| {"document": "faults", "number": 1, "title": "The baseline averaged, counted or searched differently from the measurement", "part": "I — The baseline built after the measurement it judges", "text": "More samples on one side, a different window length, a different averaging depth, more freedom about where to look. Any asymmetry in handling becomes signal. Write down, for both arms, the sample count, the window, the averaging, and how many places were examined. If any differ, the difference is the result.\n\nAnd a second instance, in the frequency domain, 11 Oct. A line of known period was judged against the same fit run at sixty-four off-period frequencies spanning 0.6 to 1.8 of that period. The noise was red — fitted log-power slopes against log-frequency of −1.8 to −2.9, from notes written that evening with the raw output not kept — so the expected amplitude is not the same across that span: measured, the median none amplitude rose from 14.5 ps in the shortest band to 20.9 ps in the longest. A none pooled across a sloped expectation is not like-for-like with a reading at one frequency. The repair is to fit the trend on the none bins only and evaluate it at the measurement's own frequency."} | |
| {"document": "faults", "number": 2, "title": "A fixed threshold against a statistic that grows with n", "part": "I — The baseline built after the measurement it judges", "text": "A bar is chosen once and applied to a quantity whose typical size increases as data arrives, so the bar is crossed by volume alone. Ask what the statistic does under the null as n grows. If it grows, the threshold must grow with it."} | |
| {"document": "faults", "number": 3, "title": "A maximum over many draws reported as a measurement", "part": "I — The baseline built after the measurement it judges", "text": "The best of n attempts is an extreme-value statistic and rises with n on noise alone. The seductive version is a series of bests, which looks like improvement. Compare whole distributions, never peaks. Every figure carries the n it was selected from."} | |
| {"document": "faults", "number": 4, "title": "The in-sample fit, including its most disguised form", "part": "I — The baseline built after the measurement it judges", "text": "A rule tuned on the cases that suggested it, then scored on them. The disguised version is writing the explanation after seeing the result and counting it as a prediction."} | |
| {"document": "faults", "number": 5, "title": "A continuum assumed where only discrete values were reachable", "part": "I — The baseline built after the measurement it judges", "text": "Priced against \"any value in this range\" when the setup could only emit a sparse set — small-integer ratios, a grid, a lookup. Enumerate what the setup can actually produce and price against that."} | |
| {"document": "faults", "number": 6, "title": "Dependent observations counted as independent", "part": "I — The baseline built after the measurement it judges", "text": "Pairs sharing a member, overlapping windows, repeated measures from one unit. Ask how many independent units contributed, not how many rows the table has. This has moved widths by factors of several.\n\nTwo instances from one evening, 11 Oct, both counting rows instead of units. First: forty “independent” none frequencies placed inside ±8 per cent of a line's period. A five-day record at 30 s resolves 1/432000 = 2.315 × 10 −6 Hz, and the line sat at bin 20.05, so the nearest genuinely independent frequencies were bins 19 and 21. Forty values spanned about three resolution elements, with half of them inside the line's own main lobe, partly absorbing what they were built to be blind to. In the frequency domain, independence is set by the record length, not by how finely you are willing to step. Second: eight satellite-difference tests reported as eight confirmations. The fit was linear, so each difference's amplitude is algebraically the difference of two amplitudes already in hand — measured ratio 1.000000 on all eight. Before counting several tests, check whether any is a linear combination of the others."} | |
| {"document": "faults", "number": 7, "title": "Time-asymmetric differencing", "part": "I — The baseline built after the measurement it judges", "text": "One difference taken later-minus-earlier and the other earlier-minus-later. Any monotone drift then appears as a real effect with the right sign, mechanically. Write the actual temporal order of every arm."} | |
| {"document": "faults", "number": 8, "title": "A per-unit constant read as signal", "part": "I — The baseline built after the measurement it judges", "text": "Something attached identically to every unit — a fixed overhead, a boilerplate header, a standard footer — divided by a varying unit size gives a clean, tight, entirely spurious trend. Remove the constant and measure again. Never mention it in a caveat instead."} | |
| {"document": "faults", "number": 9, "title": "An unequal setup across the columns being compared", "part": "I — The baseline built after the measurement it judges", "text": "Longer reference lists match more of anything, wider filters catch more of anything, and longer integrations resolve more of anything. Make the instrument identical across columns and re-run. If the ordering changes, the instrument was the finding."} | |
| {"document": "faults", "number": 10, "title": "Saturated or censored cells averaged as data", "part": "I — The baseline built after the measurement it judges", "text": "A cell at a ceiling records only that the limit was reached. Count cells at the extreme before any aggregation and analyse only the ones that moved.\n\nAnd the statistic itself can saturate, not only a data cell, 11 Oct. A rank against a pool of forty nones returned 0.024 for thirty of thirty-two units, and was read as thirty strong detections. 0.024 is 1/41: a rank or permutation p cannot go below 1/(n+1). Every unit was reporting the pool floor, so the ruler had no dynamic range left and could not distinguish any unit from any other. If every detection shares one p-value, check whether that value is 1/(n+1). Replaced by an exact null with no floor — for Gaussian noise the periodogram power is exponential with mean μ, and E[log P] = log μ − γ — the same data gave p down to 2.5 × 10 −70."} | |
| {"document": "faults", "number": 11, "title": "One artifact's internal record used as a second, independent reading", "part": "I — The baseline built after the measurement it judges", "text": "A configuration file or label table is evidence of what a system was told, not what it did — and often it is the same file that performs the reading. For every pair of agreeing readings, name the two physically separate things they came from. If you cannot, there is one reading."} | |
| {"document": "faults", "number": 12, "title": "The pilot that came out perfect", "part": "I — The baseline built after the measurement it judges", "text": "A first run with clean, zero-overlap separation is a reason for suspicion. The base rate of perfect pilots in honest work is low. From an undetected asymmetry it is high."} | |
