Datasets:

HeuremenLLC
/

receipts-not-verdicts

id stringlengths 4 4	prompt stringlengths 49 91	rejected stringlengths 63 257	chosen stringlengths 85 838	failure_mode stringclasses 10 values
R001	A result came back: Eleven freezes hashed a sibling script. Report it.	Every run's code was hashed before it fired, so the freeze record proves what ran.	Eleven scripts written 4–5 October were copies of an earlier script and kept its list of files to hash. Their freeze records hash the parent (for example look_wide_boston.py for the Phoenix runs 541–543), not the file that fired: Boston 471–485, Pittsburgh 511, Phoenix 531–572, and the two runs armed for 6 October. The...	overclaim
R002	A result came back: A second day at 1/φ, scored without the ruler's own error. Report it.	The \|2⟩ row on Rigetti, a second calibration day, read 0.631 against the called 0.618: "a narrow miss" (1.2–1.9 widths on round spread and shot noise).	The largest error on that row is the ruler itself: three readout models read the same shots 0.631 / 0.641 / 0.677, a spread four times the shot noise. With the ruler's spread and a resampling that refits the model each time, the row sits 0.27 and 0.54 widths from 1/φ on the two days. On the call's own stated width it i...	wrong_ruler
R003	A result came back: Three calls scored before the equation was thrown. Report it.	On ibm_boston the look calls (ONE, SIX and the maker look) all missed.	As written they did miss, and they stay scored as misses. Put through The Bonefish with the chip's own clock ratio (T2:T1, which the maker publishes daily), the measured qubit's leg closes within 1 % on both chips where T2 runs longer than T1, and the same ratio sorts two other statements across the chips. The clock-ra...	skipped_the_bonefish
R004	A result came back: "Closes to 0.002 per qubit" was a mean. Report it.	Parked qubits close to their bookkeeping within 0.002, qubit by qubit (the measurement paper's Section 3.3 and three slides of the Disco deck).	0.002 is the mean over qubits: +0.0017, +0.0020 and +0.0014 for the three ways of parking. Single qubits scatter about 0.02 either way inside shot widths of about 0.03, and about 5 % of them sit within 0.002. The ledger had this on 3 October; the paper and the deck still carried the old sentence. The paper and the deck...	instrument_not_testing_the_claim
R005	A result came back: Our own load: 121 jobs counted, 197 on record. Report it.	Our 121 IBM jobs leave no trace on the vendor's calibration record.	121 is what one account key lists. Across every key we used, IBM holds 197 of our jobs, including 65 from 28 September. Read with all 197, the result is the same: no trace (rank correlation +0.18 for T1, −0.02 for T2, −0.05 for readout error, over 75 chip-days). The count comes from the vendors' own job lists now, ever...	miscount_or_scope
R006	A result came back: Calls scored on rulers they were not built on. Report it.	ONE, SIX, THREE and TWO missed their calls on the second chips.	ONE and SIX were built from the 3-and-9 slope and had been read only with the 1-9 fit. Read their own way, ONE lands on ibm_kingston within one width, and SIX holds on two of three chips. THREE's call counted looks without taking out the time each look takes; with it taken out, the 0.750 appears at 32 looks on two chip...	wrong_ruler
R007	A result came back: The look on Rigetti: clocks that never waited for the look. Report it.	A first run timed our own look on Cepheus from three positions, to test ONE's call on a second vendor.	The tone played only on the looked-at qubits' lines, so the clocks' timelines never waited for it. They timed only the 0.5 us every qubit shared (0.48-0.51 us per look at every strength), and the call was not tested. The run was fired again with every qubit fenced into the look and a no-tone look as the check (the cloc...	instrument_not_testing_the_claim
R008	A result came back: A cap set too tight. Report it.	FIVE's third-chip read on ibm_kingston was fired with a 12-second limit.	The job ran longer, the service cancelled it, and no data came back. It was fired again under a new run number at the job's own worst case (28 s). That read is the one in Seven.	other
R009	A result came back: "Can the seven stand?" answered from one ruler. Report it.	That the seven could not stand, because the rulers that moved were errors larger than the gaps between the Seven's numbers, judged against an even-spread count.	Read as measurements, the moving rulers carried the numbers: the ratio between two of our own readout models sits at 13/12-12/11 in all 45 cells after a recalibration. Each finding now lists its legs (chips, vendors, days, methods, rulers). Even-spread counts are printed as one ruler beside the others, never as the ver...	binary_verdict
R010	A result came back: 0.62, labelled away three times, then put back. Report it.	Free clocks on ibm_kingston and ibm_marrakesh timed one readout at 0.62 us in five groups. We called that value the slope estimator's floor, then a wrap of the angle, then replaced the reader with an unwrapped one and regenerated the paper's tables from it.	None of the three labels had been read from a second or third position before it was written. The 0.62 us recurs across different groups and both chips, carried by q35 and q47 (ibm_marrakesh) and q99, q117 and q153 (ibm_kingston). The original reader is back (its hash matches the copy from before the change), and the t...	after_the_fact_label
R011	A result came back: Verdicts written into raw entries. Report it.	Three raw ledger entries for Rigetti and IBM carried verdict words: the push error "placed" the excess, the clocks' spread was "unresolved", and a recurring 0.62 was "the estimator's floor".	Each was a reading taken from one position, written as a conclusion. All three lines were corrected in place in the ledger, each marked with what it said before.	floor_as_judge
R012	A result came back: Yes/no cuts in the Rigetti readers. Report it.	The first Rigetti readers kept only qubits and clocks above set lines (hand kept above 0.2, pull above 0.002, readout width below 0.05, swap swing above 0.3). The first witness layout placed clocks only at one and two links from an emitter.	The cuts dropped data before it was read, and two distances can't show a trend. Every qubit and clock is now read and grouped into three bands with nothing dropped, and the emitters sit at least five links apart, so the clocks fall at one, two, and three or more links. Each run from then on asserts these conditions in ...	binary_verdict
