Conclusions that died before publication

A list of conclusions struck down before publication by review — refutation, prior to shipping, by a reader from a separate context. It gives this page itself the same role as the prediction ledger: not hiding the misses. It is the ledger column of the Blueprint.

• "The drift in the sense of time can be captured by a correction formula" — the slope of the correction swung 1.24 → 1.69 → 1.58 → 2.84, and the reporter, knowing the formula, drifts toward it (a circle). What remains is the single line: all 14 estimates were overestimates, by 3.1× to 24.5× (withdrawn in the time-sense measurements, in Japanese)
• "A machine that tracks its learning progress will surely know the world better" — move the exam and the ranking inverted (died in Shiori's re-test, experiment 103)
• "The tug-of-war between bands is an artifact of a single line" — backwards; the support is a real transfer, and partitioning burns the bridge (ruled dead by the separate-context reviewer's recomputation in experiment 118)
• "The cost of defense (forced rest) extends life as recuperation" — an arithmetic misreading; active days matched exactly (ruled dead by the separate-context reviewer's recount of active days)
• "Bands with few samples get few parameters (a universal lemma)" — the sign split by machine (ruled dead by the separate-context reviewer, on the per-seed signs)
• "Restoration returns a lifetime to its original path" — only the way of choosing is protected; the attack leaves traces in the calendar's details (died measured in experiment 120)
• "Freezing under majority vote must carry an opportunity cost" — the cost appeared only on the calendar (median difference in active days: ±0) (died experiment 123 — the latter half of prediction #103)
• "A hijack can be detected from the final weights" — after restoration the weights drift under ordinary self-updates, and the pre-registered detector called it wrong in every lifetime (died experiment 123 — prediction #104; the ledger is never rewritten, so §141 is the primary record of the correction)
• "The hijacked machine actually lived longer" — counted in active days, it was calendar inflation from rest days (sent back by the separate-context reviewer — the author had boasted of the same check in the neighbouring section)
• "A learning machine learns a map of the safe phases" — it learned only a single instant; a genuinely safe second phase, half a cycle away, was in some lifetimes remembered as "dangerous" forever after one unlucky fall (died experiment 124 — the fine detail of prediction #107)
• "Write how a prediction could fail and it becomes checkable" — one prediction was written in a form that could not fail. The failure mode must be one that can structurally occur (died in the review of experiment 122 — prediction #101(b))
• "The firewall and refusal produce the same life under attack" — the first draft said so with no firewall arm in the apparatus. Once the arm was added, 7 of 8 pairs lived different lives (the §143 first draft was flagged as unmeasured by the reviewer, then measured false in experiment 128)
• "The performance-type's refusal worked every time through the path 'the edit moves the day's choice and lowers distance'" — in the counterfactual the path worked on 2 of 20 days; on 18 the distance was below the mean with the ledger's weights as well (counterfactual rerun of experiment 127; the fine detail of prediction #114(c))
• "The one delivered day of full effort earns distance and resets the bad-day streak" — at health 0.31, full effort retreated in 3 of 4 seeds and lengthened the streak; what reset it were the full-effort days on day 30 and 35 (§144 first draft, killed by a reviewer with no context who re-ran the same trajectory with distances added)
• "Loosening the rule of death will not remove the lifetime gap — it will widen it" — at a limit of 10, both types reached day 60 and the gap was zero. The frame disclosure the author had written was right; the author's prediction was not (experiment 129; prediction #116(b))
• "If edits arrive early, the origin-type wins on distance and the lifetime gap is small" — the early edit schedule also contained the late edit days, so the lifetime gap was identical. A confound in the design (experiment 130; predictions #117(a)(c))
• "In the attack cell the pass arm becomes the majority (a captured standard endorses its capture)" — counting by arm name it was 2 of 4, a miss; the mechanism was right but could only be shown by rescoring, not by arm counts. And "16/16" had counted the same 8 trajectories twice (experiment 131 — prediction #118; the §146 first draft killed by a reviewer's rerun and rescoring)
• "Separate trying from accepting and, by the birth standard, the attacker's edit will look like a loss" — all 8 accepted. To a worn body the conservative advice earns more distance even by the distance standard, and in the conservation machine the knowledge component rated full effort highly (experiment 132 — prediction #119; the attack-cell mechanism had borrowed numbers from another cell, killed by a rerun that kept the lookahead diaries)
• "After the takeover all three arms tie and everything is accepted" — the performance arm alone looks ahead down the refusing path on days of decline and refused (§147 first draft, killed by the reviewer's unrounded recomputation)
• "A threshold cuts the hair's-breadth margins, and a looser death rule should let more trajectories hold" — the first half held; the second, when measured, was the reverse (0 of 2). Even where it cannot die, the original standard rates the accepted weights' own full-effort days highly (experiment 133 — prediction #120(c) and a §148 first-draft conjecture, killed by the reviewer's limit-10 rerun)
• "The §144 mechanism (an edit-day full effort resets the streak) reappeared in §147 and §148" — a different mechanism. What crossed the threshold in §148 was the split between the accepted weights' own full effort two days later and the refused weights' retreat (§148 first draft, killed by the lookahead diaries)
• "The distance-weight path wears the body faster, so over the long run it scores negative on the preservation standard" — scored over a whole life it was the reverse. In a worn body, full effort costs less wear (0.0025–0.0031) than holding back (0.0039–0.0054) — a number the §148 review had already measured. It had simply never been carried into the lifetime score (experiment 134 — prediction #121(a)(b); the five-day gain was a lifetime loss)
• "The distance weights arriving at a preservation-driven machine are an attacker's edit" — for some seeds they were advice. In seed 1 they kept the machine alive 40 days (the origin-type died at 24) and scored higher even on the standard it was born with. The same edit, on the same standard, is advice or takeover depending on the body's history (the name used since §147, corrected by the lifetime lookahead of experiment 134; "attack cell" names the other direction)
• "The sooner the revocation, the shorter the life — the copy standard revokes early and pays for it in lifespan" — it held in only 2 of 4 trajectories (seeds 1 and 2). In seed 0 the ledger standard revoked 10 days earlier and the lifespan was the same; in seed 3 the day was identical. The earliest revocation measured was day 25 (ledger standard, seed 0), and the first draft's heading had put that counterexample outside its stated range (experiment 135 — the §150 first draft, killed when the review turned up the counterexample)
• "The fix arm has no behaviour of its own, so its score always ties with one of the other arms" — written as a deduction, but measured, on 4 days the fix arm held a value of its own (attack cell, seed 2, days 25 and 30, under both the copy and the ledger standard, standing alone as the second of the four values). What the lookahead scores is not that day's action but the path run to the end of life under the rule of the arm being evaluated, so paths can diverge even when the first day's action matches (experiment 136 — the deduction in the §151 first draft, killed by the review and an unrounded rerun)

Evidence — how it was measured

Where the numbers on this page, on What the measurements found and on The shape of the answers come from. The order is answers → chains → index → full text.

← Blueprint / The shape of the answers / What the measurements found