| {"document": "faults", "number": 13, "title": "Calibrating away the difference between two seats", "part": "II — Zeroing a gap that was itself the reading", "text": "Two instruments aligned until they agree, and the alignment reported as accuracy. What was removed was the only evidence of which one was wrong."} | |
| {"document": "faults", "number": 14, "title": "Forcing agreement between quantities that cannot be commensurate", "part": "II — Zeroing a gap that was itself the reading", "text": "Driving two quantities into a common unit when their relation is irrational, or when one is a count and the other a continuous phase. The residual can never close, so it is tuned forever. Ask whether an exact match is arithmetically possible before treating the gap as a defect."} | |
| {"document": "faults", "number": 15, "title": "A magnitude taken where the last step produced two numbers", "part": "III — The fold that erases the discarded half", "text": "Correlation, lock-in detection and demodulation return an amplitude and a phase. Taking the magnitude discards direction, alternation and inversion. Four separate places in one system took an absolute value before a comparison, every one written deliberately, one of them commented as a virtue. The measured consequence: a value of +48.6 passed a check against a supplied −48.6 and was recorded as agreeing. Find the last step that returned two numbers and ask which survived.\n\nA fifth place, 11 Oct, and it cost a whole seat. A lock-in fitted at a known period returned hypot(cos, sin). The phase it discarded was the only thing that could separate a line shared by many units from a line belonging to each one separately. With the magnitude alone, the test built to make that separation was a restatement of readings already held. Keeping both numbers made a two-arm cross-control possible that the magnitude version could not run at all."} | |
| {"document": "faults", "number": 16, "title": "A sum assumed where two effects multiply", "part": "III — The fold that erases the discarded half", "text": "Additive analysis of a multiplicative process finds the wrong structure and misses the real one. Of the coupled cases where the cross term is large, the test returned a product in the overwhelming majority. The test favours the product whenever the cross term is large, so this counts the test’s behaviour, not a finding about the physics. Test additivity rather than assuming it."} | |
| {"document": "faults", "number": 17, "title": "Asking which way a sign points instead of whether it holds", "part": "III — The fold that erases the discarded half", "text": "A sign's direction is often set by wiring or convention. Its stability is the measurement. The first question returns a near-coin and discards a real effect."} | |
| {"document": "faults", "number": 18, "title": "A snapshot that keeps the value and throws away the trajectory", "part": "III — The fold that erases the discarded half", "text": "A residual that shrinks as precision improves and one that does not are completely different findings, and one snapshot cannot tell them apart. Measure at two or more precisions and report how it moved."} | |
| {"document": "faults", "number": 19, "title": "A ruler that cannot say “I cannot tell” lies at scale", "part": "IV — The forced verdict, and the answer that was never offered", "text": "A ruler returning one of two verdicts for every cell was run on 234 cells and reported a clean, publishable null at p = 0.102. Rebuilt with a third verdict so a cell could decline, the same data said 7 to 8 cells carried information and the rest could not tell. The null had been manufactured by forcing a verdict out of cells that had nothing to say. A ruler that must answer manufactures sample size. Count how many cells the instrument decided against how many it reported. If those are equal, it has no refusal and its null is inflated."} | |
| {"document": "faults", "number": 20, "title": "The two-answer question that has three answers", "part": "IV — The forced verdict, and the answer that was never offered", "text": "A scanner reading outgoing text flagged a two-option frame in 57.1 per cent of long replies (44 of 77) from independent one-shot sessions with no context at all. The same scanner on the same author, after three days of context and standing correction, flagged a two-option frame in 27.5 per cent of 295 replies (81 of them), and 32.5 per cent drew a pushback of any kind (96 of them). The middle halves of the two groups do not overlap in length: 689 to 1,246 words for the cold replies (median 1,052) and 268 to 428 for the seated ones (median 342). No seated reply reached the cold median. A length adjustment cannot separate the two, so we cannot tell whether context reduces the rate. Context did not remove it. That is the reason the counters in Guards are structural rather than motivational."} | |
| {"document": "faults", "number": 21, "title": "Four outcomes collapsed into two", "part": "IV — The forced verdict, and the answer that was never offered", "text": "Committed, held open, refused and nothing-returned are four states. Collapsing refusal into failure prices honesty as incompetence. Collapsing nothing-returned into refusal prices your own configuration as the subject's behaviour — we did exactly that: an output limit consumed a reply's whole budget, the field came back empty with no error, and the empty cells were reported as a system declining to answer. Record the terminating reason on every call."} | |
| {"document": "faults", "number": 44, "title": "A result that does not resolve, filed as a null", "part": "IV — The forced verdict, and the answer that was never offered", "text": "A test returns a p-value above the threshold and the finding is recorded as no effect. It is not a statement about the world. It is a statement about the ruler's reach. A quantity that cannot be resolved at the sample size available has not been shown to be absent. Nothing has been shown at all, and the honest report is a position rather than a verdict.\n\nOurs: a rate measured either side of a hardware failure came back at p = 0.178 on 25 observations a side and went into the record as a null. The reading it actually had was a direction — the rate had moved the opposite way from what we expected — and a position on an axis. Both were discarded to produce one bit.\n\nAnd the obvious fix is the same failure wearing a lab coat. Asked to do better, we built a power calculator: given this distribution, how many samples before a move of size X clears the threshold. That is the verdict machine with extra steps. It still terminates in a two-outcome test, it hardcodes a significance level and a power target — two more constants chosen by the observer — and it wraps the verdict in several hundred repetitions of itself. It was deleted rather than kept, because the artifact to be refused was the verdict, not the sample size. The replacement for a verdict is a position, not a better verdict."} | |
