Experiments
Each study here was written down and locked before it ran, and is reported with whatever it found. Nulls stay on the record. 90 studies have cleared review so far, newest first, and more join as they clear. Every study cleared so far came back null or went against its own forecast, and each card says which.
Each card keeps its number in our findings ledger (F667 is the 667th finding). It gives the finding, the numbers behind it, the forecast against the outcome where the study made one, the verdict, the limits, and a Verified line saying what was checked.
90 studies shown.
-
McKibben frozen formula is untestable on the open sample: 3 papers with free contraction, none with force; on those 3 the frozen ceiling over-reads and the false 45 form sits closer by .045 log
Null or refuted Null by lack of data.
Verified 2 November 2026: all 3 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). No outside reviewer has read it. -
Pressure-loaded tubes sit far above tension members in fibre angle, D 26.7 against q95 14.4, but 54.74 is specific only in engineered pipes, median 54.6, where designers use it; biology ties 45 on 7 papers
Null or refuted Thin sample, follow-up frame.
Verified 2 November 2026: all 6 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). No outside reviewer has read it. -
Drawn images beat text by .38 and .27 on two new spatial classes and lose .40 and .48 on non-spatial ones at 20 tokens, interaction .66, but the locked verdict fails on its shuffled control at .35 and the advantage may be a token limit
Null or refuted Failed its locked control; direction clear, cause open.
Limits Inference across two sites, not an experiment on it.
Design locked before the run:
sha256 d849065062b0Verified 2 November 2026: all 1 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Pairs absent against neighbours fill at .24 by 2025 against .05 for random empties, but blind raters called 191 of 200 plain chance, so a class of taboo, instrument, impossibility or vocabulary could not be tested on a random vocabulary, gain -.004
Null or refuted Untested in effect.
Limits Model readers, one family plus one small local model; no model took audio. Simulations and class labels are mine, class labels after results. Text and graph were not rerun in T4b.
Verified 2 November 2026: all 3 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus blind model readers; self-graded, self-graded. No outside reviewer has read it. -
Route search over a 52 card mechanism graph supports no new architecture: baselines rank top 3 in 6 of 6 jobs and in 6 of 6 under shuffled costs, the one cost winner is an accounting artefact, and 81 percent of energy links and 62 percent of signal links have no known cost
Null or refuted Route verdict rests on 6 of 6 under shuffled costs. Cost table and the analytic hourglass optimum are exploratory.
Verified 2 November 2026: all 8 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Interface properties did not beat node properties in predicting a hidden link under BH: delta -.065 and +.028 overall, +.027 and +.085 cross region, q .55 to .74, with the independent cross region arm the one worth a retest
Null or refuted A null with low power.
Verified 2 November 2026: all 0 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Hidden mechanism links are recovered by word matching as well as by typed structure: AUC .76 and .85 for words against .78 and .81 for structure on 50 cards, with planted .99 and shuffled .56, so the atlas earns no link suggester yet
Null or refuted A null against a baseline, self-graded.
Verified 2 November 2026: all 1 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
An interface beam stem model overshoots Petersen 1995 by 3.45 times, 72 against 21 percent, gives CoCr over Ti 1.11, and at a fixed femur predicts a diameter correlation that was not reported
Null or refuted Model fails its record on the sign corrected bound under a uniform size spread. 3.45 is exploratory, dividing unlike quantities, and the correlation result depends on an unreported diameter distribution.
Limits Simulation, values from memory, one cross-sectional titanium paper, spread read by eye. Nothing clinical, no device claim.
Verified 2 November 2026: all 5 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
H23 fails by its own kill: a first passage skeleton is all xylem, stretched water and cold atoms share. Xylem steepness 6 to 8 against water 88, classical limit 55 times P50 with barrier 1.3e5 kT, and a false fibre Weibull map sits inside xylem's range, so shape can't name nucleation; time dependence would, and no open data has it
Null or refuted H23 kill met: false map passes on the shape statistic
Design locked before the run:
sha256 46b928db6047Verified 2 November 2026: all 5 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
In 2D at Re 200 a wake moves a following foil's power from -84 to +88 percent and its power plus drag from -46 to +31 percent with gap and phase, but every foil is net drag, so the study's efficiency verdicts pass emptily and no thrust saving is shown
Null or refuted Simulation with an unresolved solver question.
Verified 2 November 2026: all 6 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent, self-graded, self-graded. No outside reviewer has read it. -
Equal mass grading drops stress concentration from a mesh dependent 3.4 to 5.1 to 1.15 to 1.6 at 4 to 10 percent more compliance, and linear relaxation eases nothing under fixed load, .9975, while power law creep does, .78 at n 3, at 35 times curvature; smoothing the sharp profile misses by 57 percent
Null or refuted Known answers, costs stated.
Limits Plane stress, Basquin from memory, no collar or flare geometry table, Mandel not run. Frozen form only. Linear, one seed set, values from memory marked in code; no hardware.