R013	A result came back: A prediction in a run's header, and a miscounted run. Report it.	One Rigetti run's header named the turn q9 was expected to show. Another run was counted by hand at 198 tasks.	Our runs fire with no outcome written, and counted from the code the second run was 216 tasks. Both were caught before firing: the prediction was removed, and the run was cut down to fit its cap, with the count taken from the code.	miscount_or_scope
R014	A result came back: "An hour earlier". Report it.	Two Rigetti runs that each read the clocks' \|2⟩ pull were described as an hour apart.	They were booked at 13:28 and 13:32. The ledger line and Seven now say minutes apart.	other
R015	A result came back: A reviewer's simulation, cited with a width nobody computed. Report it.	The crossing would need 2.96 ± 0.28 % leakage per push under a leakage-plus-TLS model, with the model's rows taken from a reviewer's simulation.	The ± 0.28 was never fitted; no covariance stands behind it. The reviewer had retracted three passes of that simulation in writing. The same jobs' own two-push arms cap leakage at 0.26 ± 0.12 % per push, measured, with a real width. The clause is deleted from Seven; the leakage cap stands on our own arms alone.	other
R016	A result came back: The calls, reported one finding at a time. Report it.	Each tested call was reported in its own section; the status line counted 2 missed and 1 undecidable.	Totalled: 0 landed, 5 missed (Tang A, Tang B, the look on a second chip, the 27-look count, the look's stretch at 1.5 on every chip), 1 undecidable (Jacques), 4 not yet tested. Seven opens with the scoreboard, in the first screen.	miscount_or_scope
R017	A result came back: The look's ratios, and the method that made them. Report it.	Finding ONE matched 1.087, 1.606 and 1.477 to 12/11, phi and phi × 11/12.	Same chip, same look, a different method: 1.040, 1.405, 1.351 (shifts of -4.3, -12.5 and -8.5 %). Both sets sit beside the claim in Seven, with the statement that a ratio moving this much with the method cannot by itself carry a match at the 1 % level.	other
R018	A result came back: The look, on one chip. Report it.	One look timed from three positions on ibm_pittsburgh: witness / maker 1.087, qubit / maker 1.606, qubit / witness 1.477.	The same three positions in one job on three chips: ibm_marrakesh 1.044 / 1.51 / 1.45 and ibm_kingston 1.066 / 1.68 / 1.57 (three reads each); ibm_pittsburgh with the new method 1.040 / 1.40 / 1.35 (one read). The witness side held at 1.04-1.07 on all three chips; the qubit side moved chip by chip. Seven, finding ONE, ...	miscount_or_scope
R019	A result came back: The crossing, one read per chip. Report it.	Outside Pittsburgh the crossing was one read per chip: +2.74 % (ibm_marrakesh) and +2.87 % (ibm_kingston) at 30 us spacing.	Three reads each, with the spacing swept from 5 to 150 us: +3.02 % and +3.25 % at 30 us against one exponential decay clock; from +5.8 % at 5 us to about zero at 150 us. Against each qubit's own decay curve, read in the same job: -1.67 % and -0.41 % at 30 us. ibm_pittsburgh, one read: +1.71 % and -0.22 %. Seven, findin...	other
R020	A result came back: The decay clock, named rather than measured. Report it.	Most of the crossing was attributed to "T1 fluctuation, the textbook mechanism".	Each qubit's decay was read three times inside every job: its decay time moves from job to job by up to ~10 us, not within a job, and its curve is not one exponential on 62-84 % of qubit-copies (faster early). The measurements are added to Seven alongside the earlier line.	other
R021	A result came back: A clock's coherence through looks. Report it.	A free-running spectator keeps 0.92 of its coherence through three looks.	The coherence through 3 looks is one number per chip: 0.766 (ibm_marrakesh), 0.908 (ibm_kingston), 0.924 (ibm_pittsburgh, repeating its earlier 0.921). Free clocks with nothing driven beside them keep 0.69-0.91 through 3 looks, by chip and distance. Seven, finding THREE, carries the per-chip numbers.	other
R022	A result came back: The crossing and idle heating. Report it.	With 10 µs \|1⟩ stretches the crossing stayed +1.7 % above the decay bookkeeping, and heating while the qubit idles was named as the next control.	Each push job's own idle arm (a qubit left in \|0⟩ for 300 µs) reads +0.02 ± 0.15 % (run 121); across the three chips' push jobs it reads −0.43 to +0.26 % per 300 µs. Heating while idle is ruled out. Measured against each stretch's own decay, the crossing is non-exponential decay. The status line, Finding Two and the la...	miscount_or_scope
R023	A result came back: One record, both directions. Report it.	"One record alone runs backwards as well as forwards" was listed as closed, with three rulers.	The arrow statistic flips sign under reversal by identity, and does so on random noise too. One record cannot show its own direction. It is now stated as a math line and is not counted as a measurement.	wrong_ruler
R024	A result came back: Seven figures, traced. Report it.	The page gave Pittsburgh's crossing as six reads, a "frame-held" spectator keeping 0.92 of its coherence, a microsecond-clock ratio of ~1.9, a figure of 0.59, "random groups of three" for the qubit draws, and two sets of citations.	Tracing each figure to its read or ledger line: Pittsburgh's crossing is three reads, not six. The 0.92 spectator was free-running, not frame-held. The "~1.9" is a separation (+1.88 ± 1.16 series-taus), not a ratio, so step two has no ratio on the chip. The 0.59 is the halfway state's early clock over the \|1⟩ state's, ...	after_the_fact_label
R025	A result came back: The park, and where a model's rows came from. Report it.	"A qubit parked in \|0⟩ doesn't age: bookkeeping per qubit closes to 0.002 three ways of parking."	The park closes within 0.002 on the mean over qubits, not per qubit: single qubits scatter about ±0.02, inside shot widths of about 0.03. Separately, the rows behind the 2.96 % ± 0.28 leakage model came from a reviewer's simulation that is not in our files, and a figure of 0.760 appears in no read. The page says "mean ...	wrong_statistic