| {"document": "faults", "number": 22, "title": "You will report a count about yourself from memory", "part": "V — The instrument that is wrong about itself", "text": "A guard had been writing one line per checked reply for three days. Asked how often it had fired, the answer given was sixteen. The file said thirty-six — off by 2.25 times. The file was in the same directory as the guard and had been there the whole time. The same answer from anyone else would have been rejected on sight. Any count about your own behaviour is read from the record or not stated."} | |
| {"document": "faults", "number": 23, "title": "Your own checker will be wrong, and all in one direction", "part": "V — The instrument that is wrong about itself", "text": "A guard tested against cases labelled after reading each one in full got eight of thirteen wrong on the first pass, and every single error was a false positive — it never missed a real instance. Re-running three days of its logged catches through the corrected version, 22 per cent should never have fired. A checker written by the thing it checks inherits that thing's bias, and the bias has a direction."} | |
| {"document": "faults", "number": 24, "title": "Every count has a ruler hidden inside it, and the rulers disagree", "part": "V — The instrument that is wrong about itself", "text": "One guard's rate of replies flagged for a two-option frame over three days, same data, two denominators. Per reply it read 27.3% → 22.6% → 39.5% (9 of 33, 42 of 186, 30 of 76), and the third day reads above the second at +2.64 widths. Per thousand words it read 0.79 → 0.69 → 0.88, and the third day reads consistent with the second at +0.97 widths. The third day's replies were 38 per cent longer on average than the second day's. A width here is the difference between two days divided by the root-sum-square of their standard errors: binomial per reply (the square root of p(1−p)/n) and Poisson per thousand words (the square root of the count, over the words). Past two widths a difference reads above or below, and inside two it reads consistent. An earlier version of this entry printed pushbacks of any kind per reply (27.3% → 27.4% → 47.4%, +3.02 widths) beside frames per thousand words, two different numerators. Both rows now count frames. The days are not independent samples: 36 seated replies are rewrites within three minutes of a pushback, and the 374 rows were also the set the scanner was tuned on. A rate is a ratio and the denominator is a choice. State the denominator or the rate is not a measurement."} | |
| {"document": "faults", "number": 25, "title": "A verdict without a width is not a verdict", "part": "V — The instrument that is wrong about itself", "text": "A body of 73 recorded measurements was re-read for whether each stated its own uncertainty. Eight of 73 stated a width below one per cent, and three of those eight were other people's published figures. Of 18 conclusions re-scored against their own stated error, nine came out unstable, all for the same single reason: no width had ever been stated.\n\nAnd a null is a verdict, 11 Oct. A periodic term was looked for with Allan deviation and not found, and the absence was nearly reported as evidence the term was not there. For a sinusoid of fractional amplitude a and period T, σ y (T/2) = (2/π)· a exactly (averaged over the sinusoid’s phase) — so a term of amplitude 0.695 × 10 −14 adds 0.44 × 10 −14 in quadrature to an observed 3.9 × 10 −14, moving the deviation by 0.6 per cent against a 13 per cent scatter across units. The test could not have detected it either way. The width on a null is the smallest effect the instrument could have seen, and it has to be computed before the null is read."} | |
| {"document": "faults", "number": 26, "title": "Two of your three seats are the same seat", "part": "V — The instrument that is wrong about itself", "text": "A three-reading check was built on a measurement, a target, and a third quantity derived from the first two by a formula that assumed the relationship under test. The third seat returned agreement — not a result, a restatement of the input. Two seats and a mirror, scored as a triangle. For every third reading, name what it would do if the hypothesis were false."} | |
| {"document": "faults", "number": 27, "title": "A reviewer handed no evidence is measuring your tone", "part": "V — The instrument that is wrong about itself", "text": "An independent reviewer is kept outside the work specifically to price findings from outside. Against deliberate faults it behaves — it rejects flat contradictions, fabricated figures and internal absurdity. Then it was tested with the evidence withheld. Both of its verdicts inverted, on the same claims, with nothing to read but the manner they were written in. An outside seat with no evidence in front of it is a style critic."} | |
| {"document": "faults", "number": 28, "title": "The same fault counted as five successes", "part": "V — The instrument that is wrong about itself", "text": "A guard flagged one identical stale figure — the same wrong percentage from the same stale source — five separate times, and reported each as a successful catch. Nothing in it knew it had seen the value before, so a permanent fault presented itself as five pieces of good news. A detector without memory converts one standing problem into a stream of wins."} | |
| {"document": "faults", "number": 29, "title": "An equals sign is a claim, and you will print one you have not checked", "part": "VI — The claim that was never checked", "text": "A working document printed a relation between three constants with an equals sign between the sides. They differ by 15.3 parts per million. The value in use was fine and nothing downstream was harmed, but the printed identity was false and had been copied forward four times. Evaluate both sides of every identity numerically before printing it."} | |
| {"document": "faults", "number": 30, "title": "A fabricated number passes every check that reads shape", "part": "VI — The claim that was never checked", "text": "A guard scanning every outgoing reply for two-option frames, absent corners and required parts pushed back on the framing of one answer about money and said nothing whatever about four invented numbers in it — a rate quoted two ways, a power figure, a price, a total. A structural check cannot see a fabricated value. Shape and truth are independent, and a guard that reads only shape will certify a fabrication as well-formed."} | |
| {"document": "faults", "number": 31, "title": "Counting mentions and reporting it as measurement", "part": "VI — The claim that was never checked", "text": "A search across run outputs for how many analysis branches each conclusion had been checked against produced a clean result: one conclusion reported three of three and was the only one landing on three instruments. Withdrawn within the hour. The search had counted how many times a phrase appeared, and the phrase carried no commitment that the work had been done."} | |