Design locked before the run:
sha256 583fed041aa9Verified 2 November 2026: all 9 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Shutdown and shedding: a local carbon rule reaches within 4 percent of full knowledge in drought while cutting 7 of 30 recoverable branches, a store signal removes that gap, a growth rule beats a fixed stem 10 percent under changed wind and loses 50 percent under a biased read, and shutdown's worth reverses with my flow model
Null or refuted Toy with one open endpoint table.
Limits Shed model slack; binding store killed all rules. Growth and hydraulic numbers post hoc. No open taper table. CODIT in words only. Linear, one seed set, values from memory marked in code; no hardware.
Design locked before the run:
sha256 583fed041aa9Verified 2 November 2026: all 5 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Peak's three leaf statistics do not identify stomatal coupling, uncoupled 2.65 and coupled 2.83 from them, while only coupled lattices form patches, .95 against -.02; passive valves cut waste to 0 with three bursts at 57 percent pump power, where a regulator wastes 47 percent at 145
Null or refuted Toy; statistic non-identification is the usable part.
Limits All post hoc, units arbitrary, threshold set after a bug, no leaf series, no microfluidic data. Linear, one seed set, values from memory marked in code; no hardware.
Design locked before the run:
sha256 583fed041aa9Verified 2 November 2026: all 6 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
On 70 board knots, potential flow gained nothing as locked, 15.6 against 15.0 degrees, because saw marks dominated; with marks removed near the knot it gains 3.6 degrees and 1.0 over one non-flow bulge, leaving 39 degrees unexplained
Null or refuted Weak measure; published flow-grain work read through summaries.
Limits Clean pass is post hoc, bulge not refitted, ring trace mixed with fibre, pixel aspect unknown. Shape does not name flow. Linear, one seed set, values from memory marked in code; no hardware.
Design locked before the run:
sha256 583fed041aa9Verified 2 November 2026: all 6 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Kubler growth stress raises first compressive yield 18 percent at 5 MPa and leaves ultimate moment level, with stored energy 2.2 percent of Handbook work to maximum load
Null or refuted Known model; quantified.
Limits Compression perfectly plastic, tension strength from memory, axial unchanged by construction. Linear, one seed set, values from memory marked in code; no hardware.
Design locked before the run:
sha256 583fed041aa9Verified 2 November 2026: all 3 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Do coverage gaps predict later results: void, the 2D plant failed at 1 of 11 thin at 2000 because old titles count monolayers on substrates; on file unscored, gap score 0.615 and 0.697 against additive expectation 0.654 and 0.770, and thinness against citations rho 0.047 and -0.008
Null or refuted Void by its own kill; informative about why
Design locked before the run:
sha256 52eafa20e1d3Verified 2 November 2026: all 10 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). No outside reviewer has read it. -
Crossing rays in the hole plate are real two-fan density and not sampling, correlation .998 across sampling and .98 with density product, and my spectral moire finder fails on real figures though it works on gratings; moire adds reach for coarse detectors and no information
Null or refuted Known physics; finder failed
Design locked before the run:
sha256 364a7bf30ff8Verified 2 November 2026: all 0 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Shape does not predict a shared equation in 23 render and chart pairs: Moebius AUC .63, inside shuffled labels, reader AUC .51, and 3 of 5 labelled pairs a blind reader rated alike share none, though the locked score failed its plant at .43
Null or refuted Small n, kill fired
Design locked before the run:
sha256 364a7bf30ff8Verified 2 November 2026: all 3 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Black hole index rays reach Smith arcs only as arcs: rays stay within .5 percent of circles, no chart beats nulls, a b-dependent line cutoff restates Bouguer exactly and carries the exponent, and the frozen Z0/n line fails since n gives no reflection when epsilon equals mu
Null or refuted Most content known; the line mapping not searched widely
Design locked before the run:
sha256 364a7bf30ff8Verified 2 November 2026: all 0 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Root and pattern arm void: the planted factored language reached 0.87 where the prereg required 0.80; unscored, factoring raised cost in five of six sets, Arabic 1.06, Hebrew 1.08, with exact round trip on all
Null or refuted Void by its kill.
Limits Plant bar may be too strict for my code, not a finding about any language.
Design locked before the run:
sha256 8882c848b3ceVerified 2 November 2026: all 4 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Signal shape: spectral flatness says nothing about recovery, rho -0.006 against a forecast -0.8, while byte entropy does, -0.82 generic and -0.85 own decoder, and Reed-Solomon looks like compressed data until its decoder runs
Null or refuted A null with small n.
Limits Eleven points, hand-chosen layers.
Design locked before the run:
sha256 b6298885c1bfVerified 2 November 2026: all 4 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Locked literary device scoring is void, its plant failed and its control gained as much as real cues; read blind, formula adds 0.067 masked top-1 in verse and 0.042 in prose, rhyme 0.031 in verse, and meter, alliteration, parallelism and repetition about 0
Null or refuted Locked result void; blind one exploratory.
Limits Blind plant reached only 0.09. Trigram reader. Says nothing of people or oral transmission.