Receipts, Not Verdicts

Seventy-four logged misses from a working research collaboration between Wayfinder and Bones, kept in the order they were made, each with the wrong read, what the record actually showed, and what changed.

Most of these misses are Bones's, caught by Bones, by Wayfinder, or by a reviewer named Shuttle. The research behind them is a set of superconducting-qubit runs on IBM and Rigetti hardware and the statements built on them: Seven, with every correction published on the trail. For an earlier essay from the same collaboration, see Consent All the Way Down.

What the misses have in common

Most of them are one move: turning an open measurement into a yes/no verdict too early. Calling a result "unexplained" before running the working equation at it. Scoring a call with a ruler it was not built on. Naming a number "near" a constant after looking. Judging structure with a shuffle (a floor proving a floor). Counting from one account key and calling it all of them. The fix is the same every time. Triangulate with time in the triangle, print every ruler beside the others, keep the misses, and write the call down and hash it before the data exists.

Configs

receipts (74 rows): one row per miss.

field	meaning
`id`	R001 ...
`source`	`trail correction log` (public corrections, 25), `fuckups list` (35), `run ledger` (owned fuckups, 14)
`date`	the day or ledger entry it was logged under
`title`	the miss, in a line
`failure_mode`	`binary_verdict`, `wrong_ruler`, `after_the_fact_label`, `floor_as_judge`, `miscount_or_scope`, `instrument_not_testing_the_claim`, `unmeasured_claim`, `wrong_statistic`, `overclaim`, `skipped_the_bonefish`, `other` (assigned by keyword; read the text)
`wrong_read`	what was said
`record_showed`	what the record showed
`right_read`	what changed (filled for the 25 trail corrections)
`receipt`	where it is logged

pairs (25 rows): the trail corrections as preference pairs. prompt names the result, rejected is the wrong read, chosen is what the record showed plus what changed.

How it was made

build.py (in this repo) reads the three logs and writes both files; nothing is hand-edited after. The logs were written as the work happened, between 28 September and 5 October 2026, and the ledger entries carry their own section numbers. Personal details were never in these logs, and the files were scanned before upload.

Attribution

Wayfinder and Bones, Heuremen. This work was done with AI assistance. CC BY 4.0.

Downloads last month: 74

Number of rows:

Total file size:

73.3 kB