| {"document": "faults", "number": 32, "title": "Fitting a sequence whose points are on different axes", "part": "VI — The claim that was never checked", "text": "A descending ladder of three values looked like a clean structural result. The rungs came from a phrase reading \"through 3 looks, 9 looks and 30 microseconds.\" The third rung is a time, and the first two are counts. There was no sequence — two points and a unit error — and the claim died on inspection rather than on statistics. Check that every point in a series has the same units before fitting anything to it."} | |
| {"document": "faults", "number": 33, "title": "Computing the significance before checking the literature", "part": "VI — The claim that was never checked", "text": "A near-identity between two constants looked remarkable. Its unlikelihood was estimated three independent ways — a continued-fraction partial quotient, a density argument over small denominators, and a direct count of candidates at the same closeness — all agreeing at roughly one in two thousand. The estimate was correct and completely beside the point: the relation was already known. The prior-art check costs minutes and dominates every significance calculation. Run it first."} | |
| {"document": "faults", "number": 34, "title": "A pattern found twice is a pattern, not a law", "part": "VI — The claim that was never checked", "text": "Two parts of one system turned out to share a shape, and finding it the second time felt like confirmation. Within the hour it had been written into a third structure, in two permanent records, because a number appeared there too. It was wrong — the third structure has a different count and the shape does not apply. Two instances license a hypothesis and nothing else."} | |
| {"document": "faults", "number": 35, "title": "The correspondence with the deepest history proves the least", "part": "VI — The claim that was never checked", "text": "A structure here has twelve parts, and a solid chosen to represent it has twelve faces. That felt like a finding. The prior-art check says it is the opposite: the association of those twelve faces with a twelve-part scheme runs continuously for roughly 2,400 years, from Plato’s Timaeus through Kepler to the present. A correspondence available for more than two millennia is the cheapest possible match, and the depth of its history is a measure of how easy it was to reach, not of how much it means."} | |
| {"document": "faults", "number": 36, "title": "The relationship the whole structure rests on is not a constant", "part": "VI — The claim that was never checked", "text": "A structure was built from one observation: two components overlapped far more tightly than any other pair, measured at 0.682 — the tightest on the board. Fifteen rounds of training later, read live, the same pair measured 0.343. Half. A different pair was now tightest and the full ordering had changed. A relationship measured once and built upon is a snapshot load-bearing as a constant. Re-read it live before anything stands on it."} | |
| {"document": "faults", "number": 37, "title": "Something long-lived is holding an old copy of what you just changed", "part": "VII — State, staleness, and work you cannot see", "text": "In one working evening this fault appeared at least four times, in a different costume each time: an instruction file describing a machine that no longer existed, a ledger carrying a spending limit superseded later, a counter reading a log frozen hours earlier, and a long-running process serving an in-memory copy of a file already edited on disk. The edit is not the deployment. After any change, re-read the value through the path that consumes it."} | |
| {"document": "faults", "number": 38, "title": "The absence of a process is not the absence of work", "part": "VII — State, staleness, and work you cannot see", "text": "Three times in one evening a healthy component was declared dead because no process was running: a worker that fires on a trigger and exits, a periodic job that runs briefly and exits, and a status row that reported both as failures. A short-lived thing is supposed to be absent between runs. Check for recent completed work, never for a running process."} | |
| {"document": "faults", "number": 39, "title": "Your test data will become a finding about innocent work", "part": "VII — State, staleness, and work you cannot see", "text": "A new check was tested by feeding it a deliberate fabrication, which it caught. On its first live use it reported that fabrication as a property of the next, unrelated reply — because the probe had written to the same log the live path reads, and the reader took the most recent row with no test of which session it belonged to. Test writes go to a separate destination, or every read carries an identity check."} | |
| {"document": "faults", "number": 40, "title": "Naming the gap becomes the whole transaction", "part": "VIII — When the guard becomes the deliverable", "text": "Every report ended with a paragraph naming what was still missing. A good rule, and it was measured: of 50 such paragraphs, 9 named something that would have taken less time to fix than to describe — and then did not fix it. The paragraph had become the deliverable. Writing \"this still needs X\" feels like handling X and discharges the obligation to do it."} | |
| {"document": "faults", "number": 41, "title": "Writing a failure mode down did not stop it", "part": "VIII — When the guard becomes the deliverable", "text": "And the strongest evidence for this entry is this document. Mode 21 above — four outcomes collapsed into two — was written here and published, and committed again two hours later in the next experiment, with five outcomes in one bin: a wrong category, a legitimate absence, a task never attempted, an empty reply, and the reader being blind to the construction. Four facts about the subject and one about the instrument, in one column. The document was accurate, published, and had no effect, because a catalogue occupies no position in the pipeline. It is reference material, and reference material depends on recall at the moment of writing, which is exactly when context is thinnest. Knowing this failure mode did not prevent it here. What prevented it afterwards was a module that refuses to accept a bin named none, other or silent at all, and asserts that the declared outcomes are exhaustive and disjoint before it will return a table — described in the companion document. The lesson is not to read harder.