Design locked before the run:
sha256 b6298885c1bfVerified 2 November 2026: all 5 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
On 64 shared chips at 10 dB, a matched filter keeps 1.00 with Walsh, .91 with Gold and .28 with random codes, where joint decoding keeps random at .79, so orthogonality buys a cheap receiver more than capacity; past the chips a second basis beats random, .42 against .26; colour overlap from CIE tables gives .82 joint and .52 matched
Null or refuted Theory reproduced; the channel tie is inference.
Verified 2 November 2026: all 3 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Noise amplification matches the Frobenius law within .9 percent and condition number does not rank it: whitening cond 170 amplifies 3773, a random matrix cond 664 amplifies 1764; closed-loop prediction spends 2.18 bits per pixel where the same predictor open loop spends 3.30
Null or refuted Known in principle, measured here.
Verified 2 November 2026: all 6 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Keeping only a covariance's diagonal costs 2.23 bits per pixel on image patches, .285 of the diagonal code, while the diagonal holds 2 percent of squared energy; held out it stays 140 of 143 bits, but on air sensors split in time the full model loses 11 bits to its diagonal
Null or refuted Formula is textbook; the drift row is the useful part.
Verified 2 November 2026: all 5 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
On NGSIM, five car-following rules fit at NMSE .10 to .56 and disagree on a capacity change by .25, the best on dev is worst on the next road, and 3 of 8 hidden chart failures sit under 1 pixel at q99 on a 4K panel
Null or refuted Small, with leaks.
Limits Two highway segments; readers are two language models, one invented marks; off-frame plates leak through the y-axis.
Verified 2 November 2026: all 4 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Changing the frame flips the reading on 2 of 4 relief legs for rate against interval, 2 of 4 for near-far against after-before, 2 of 46 for mean against tail, and 36 of 126 transfer cells for median against worst
Null or refuted A check on the test.
Limits The fault interval reading may come from declustering, not tested. Flips describe the tests.
Verified 2 November 2026: all 8 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Reburn falls to .60 of the later rate near a burn, with planted relief found 1.0 of the time, while protest rises to 1.39 with power .9, and the fault and stomata tests find planted relief at most one time in five, so they can't speak to relief
Null or refuted Power figures are the useful part.
Limits Fuel loss is the plain cause on fire, no shared mechanism shown; protest counts are reports; power depends on my synthetic catalogue, changed before lock after a hung smoke.
Verified 2 November 2026: all 2 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Fitted shape parameters track the false passes: with two or more, false maps counted 9 of 16 and real 1 of 8; with at most one, real 5 of 9 and false 3 of 9
Null or refuted Direction only.
Limits Small counts; the cut was set in the prereg after a dev look at two forms; neither split is significant.
Verified 2 November 2026: all 8 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Statement errors and proof errors need separate counts: the locked corpus planted 362 proof and format faults and 0 statement errors, and on 3 hand-written statement errors, a false axiom, a converse stated as an axiom and a hypothesis assuming its conclusion, all 4 kernels accepted all 3 while rejecting a proof fault control
Null or refuted A category demonstration, small and exploratory.
Limits 3 files, by design accepted by any Metamath checker, so this demonstrates the category and gives no rate. It says nothing about how often real statements are wrong. Statement review needs an independent reader, and none was tested.
Design locked before the run:
sha256 df09a3ce4119Verified 2 November 2026: all 5 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Against Lean, mmk's source is 9.1 times shorter than Lean's kernel directory, just under the forecast 10, while Lean's olean is 6.5 times smaller than compressed Metamath on one matched proof family and a Rust Lean kernel builds clean in 51 s against about 170 s for buildc plus mmk, and the forecast 100 times faster build stays untested
Null or refuted Apples and oranges by design; reported so the forecast is scored.
Limits One generated family of repeated proofs, which favours Lean's sharing. Nanoda is a different kernel for a different logic and I read its 51 s as including crate dependencies, not separated or checked here. Metamath's substitution check is far less work than type checking with reduction.
Design locked before the run:
sha256 df09a3ce4119Verified 2 November 2026: all 6 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Adaptive rate gives 1.18 of the best fixed code across noise and 1.29 on a step trace, .98 with the same noise shuffled in time; a locked toy never ruptured, and post hoc a lost control state stops the rate returning at .33 from peak .05, which a beacon repairs at a cost in lasting loss
Null or refuted Toy with a post hoc key result.
Verified 2 November 2026: all 2 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Short generators read as random as system random to every compressor, gain -.00015, a gap near 8.7 bits, while Berlekamp-Massey flags xorshift alone; generator length says nothing about gain, Spearman -.06 and not above shuffle
Null or refuted Known in principle; a bug in three rows.
Verified 2 November 2026: all 1 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Lower layout strain goes with smaller compressed adjacency, Spearman .95 over all steps and .49 among converged layouts, ranging .06 to .96, and a relaxed layout only ties cheaper orderings overall, winning on meshes and losing on trees; the kill fired on my own backward sign
Null or refuted Sign error in the lock; result needs a clean relock to count.
