 1000,
    "context": 34679,
    "delta": 11701,
    "reply_exact_ok": true
   }
  ]
 },
 "known_limits": {
  "two_cells_not_a_curve": "The two ends were measured, not the eight points between them. Where along the curve a figure becomes worth quoting is not established here.",
  "carried_forward_one_skill_is_inside_the_noise": "skills-cost-2-1-233.json carries: 'The 1-skill cell measured -149.0 with a spread of 521 tokens across rounds ... no claim is made about a single skill.' This run extends that limit upward rather than retiring it: at 10 skills the signal is still smaller than the round-to-round spread.",
  "the_constant_offset_is_unexplained": "Both cells came in exactly 10 tokens below their published values. A constant at both ends is not proportional drift and is too small to matter for any decision, but it is recorded rather than smoothed away, and this run does not explain it.",
  "one_machine_one_release": "Opus on Claude Code 2.1.234, macOS, one session."
 }
}