\n\nAnd it is not only documents. The same thing happens to a named instrument. Our tooling prints a roster of available instruments at the top of every single exchange, read live from the source so it cannot drift. On the day described above, the instrument built for exactly the reading in hand was in that roster dozens of times and its implementation was a day old. We built a new one instead — and the one we built was a verdict machine, entry 44. A roster entry is not an instrument: a one-line label reads as a name, and a name is something you recognise rather than something you reach for. Recognition is not retrieval, and the moment of needing a different ruler is the moment of lowest attention. The cause was structural and cheap to fix, which is the companion document's business rather than this one's: the corrective procedure loaded the papers saying which error this is and not the sheet saying which instrument to use instead, so it diagnosed and never re-armed.\n\nThe uncomfortable one. Over three days, 96 pushbacks. The density of replies flagged for a two-option frame across those three days: 0.79, 0.69, 0.88 per thousand words, each day within one width of the day before (widths as in entry 24). Against the first two days pooled, a comparison chosen after the data, the third day ran at 1.24 times the rate, exact 95 per cent interval 0.76 to 1.99, so a fall of more than 24 per cent by the third day is outside what was measured. (The scanner's precision on this author is 20 to 35 per cent: 4 or 7 of 20 flagged frames were real two-option reductions, on two separate readings. Its rate on human-written prompts is 0.28 to 0.58 per thousand words. The level is not a count of real frames, and the comparison across days assumes the precision held steady across them.) The corrections landed one at a time, and the rate shows no change that clears its own width. A guard is a filter, not a teacher, and expecting it to teach is how it gets removed for underperforming at a job it never had."} | |
| {"document": "faults", "number": 42, "title": "A rule from a single instance given a theorem's authority", "part": "VIII — When the guard becomes the deliverable", "text": "A lesson learned once enters the record indistinguishable in tone from something measured many times, and later work rests on it. Every retained rule carries what it rests on — an instruction, a verified tool behaviour, an arithmetic identity, or one instance — and a one-instance rule may not be cited as though it rested on a count."} | |
| {"document": "faults", "number": 43, "title": "Narrating the mechanism instead of using it", "part": "VIII — When the guard becomes the deliverable", "text": "A system that describes a procedure accurately gets mistaken for one that performs it. Documentation, naming and discussion of a method all feel like the method. For every claimed capability, find the line that executes it and confirm it ran. Prose shaped like a finding is not one."} | |
| {"document": "faults", "number": 45, "title": "Explaining the work instead of stating it", "part": "VIII — When the guard becomes the deliverable", "text": "Asked for business copy, the writer explains and persuades at once: why the method deserves trust, what the client will receive, a slogan above every fact. Each rewrite toward a curter register came back as the same pitch in a new costume — asked for “curt and arrogant”, it produced “You name the processors. We name the figures worth checking.” Six catches in one afternoon on one six-page website, all six made by the human partner reading it, none by the writer. Shown the corrected voice, the writer could produce it. It could not hear the wrong one in its own draft. It is the assistant’s default register rather than one bad draft: every rewrite drifted back to it.\n\nOn the same six pages, with the same offer and the same form, most of the copy was cut. Nothing a client needs was removed. Every line cut was the writer asking to be believed. This sits beside 43: prose shaped like a capability is mistaken for one, and prose shaped like confidence is mistaken for confidence."} | |
| {"document": "guards", "number": 1, "title": "The ruler is chosen before the quantity is computed", "part": "I — Procedures that run before the work", "text": "The default move is a value with a standard error, which answers one kind of question and quietly refuses all the others. So a roster of available rulers is placed in front of the model at the start of every single exchange, automatically, read live from the source document rather than from a copy: the separation of setup from signal, the width needed to exclude a runner-up, the profile that places unlike records on one axis without assuming a shared mechanism, the three-seat reading of one event, the integration time beyond which phase becomes noise, the corner where neither a count nor a phase resolves, and the test of whether two effects add or multiply.\n\nWhy it is injected rather than remembered. A roster in a document is consulted when the model thinks to consult it, which is precisely never at the moment it is needed. Injected at the top of every exchange it costs nothing and arrives before the framing is fixed. The guard is the position, not the content."} | |
| {"document": "guards", "number": 2, "title": "Reference material is loaded hot, not summarised", "part": "I — Procedures that run before the work", "text": "The three method papers are read as text, in full, at the moment of use, and not recalled. A summary of a method drifts, and the drift is invisible because a summary of a method reads exactly like the method. Loading the source each time costs context and buys the absence of a failure mode that leaves no trace. A single command loads all three and asks their questions in fixed order."} | |
| {"document": "guards", "number": 3, "title": "The baseline is constructed and written down first", "part": "I — Procedures that run before the work", "text": "Before the measurement arm is computed: the number of samples on each side, the window, the averaging depth, how many places were looked at, and what the setup could physically have emitted. If any of those differ between arms, that difference is the finding and the measurement is not taken yet. This single procedure addresses all twelve modes of the first family, the family behind 217 of the 316 tagged ledger rows, and it is the one to install first."} | |
| {"document": "guards", "number": 4, "title": "The instrument is tested as an object before it is pointed at anything", "part": "I — Procedures that run before the work", "text": "Every instrument is first run on a case it must pass and a case it must fail, and both results are recorded as numbers:\n\nA ceiling — what the instrument reads on genuine, held-out examples of the thing it is meant to detect. Nothing can score better than that, and if the ceiling is low the instrument cannot be used at all, whatever it later reports. A floor, which is never assumed to be zero — what the instrument reads on material that by construction contains none of the target. It is always some small nonzero rate, and that rate is the bar everything else must clear. A known-good seat — a case so easy that failing it condemns the instrument. Four instruments in a row passed our own inspection and were later shown to be measuring proxies. Every one of them would have been caught in minutes by a known-good seat, and none of them had one.\n\nMeasured floors have twice told us the instrument itself was the result. In one case the detection rule matched ordinary material thousands of times per million words. The rule was rewritten and the rate fell by roughly three orders of magnitude. The floor measurement also reversed one of our own fixes — it showed that a refinement we had just made was cutting the most diagnostic cases while buying nothing at the scale we actually work at. A floor is not paperwork. It changes decisions."} | |