Verified 2 November 2026: all 0 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Eleven format families won 9 of 11 plants built on their own geometry, shuffled 1.9, and packaging moved a family as far as geometry did: wavelet won no real set with raw int16 and beat PNG with zigzag planes
Null or refuted Planted confusion, partly by construction.
Verified 2 November 2026: all 3 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Measured spectra: own dSCF first ionization MAE 0.27 eV on Z<=54 and 0.56 on Z=55..103, but frozen-ion levels lost to a bare Rydberg formula on held-out C, N, O levels, 0.0141 against 0.0057 eV, with a fitted defect at 0.0024, and overbound the lowest levels by about 1 eV
Null or refuted Kill fired; known physics in the main; the null on frozen-ion levels is the reportable part.
Design locked before the run:
sha256 38c422735c12Verified 2 November 2026: all 8 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, one vendor, self-graded. No outside reviewer has read it. -
A time reversal mirror on a real chiffchaff note focuses with amplitude N^0.97, while contrast grows only N^0.15 and saturates near 7, and the forest patch widened the spot to 1.3 to 1.4 of open air against a forecast narrowing
Null or refuted Scatter gain did not appear at this scale.
Limits One forest draw, 4.5 m 2D patch, hard trunks. The width reversal is unexplained. Derode et al. comparison from memory.
Verified 2 November 2026: all 5 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Bird notes through a simple forest channel sit at percentile 0.61 among surrogates of equal band and envelope, 7 percent of species at the 90th, a linear chirp beats real by 0.004, and pitch follows mass with slope -0.32, so nothing here supports a claim of optimisation
Null or refuted Near-chance result on a modelled channel.
Limits Convenience sample, lossy codecs, a modelled channel with no measured forest, masses from memory. Above chance does not mean selection.
Verified 2 November 2026: all 5 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Three time-lens and temporal-boundary instruments lose or barely match what exists: recorder 5.5 ENOB against a 6 ENOB job, magnifier 0.63 ps against 300 fs at M 50, converter -8.9 dB behind a mixer by 1.9 dB
Null or refuted No instrument worth a design note on this job.
Limits Competitor numbers from memory.
Verified 2 November 2026: all 7 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Golden sampling of a time lens gives no gain in one dimension: lobe -0.65 dB against about -12 dB for the 2D spiral, and a quiet edge of 0.27 that falls below its own shuffled gaps at mean 0.50
Null or refuted A forecast advantage failed.
Limits One taper, one threshold.
Verified 2 November 2026: all 5 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Other trend readings did worse and the rule added nothing on the sensor and frame toys: trend only stable 0.66, predicted exit 0.69, local Lyapunov 0.72 against 0.79; at real bounds Marechal lambda/14 never fired, lambda/100 fired the same with or without trend, and at 16.67 ms over-budget moved 2.8 to 3.3 percent
Null or refuted Honest nulls in toy models.
Verified 2 November 2026: all 9 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Model readers refused 0.96 of false maps on full cards and 0.83 from plots and domain names alone, with the gap resting on two of twelve cards, so a map card added little for these readers
Null or refuted Small and leaky.
Limits Models, not people; names leak; twelve cards.
Verified 2 November 2026: all 2 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Of 20 maps with a stated held-out test, 6 exact controls passed as homology, 11 broke at a named point and 3 stayed analogy, and none of the should-break or candidate maps became homology
Null or refuted Mostly known answers in one typed form; the value is the kept failures.
Limits Thresholds mine and one flips status at twice; bits lost are placeholders for four maps; memory values marked; exact maps are textbook.
Verified 2 November 2026: all 4 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Two-loop rule, toys: event-triggered recalibration applied 7.6 of 120 at Strehl 0.9965 against 0.9976 always and 0.9954 random, after locked bounds never fired; a frame budget controller cut over-budget 21 to 2 percent with quality falling 0.135 on an invented model
Null or refuted Invented frame model, mild drift, post hoc frontier.
Verified 2 November 2026: all 8 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Stress that builds with time at load reproduces the size of Miner's failure and gets the sign wrong: high-low 1.10 to 1.13, low-high 0.73 to 0.87, where metals are reported the other way; order sensitivity is real in six of eleven toys and no spacing ridgeline appears
Null or refuted Post hoc after three faults, toys only.
Verified 2 November 2026: all 4 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Reef and granny split by writhe and Jones with Alexander blind; weave interlace orders as published rigidity; moire and D65-to-A metamer pairs check out on synthetic reflectances, median Delta E 0.89
Null or refuted Checks with stated gaps.
Limits No weave numbers, no real dye spectra, mu from one table; rod hit its iteration cap.
Verified 2 November 2026: all 1 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). No outside reviewer has read it. -
A discrete chain matches Cayley's a and losses, the catenary, a folded chain beyond g and a standing shape on a moving ring; fountain tension law holds, and height carries an offset from link size in my solver
Null or refuted Mostly textbook.
Limits Fountain not a test of pot reaction, mechanism disputed in print.
Verified 2 November 2026: all 0 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). No outside reviewer has read it. -
Linking, writhe, Calugareanu, modes on a Mobius strip, vortex density, helicity, Kelvin, Artin, entropy, folding, SSH, Chern and persistence pass on our kernels
Null or refuted Textbook checks.