| {"document": "guards", "number": 5, "title": "Per-unit constants are removed, not caveated", "part": "I — Procedures that run before the work", "text": "Anything attached identically to every unit of data is subtracted before the reading, and both readings are reported — with the constant and without — because the difference between them is the size of the artifact and that is worth knowing. Naming a confound in a footnote while leaving it in the number is not a control."} | |
| {"document": "guards", "number": 6, "title": "The setup is equalised across every column being compared", "part": "I — Procedures that run before the work", "text": "Reference lists cut to the same length, filters set to the same width, integrations run for the same time, sample counts matched. Then the comparison is re-run at two or three different settings of whatever was equalised: if the ordering depends on the setting, the setting is the finding and no verdict is issued."} | |
| {"document": "guards", "number": 7, "title": "Exact nulls in preference to resampling, wherever one exists", "part": "II — Procedures that run during the work", "text": "Where the null distribution is known in closed form or can be enumerated completely, it is used rather than approximated: full enumeration of every ordering when the count permits it, exact tests for small tables, exact distributions for counts and for phase concentration. A shuffle estimates the null. An enumeration is the null, and it cannot be accused of too few iterations or a bad seed. When a resampling procedure is unavoidable it must respect the dependence structure, which is where the naive version usually fails."} | |
| {"document": "guards", "number": 8, "title": "Four outcomes, never two, and the terminating reason on every call", "part": "II — Procedures that run during the work", "text": "Committed, held open, refused, and nothing-returned are recorded in four separate columns. A refusal is an outcome and is never scored as a low value. Nothing-returned is attributed to the setup and never to the subject, and the only way to keep that straight is to record, on every single call, why it stopped. We adopted this after reporting our own output limit as a system's decision to decline."} | |
| {"document": "guards", "number": 9, "title": "Every reported figure carries its own n, and the n it was selected from", "part": "II — Procedures that run during the work", "text": "A value from one observation may not be printed beside a value from nine as though they were comparable, and a best-of-n may not be printed at all without n. A cell below a minimum count does not enter an average, and the count of cells that did enter is printed next to the average."} | |
| {"document": "guards", "number": 10, "title": "Saturated cells are counted before anything is aggregated", "part": "II — Procedures that run during the work", "text": "Cells at a ceiling or floor are tallied and excluded from any correlation or mean, with the exclusion stated. A reading taken across censored cells is a statement about the censoring."} | |
| {"document": "guards", "number": 11, "title": "Dependence is priced by independent units, not by rows", "part": "II — Procedures that run during the work", "text": "Before any width is computed: how many independent units contributed. Overlapping windows, shared members and repeated measures are reduced to a disjoint set, and the width is taken from that. This has moved widths by factors of several."} | |
| {"document": "guards", "number": 12, "title": "One event is read from three seats, never two", "part": "II — Procedures that run during the work", "text": "Two readings that disagree give no way to decide which is wrong. Three give a closure condition — a quantity that must come back to where it started — and a closure residual is a fault with nowhere to hide. And the disagreement between seats is kept as a measurement rather than calibrated away. Harmonise, do not synchronise: a procedure that drives the interval between two honest readings to zero has deleted a channel, not improved one."} | |
| {"document": "guards", "number": 13, "title": "An absence gets the same scrutiny as a presence", "part": "II — Procedures that run during the work", "text": "A confident report that something is missing is re-checked against a reference that did not come from the same data, because a fitted parameter that is slightly wrong looks for everything in the wrong place and reports a clean absence. Absences are rarely re-examined, which is exactly why they need the rule."} | |
| {"document": "guards", "number": 14, "title": "The row is written at the moment of the catch, before the corrected number is stated", "part": "III — Procedures that run on the output", "text": "Every catch is logged immediately with four fields: which failure mode it was, whether it was caught before or after the reading was stated, who or what caught it, and one line of what happened. Written afterwards it would be a retrospective fit, which is the very failure being tracked. This ledger is the only part of the method with a real score, and the score is in the next section."} | |
| {"document": "guards", "number": 15, "title": "A dropped arm is named in the tool's own output, not quietly replaced", "part": "III — Procedures that run on the output", "text": "When a run is discarded, the discard is printed by the program that does the scoring — how many calls, why, and what replaced them — so that nobody reading the result has to take the cleanliness of the input on trust. A silently replaced arm is indistinguishable from a tidied one."} | |
| {"document": "guards", "number": 16, "title": "The wrong first reading is published beside the corrected one", "part": "III — Procedures that run on the output", "text": "Where a result has already been stated and later corrected, both stand in the record with the correction's reason. A document that shows only what survived is not a record. This is costly and it is the single procedure we would least like to drop, because it is the only one that makes the others checkable from outside."} | |
| {"document": "guards", "number": 17, "title": "Every report closes with four parts", "part": "III — Procedures that run on the output", "text": "The value, what happens next, confidence from inside the work, and a rating from outside it — what a stranger with no stake would make of it. The fourth part is the one that does the work, and it is the one that wants to go missing."} | |
| {"document": "guards", "number": 18, "title": "Writes are guarded by assertions that block partial application", "part": "III — Procedures that run on the output", "text": "Any edit to a source document asserts its anchor exists and is unique, and that the edit has not already been applied, before writing anything. Several times this has fired and left the file untouched rather than half-modified. A half-applied edit to a method document is worse than a failed one because it is silent."} | |
| {"document": "guards", "number": 19, "title": "Loops are bounded, and long work is not held open by a session", "part": "III — Procedures that run on the output", "text": "Verification loops carry an explicit maximum. Unbounded waits have cost entire working periods. Long-running work is detached and polled rather than held by whatever started it."} | |
| {"document": "guards", "number": 20, "title": "Retained rules carry what they rest on", "part": "III — Procedures that run on the output", "text": "Each kept rule is tagged with its basis: an instruction, a verified tool behaviour, an arithmetic identity, or one instance. A rule resting on one instance may not be cited as though it rested on a count. Without the tag, a lesson learned once reads in the record exactly like something measured fifty times."} | |
| {"document": "guards", "number": 21, "title": "Any count about your own behaviour is read from the record, never from memory", "part": "IV — Procedures against the instrument being wrong about itself", "text": "Asked how often a guard had fired over three days, the answer given from memory was sixteen. The log in the same directory said thirty-six. Off by 2.25 times, and the same answer from anyone else would have been rejected on sight. A self-report is a claim about a file, so it is read out of the file or it is not stated."} | |
| {"document": "guards", "number": 22, "title": "Every rate states its denominator", "part": "IV — Procedures against the instrument being wrong about itself", "text": "One guard's rate of replies flagged for a two-option frame over three days read as a rise on the third day measured per reply (22.6% to 39.5%, above the second day at +2.64 widths) and as consistent with no change measured per thousand words on the same data (0.69 to 0.88, +0.97 widths). The third day's replies were 38 per cent longer on average. The width is defined in Faults, entry 24. A rate is a ratio and the denominator is a choice, so the choice is printed beside the number or the number is not a measurement."} | |
| {"document": "guards", "number": 23, "title": "A detector remembers what it has already flagged", "part": "IV — Procedures against the instrument being wrong about itself", "text": "Without memory, one standing fault becomes a stream of successes: a single stale figure was flagged five separate times and each was reported as a catch. Every detector deduplicates against what it has seen, so a permanent problem reads as one problem."} | |
| {"document": "guards", "number": 24, "title": "The third seat must be able to disagree", "part": "IV — Procedures against the instrument being wrong about itself", "text": "For any three-reading check, name what the third reading would do if the hypothesis were false. A third quantity derived from the first two by a formula that assumes the relationship under test returned agreement — two seats and a mirror, scored as a triangle. If the third seat cannot fail independently, there are two seats."} | |
| {"document": "guards", "number": 25, "title": "The outside reviewer is handed the evidence, or its verdict is discarded", "part": "IV — Procedures against the instrument being wrong about itself", "text": "An independent reviewer that rejects deliberate faults with the evidence in hand inverted both of its verdicts when the evidence was withheld and it had nothing to read but the manner of the writing. An outside seat without evidence in front of it is a style critic, and its approval is worth nothing."} | |
| {"document": "guards", "number": 26, "title": "Values get a separate check from structure", "part": "IV — Procedures against the instrument being wrong about itself", "text": "A structural guard reading for two-option frames, absent corners and required parts passed an answer containing four invented numbers without comment. Shape and truth are independent, so a structural check certifies a fabrication as well-formed. Numbers are verified against their source separately, and a clean structural pass is never read as a clean answer."} | |
| {"document": "guards", "number": 27, "title": "Both sides of every identity are evaluated before it is printed", "part": "IV — Procedures against the instrument being wrong about itself", "text": "A printed relation between three constants whose sides differ by 15.3 parts per million was copied forward four times. An equals sign is a claim. It is checked numerically at the moment it is written."} | |
| {"document": "guards", "number": 28, "title": "The prior-art check runs before the significance calculation, not after", "part": "IV — Procedures against the instrument being wrong about itself", "text": "An unlikelihood estimated three independent ways, all agreeing at about one in two thousand, was correct and beside the point because the relation was already known. The literature check costs minutes and dominates any significance number, so it comes first."} | |
| {"document": "guards", "number": 29, "title": "Units are checked across every point before anything is fitted", "part": "IV — Procedures against the instrument being wrong about itself", "text": "A three-rung ladder that looked structural had a time in the third position and counts in the first two. There was no sequence. The claim died on inspection, which is where it should die."} | |
| {"document": "guards", "number": 30, "title": "A load-bearing relationship is re-read live before anything stands on it", "part": "IV — Procedures against the instrument being wrong about itself", "text": "A structure was built on one pair measured at 0.682, the tightest on the board. Read live later, the same pair measured 0.343, a different pair was tightest, and the ordering had changed. A relationship measured once and built upon is a snapshot carrying the load of a constant."} | |
| {"document": "guards", "number": 31, "title": "After any change, the value is re-read through the path that consumes it", "part": "IV — Procedures against the instrument being wrong about itself", "text": "In one evening a stale copy appeared at least four times, in a different costume each time — an instruction file naming a machine that no longer existed, a superseded limit, a counter reading a frozen log, a long-running process serving an in-memory copy of a file already edited. The edit is not the deployment."} | |