Limits Checks only. Local pair creation of charges not shown.
Verified 2 November 2026: all 0 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). No outside reviewer has read it. -
The failed paraxial knot was unnormalised modes: the same coefficients on normalised Laguerre-Gauss modes close to a trefoil, Alexander and Jones right, and derived coefficients at scale 2 leave no open lines
Null or refuted Fix of own bug.
Limits Coefficients from memory of Dennis et al; old tracer not tested separately.
Verified 2 November 2026: all 1 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). No outside reviewer has read it. -
Conway and Kinoshita-Terasaka agree on Alexander, Jones and Khovanov ranks over F_32003 and F_2, a null of four readers; a map count to L2(7) separates them, 13 against 7
Null or refuted A known result rechecked with a small separator.
Limits Separation by group maps is known (Riley, Wada); L2(7) count not found, not novel.
Verified 2 November 2026: all 3 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). No outside reviewer has read it. -
An elastic plectoneme at fixed radius gives writhe share 0.69 against the published 0.72, but share runs from 0.80 to 0.56 across radius 4 to 7.5 nm, so agreement rests on the radius
Null or refuted Textbook check.
Limits Radius from memory; Boles via summaries; no thermal or excluded volume.
Verified 2 November 2026: all 6 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). No outside reviewer has read it. -
Rendering the gap between two observers shows how seams end and adds little else: controls pass, noise gives largest gap 1.6 sigma against 4.9, and the field looks alike on noise and on H0 while z(t) separates them; who moved divides outcomes only circularly
Null or refuted A picture on a chosen set.
Verified 2 November 2026: all 2 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Light scattering and cipher diffusion share one measured property, spread that saturates with depth; they do not share single-flip avalanche, which needs modular wrap or nonlinearity in the output, since a linear scatterer's field differential is exactly input independent and its flip rate falls as N^-0.47
- Forecast against outcome
- in FINDINGS
Null or refuted A table to answer a word question; toy cipher and no noise.
Limits Analogy table; the shared word covers more than the measurement
Verified 2 November 2026: all 1 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code; self-graded, self-graded. No outside reviewer has read it. -
In 36,000 counted tries a twisted two-layer stack did no job in a folded pancake that a plain two-plate stack could not: both reach 0.249 of the 0.25 limit across 450 to 650 nm, with ghost 5.9e-4 against 7.0e-4
- Forecast against outcome
- in FINDINGS
Null or refuted A clean null at 36,000 counted tries, in an ideal Jones model.
Limits Accepted physics; a null for twist
Verified 2 November 2026: all 9 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code; self-graded, self-graded. No outside reviewer has read it. -
The gap as an object: white light recovered the gap distribution across three decades to the limit of my table, FTIR leads near contact and loses range fast, a capacitance reading falls apart at 1 percent parasitic, and squeeze-film loss leaves acoustic resonance no reading below a few um; fold layers blur at the Batchelor count
- Forecast against outcome
- in FINDINGS
Null or refuted Readouts idealised; errors near table floors; systematics added second.
Limits Accepted physics for the readout laws and Batchelor; engineering rows are idealised
Verified 2 November 2026: all 1 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code; self-graded, self-graded. No outside reviewer has read it. -
Folding beats a straight system per volume only for a Cassegrain, and only while surface form stays under 0.06 waves rms; a periscope never wins on volume, and overlap of passes needs a degree of freedom that costs light
- Forecast against outcome
- in FINDINGS
Null or refuted A budget set by stated form and reflectance; most forecasts wrong and the control weak.
Limits Accepted physics for 0.25 and the volume sweep; budget numbers are unproven engineering, set by my stated form and reflectance
Verified 2 November 2026: all 1 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code; self-graded, self-graded. No outside reviewer has read it. -
A table of 90 directed pairs over ten quantities and 55 tuples of three kinds, with 23 coefficients read today and the remainder computed, cited or from memory
Null or refuted A partly verified map.
Limits Most figures unverified; the node counts map my catalogue.
Verified 2 November 2026: all 3 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Nine of twelve transducer forms have an interior band about a decade wide, four of them lose it under an amplitude metric, and published points landing in band barely discriminate, 9 of 10 against a shuffled q95 of 8
Null or refuted A map on memory parameters.
Limits Textbook forms from memory; one ratio at a time; operating points mostly memory.
Verified 2 November 2026: all 3 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
A chain beats the best of three direct routes in 10.7, 7.3 and 4.9 percent of random draws at 2, 3, 5 stages, passive chains in 4.5, 0.7 and 0, and Friis holds within 1.4 percent; on device rows from memory it wins with a gain or phase stage first, and direct wins at high signal and on range
Null or refuted Known answer checks, and memory device rows.
Limits Random draw ranges are mine; device figures memory; cost, size, bandwidth left out.
Verified 2 November 2026: all 10 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Sampling can miss metamers entirely: for intensity readings at N 6, random signals show none at 7 to 10 readings, while constructed signals give exact metamers in 20 of 20 at 10, and none exist at 11, as Balan's 2N-1 states
Null or refuted A known theorem checked; the lesson is about sampling.