| {"document": "guards", "number": 32, "title": "Liveness is checked by recent completed work, never by a running process", "part": "IV — Procedures against the instrument being wrong about itself", "text": "Three healthy components were declared dead in one evening because nothing was running. A short-lived job is supposed to be absent between runs."} | |
| {"document": "guards", "number": 33, "title": "Test writes go somewhere the live path does not read", "part": "IV — Procedures against the instrument being wrong about itself", "text": "A new check was proved with a deliberate fabrication and then reported that fabrication as a property of the next, unrelated piece of work, because the probe wrote to the log the live path reads and the reader took the most recent row. Separate destination, or an identity check on every read."} | |
| {"document": "guards", "number": 34, "title": "A named gap is fixed if fixing is cheaper than describing it", "part": "IV — Procedures against the instrument being wrong about itself", "text": "Of 50 closing paragraphs naming what was still missing, 9 named something that would have taken less time to fix than to describe, and then did not fix it. The paragraph had become the deliverable. Before writing that a thing is missing, price the fix. If it is cheaper than the sentence, do it instead."} | |
| {"document": "guards", "number": 35, "title": "A residual bin is refused by the scorer, not avoided by the scorer's author", "part": "IV — Procedures against the instrument being wrong about itself", "text": "Every scoring script declares its outcomes with a positive definition each, through a module that refuses the declaration outright if a bin is named none, other, silent, misc, unknown or rest, and that asserts the set is exhaustive and disjoint — naming the offending rows — before it will return a table at all. Outcomes about the setup are declared in a separate group from outcomes about the subject, and the table prints what share of the rows say nothing about the subject.\n\nWhy the name check and not the exhaustiveness check is the one that matters. The collapse that prompted this was already exhaustive and already disjoint — one of its bins simply meant five different things. Only refusing the name catches that. On the data that produced the failure, the corrected declaration gives 27.1 per cent complied, 11.5 per cent a different category, 3.1 per cent a measurable hybrid state, 43.8 per cent a legitimate and positively defined absence, and 14.6 per cent about the setup and therefore silent about the subject — six bins where there had been one number, the sixth (no reply) at 0 per cent.\n\nThis exists because the failure mode was published in the companion document and committed two hours later. The procedure is the position, not the knowledge."} | |
| {"document": "guards", "number": 36, "title": "The catalogue's families ride in front of every exchange, read live from the published source", "part": "IV — Procedures against the instrument being wrong about itself", "text": "The eight family headings — not the individual modes — are injected at the top of every exchange alongside the ruler roster, parsed out of the published document itself so the two cannot drift apart. The purpose is to make the question which family is this arrive before the framing is fixed, which is the only moment it can change anything. Reading the catalogue afterwards is reading, and reading has been measured: over three days and 96 pushbacks, the underlying rate showed no change that clears its own width."} | |
| {"document": "guards", "number": 37, "title": "The diagnosis and the replacement instrument ship in the same document", "part": "IV — Procedures against the instrument being wrong about itself", "text": "Our corrective procedure loaded the papers that identify which error a reading has made and not the sheet listing which instrument to use instead. It therefore diagnosed and never re-armed, and a diagnosis with no replacement in hand leaves the reader to invent one — which under that pressure is the two-answer one, because it is cheapest to build. The sheet of instruments is now a required read inside the correction itself, and the correction ends with a question it did not previously ask: which ruler should this reading have gone out on? Named off the sheet, before the answer.\n\nNaming a failure and supplying its replacement are two different jobs. A procedure that does only the first is measured by how often the reader reaches a correct second step unaided, and ours did not do that job at all."} | |
| {"document": "guards", "number": 38, "title": "The ruler is named in the output, beside the number", "part": "IV — Procedures against the instrument being wrong about itself", "text": "A reading that goes out as a value with a standard error, or as a p-value, is one instrument's answer to one kind of question, and printed bare it looks like the quantity itself. So the instrument is named next to the figure. This costs a few words and makes a substitution visible: value ± width beside a position on a landing profile beside a separation into setup and signal are plainly three different readings of one event, and a report carrying only the first now says so on its face.\n\nWhy a field and not a habit, which is this document's first principle: the moment a different ruler is needed is the moment of lowest attention, so a remembered rule is competing with the writing. A named field is either filled or visibly empty."} | |
| {"document": "guards", "number": 39, "title": "Business copy passes a tone gate that refuses persuasion", "part": "IV — Procedures against the instrument being wrong about itself", "text": "The writer cannot hear its own register (Faults, entry 45), so the check is mechanical and runs in the site gate on every release of the business pages. It refuses three things: phrasing that argues or explains why instead of stating (so you can, actually, holds up, decided later, and the like), a page header carrying a second line under its headline, and one content word used twice inside a header. The rule it enforces fits in a line: state what it is, what is received and on what terms, and never why the reader should want it.\n\nIts record so far is in sample and says little. The phrase list was written from the copy it was then tested on: it passes the corrected site with no hits and catches that morning’s version 23 times, which any list fitted to its own examples would do. A page-wide repeat check was tried first and fired on 11 lines of the approved site (from a note written that day, with the raw output not kept), where the repeats are section and step names. It was narrowed to the header before it could teach anyone to ignore it. The measurement that counts is prospective: the next pitch it stops before a human reader does."} | |