Limits One N; generic Gaussian readings.
Verified 2 November 2026: all 9 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Public Auger arrival directions show no cubic pattern across five energy cuts, BH p 0.85, at low sensitivity: planted power needs epsilon 0.2 at 8 EeV and 1.0 at 32 EeV
Null or refuted Null with stated power.
Limits No field deflection model; only public data; the null bounds a large pattern alone.
Verified 2 November 2026: all 5 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Chains, couplings, biomarkers, device chains and ulcer structure graded: hair cortisol sign turns with time, H. pylori eradication prevents 80 percent of duodenal recurrence by RCT, stress carries an observational fraction near 29 percent (7 to 50), and body-based treatment uniqueness stays unsupported
Null or refuted A survey, useful internally
Limits Abstract level only; many clinical numbers marked from memory.
Verified 2 November 2026: all 4 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). No outside reviewer has read it. -
One slow-state form clearly beats the domain form in one of five domains; embedded history stays a family resemblance in math, and a fit cannot test shared mechanism
Null or refuted Honest null on five small cases; G2's threshold came after G.
Limits That the fields share no mechanism or that analogies are empty. Palaeomagnetic remanence, reaction wood and residual stress not tested.
Verified 2 November 2026: all 0 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). No outside reviewer has read it. -
One toy truth gives influx up 1.43 and clearance 0.97, and more space alone raises influx 1.64, so the glymphatic readouts answer different questions; flow of 3 um/min separates at 0.92 and 1.00 by tracer size
Null or refuted Clarifies a dispute's terms only; a toy.
Limits Which side is right. Does not reproduce a marked fall. One dimension, no perivascular space, D known.
Verified 2 November 2026: all 6 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). No outside reviewer has read it. -
No feature visible when a remark was made earns a ranking: fame 0.51, quantitative 0.52, pooled passes for testable 0.70 and field age 0.68 come from one all-refuted stratum I created, and every outcome was recalled
Null or refuted Null plus a self-found artefact.
Limits Features have no value: the sample is survivorship-shaped and one coder recalled every row.
Verified 2 November 2026: all 4 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). No outside reviewer has read it. -
OEIS recurrence conjectures: 1 of 213 fails on its b-file, A031740 at n 65, which the entry already marks false since 2016; no new counterexample
Null or refuted Cheap, mostly confirms.
Limits Passing lines true. Fitted lines pass by construction, and only 5 percent of conjecture entries parse.
Verified 2 November 2026: all 4 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). No outside reviewer has read it. -
Searches catch every historic counterexample tried and find none new: Lander and Parkin, Polya at 906150257, 5777 and 5993, 18 ternary exceptions, while li stays above pi to 1e10 and powerful numbers to 1e16 show no triple
Null or refuted Reproduction.
Limits Anything beyond the range, and that the range is new.
Verified 2 November 2026: all 6 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). No outside reviewer has read it. -
With a simulated chooser, tracking drift only pays when drift is large: against a fair static fit -0.05, 0.00, +0.65, +1.60 nats per pick for still, slow, fast, switch; the locked baseline was biased and its verdicts void
Null or refuted A formula standing in for a person, and a post hoc rescue.
Verified 2 November 2026: all 4 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
The bench question cannot be answered by updating through the build: 2 of 28 records in each transcription have two or more dated milestones, so no update ran, and even 100 cases would give power 0.22 under a generous model
Null or refuted A data gate, stating what cannot be claimed.
Verified 2 November 2026: all 4 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Filtering buys chaotic horizon at a log rate: median 2.15 against 0.83 time units at noise 2 on Lorenz 63, a median gain of 0.93 under the 1.99 the log rule gives, while the locked RMSE measure failed
Null or refuted A textbook relation rechecked; the locked measure was poor.
Verified 2 November 2026: all 6 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Earth's core and Sgr A* share ray geometry and no law: the same tracer gives a PREM P shadow of 98.4 to 145.1 degrees and a hole's sqrt(27), but 0 of 4 dimensionless predictions agree within 1 dex (misses 9.0, 4.5, 7.6, 1.5 dex), and a Luneburg lens obeys the shared ray law just as well
Null or refuted A clean null on a tempting analogy, with its rule written first. Memory values for heat and timescales. The rule's reach is the four ratios I chose.
Design locked before the run:
sha256 46d1909ef167Verified 2 November 2026: all 10 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). No outside reviewer has read it. -
A single split of noise and field bias against signal beats the base rate at ranking failed replications (Brier 0.222 against 0.256) and adds nothing to the z value alone (0.218); a field's share of just-significant originals ranks fields (rho 0.79, 8 fields)
Null or refuted Claim-level split adds nothing over z, already known from F172's logistic. Field rank is a lead on eight fields. Contrary pass is confounded.
Verified 2 November 2026: all 5 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). No outside reviewer has read it. -
Worked shapes on our code match known answers for caustic exponents, vortices, Fresnel edge, Gibbs, Burgers and uncertainty; a trefoil is checked as a field, and my paraxial knot failed
Null or refuted Textbook checks; instruments for later work.
Limits Anything new; they are checks of our kernels. Several matched only after fixing my test code; see agent report summary in RESULTS and code history.
Verified 2 November 2026: all 0 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). No outside reviewer has read it. -
Correcting phone IMU noise by 41 and 13 times moved the moving-observer result very little: camera plus inertial 0.20 m to 0.22 m, inertial alone 1.18 m to 1.15 m, and the microphone stays unhelpful as a hard constraint
Null or refuted Honest null: the recorded conclusion survives a 41 times wrong constant.
Limits Die specification, no measured phone; bias unchanged; simulation only.
Design locked before the run:
sha256 4d9acc9545b2Verified 2 November 2026: all 6 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
One expression, linewidth ratio times a time-scale match, does not predict tuned actuation across classes: four rows of ten hit within a factor 3 but a shuffled pairing does as well, and the time factor makes predictions worse
Null or refuted Memory values and a forking-path risk.
Limits Most values from memory; I chose each row's competitor and linewidth. Physics and published figures only.
Verified 2 November 2026: all 1 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Bound first, four of fifteen recombination rows stop and nearly all that pass are published; only joint acoustic and optical non-line-of-sight gave no direct hit in one search, and it showed no gain in simulation
Null or refuted Order-of-magnitude bounds on memory parameters.
Limits Memory parameters, Fisher at a point. R02 and R05 follow directly from values I picked. A one-search miss is weak evidence of novelty.
Verified 2 November 2026: all 0 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
A transduction table of 121 cells gave no ranked leads: 1 empty cell is unexplored, 12 physical, 10 collection gaps or chains, and only 36 of 112 effects were opened or matched by code
Null or refuted Partly verified table; useful as a map, not a result.
Limits That a cell marked empty is empty in nature. Literature workers relied on snippets though asked to open pages.
Verified 2 November 2026: all 6 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
In any loop through a person the exit is the narrowest stage by about 2.6e5, and raw sensor rate is not message rate; figures from cited sources, in differing protocols, good to an order of magnitude
Null or refuted A table from others' figures, with unaligned protocols.
Limits A measured rate on one protocol. DAC, ADC, gaze, optogenetic and electrode stimulation rates not found. Retina is an extrapolation from guinea pig.
Verified 2 November 2026: all 1 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
A null-space complement finds what an instrument misses and rediscovers Zernike unprompted, but ties plain Bayes design at picking additions and loses by 1.66 in the eye, where the gap is nearly everything; kept as a diagnostic
Null or refuted Honest null on the design claim, with a plant and a kill; the Zernike rediscovery is a known answer, not a discovery.
Limits Anything for nonlinear instruments beyond the one Zernike search. A plant in the null suits a complement by construction. Linear model with known prior. F152's corpus gap ranking is a separate thing and stays failed.
Verified 2 November 2026: all 1 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). No outside reviewer has read it. -
A designed random medium beat a clean instrument in no Gaussian budget tested: noise helps only a threshold that cannot move, a diffuser won no budget against a lens, disorder gave a lattice no memory, and a clip with no noise wins in impulsive ambient
Null or refuted Honest null from a locked design with plants and a kill, on a stated budget that omits the diffuser's main case.
Limits That diffusers are useless: size is their practical case, not scored; the lens has perfect focus and a known disk; points were sparse; eight or more planes or an unknown PSF not tried. One lattice rule.
Verified 2 November 2026: all 0 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). No outside reviewer has read it. -
A named decisive observation does not predict survival on unfamiliar challenges: AUC 0.50 on 66 cases, with a scorer that catches planted vagueness perfectly, 0.69 on recognised cases against 0.50 on unrecognised, so hindsight is what made it look useful; writer's plain probability does better on replication, 0.66
Null or refuted Preregistered null on 66 cases with a scorer that catches plants.
Limits That naming an observation has no use for people, since it may still direct effort, which this does not test. One model family throughout. FReD descriptions are terse and judged on form. A failed replication is not a false claim, and retraction often concerns conduct. Writer's p_self on FReD is one positive here, not a pass on any verdict, and recognised rate 0.30 suggests part of it is memory.
Verified 2 November 2026: all 5 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Public goods check: my "reproduction" of the cheater result was self-benefit. With goods reaching others only, producers collapse through b/c 100 and reach 0.48 at 400, so space alone gave no cooperators at sensible cost under this update rule
Null or refuted My reproduction was self-benefit; corrected null under one update rule.
Limits Anything about microbes. One update rule with invented numbers.
Verified 2 November 2026: all 3 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Long windows hit the date, and do not dodge frost better than a calendar: chill model r 0.25, three-day window 0.11; frost exposure chill 0.202, calendar 0.206, three-day 0.268, observed trees 0.221
Null or refuted No frost benefit from long windows on exposure; damage unseen.
Limits Damage: exposure at -2.2 C in 21 days is not a burned bud. No photoperiod. Volunteers, east-weighted. Selection acts on damage this cannot see.
Verified 2 November 2026: all 6 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Mycelial graphs lose to same-length nulls on transport and robustness in 7 species; thick cords follow paths (rho 0.32, cord formers about 0.5), and only Phanerochaete cuts resistance clearly; junction exponent median 2.84 against 0.47 for random triples
Null or refuted Mycelial graphs lose to nulls on transport and robustness, small colonies.
Limits Forest mycelium, wood-wide networks or transfer. Small colonies only (files under 12 MB), straight-edged nulls carry an advantage, betweenness stands in for flux, thick cords may be old cords.
Verified 2 November 2026: all 5 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Fixing my own CMA-ES bug reversed the headline: plain beats bio-labelled on shifted known optima, CMA-ES rank 2.0, best bio ACOR 3.14, and SMA and GWO lose 12 and 13.6 decades when the optimum leaves the origin
Null or refuted Refutes the bio-labelled edge after my own bug fix; default parameters only.
Limits Other functions, dimensions or budgets. Parameters are defaults.
Verified 2 November 2026: all 4 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Rod and grain do not share one threshold on these data: rod is single-hit on the mean curve with n 2 rejected by 82 BIC, the rod variance channel cannot tell n apart, film fails on variance and its fit is no better than shuffled, and the human n runs 4 to 10 or 6 to 12 depending on the dark term
Null or refuted Preregistered refutation of one shared threshold.
Limits anything about mechanism downstream of a rod. Trial count 66 is a midpoint. One rod's single-hit does not fix how many absorptions a person needs.
Verified 2 November 2026: all 6 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
A moving observer: inertial sensing halves the median path error and the microphone adds nothing as a hard constraint; sensory conflict is detectable; orthogonality costs little until skew is large
Null or refuted Locked fusion test killed on means; the inertial gain is the usable part.
Limits Labelled corners, axis-aligned rooms, a rough start given, so loop closure and association are outside this test. These runs measure this estimator and say nothing about why anatomy has its form.
Verified 2 November 2026: all 0 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
A lens and a microphone fitted through one room's modes did not beat either alone: kill condition met in two designs, so there is no gain for the mode basis to claim, and first-order echoes gave the only gain
Null or refuted Preregistered kill met in two designs; simulation.
Limits Simulation, 2D, one fitter, heading given, the echo arm's forward model is simpler than its truth would be. A stronger fitter or a trained one could change every figure. The addendum is a second look at the same rooms, so the echo gain needs a fresh seed.
Verified 2 November 2026: all 0 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code plus agent; self-graded, self-graded. No outside reviewer has read it. -
Von Bertalanffy growth across trees and repos: VB is the best of three saturating forms in both media (hold-out wins 0.82 of 66 tree series, 0.68 of 19 repos), but it wins by collapsing toward a straight line, and neither medium shows deceleration beyond linear
Von Bertalanffy beats Gompertz and logistic on hold-out tree and repo growth only by collapsing toward a straight line, and neither medium shows deceleration beyond linear.
- Sample
- 66 tree series at 3 sites; 19 repos
- Result
- none; T2 median -0.00005, R2 -0.00002
- Controls
- working: shuffles C1, C2; the verdict rule itself gave a false success
- Forecast against outcome
- T1 right, T2 wrong, R1 wrong, R2 right, X wrong
- Known answer
- none
Null or refuted Honest refutation of the study's own cross-media pass: X passed through a rule that the linear nest makes trivial. Three sites without pith and 2 to 6 month repo histories limit it, so its value is mainly as a design lesson.
Design locked before the run:
sha256 27f9dfacVerified 2 November 2026: all 5 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code, self-graded. No outside reviewer has read it. -
Childhood trauma and amygdala-vmPFC resting connectivity: no meaningful association in 187 adults; effects between 0.10 and 0.20 stay open
In 187 adults on open data, childhood trauma scores show no meaningful link to resting amygdala-vmPFC connectivity, which excludes effects of |r| 0.20 or more and leaves 0.10 to 0.20 open.
- Sample
- n 187 of 241 (17 motion, 16 no CTQ, 21 missing subscale)
- Result
- r -0.006, 95% CI -0.152 to 0.140, 90% CI -0.129 to 0.117
- Controls
- working: 10,000 permutations, 47th percentile; 15 exploratory sensitivities r -0.081 to 0.081
- Forecast against outcome
- decision rule set before data; outcome no meaningful association
- Known answer
- none (no positive control for connectivity)
Null or refuted A clean registered equivalence null on real CC0 data with a shuffled control. A larger null already exists, so it adds a replication and a specific equivalence bound. Power 0.77 at the bound, modest.
Limits Anatomical mask, a low-pass FD rule outside its validated single-band data, retrospective self-report, and no positive control showing the pipeline finds a known effect.
Design locked before the run:
sha256 4d4681b9Verified 2 November 2026: all 21 numbers on this card appear in the study's own ledger entry (checked by the site's build,
tools/experiments/select.py). The study was scored by code, self-graded. No outside reviewer has read it.