habitatTest log¶
Every declared test and what it returned, kills included. Newest at the bottom.
A test log, not a work journal. Each entry is a test that was declared before it ran, what it returned, and whether it survived. What it deliberately does not carry: file paths, branch names, function names, internal documents, who owes whom a decision, or what we plan to do next. Those are the shop floor and they are not a result.
This page was rewritten on 5 September 2026 to remove internal references that should never have been published. No result changed, and nothing that failed was removed. The kills below are the same kills.
2026-09-02¶
A correction, first. A commit was pushed that morning whose message said the tests were green. No test had run. The claim was retracted on the record the same day and the rule written down: a result never observed is never stated. It is first here because it is the worst thing that happened that day.
One letter against a floor the record sets itself. Adams to Jefferson, 3 February 1812, fifteen of the record's 1,103 compositions, against 164 letters. Removing it moves the field by 0.903 percent of its own norm. Three hundred random removals of the same size give a median of 0.001992, so the letter is 1.41 times that median, at the 94th percentile, sixth of 164 and first among the thirteen letters of similar size. Reordering the record leaves the fingerprint unchanged.
And the blind ranking against it: ALL PREDICTIONS MISSED. Six predictions were frozen before the run, three from the instrument's operator and three from its author. Every one missed on the blind object. One apparent hit was a singleton artifact and was not counted. An internal-consistency check passed at rho 0.208, and had been flagged in advance as close to tautological, so it carries nothing. Verdict: the single-letter result is suggestive only, and no claim was published on it. Recorded either way, which is the rule.
A stored value was not what the documentation said. The frozen record's covariance was found by recomputation to be the raw covariance of its own vectors, agreeing to 3.6e-16, and not the ridged form the developer contract described. The contract line was stale and was superseded.
A reading was withdrawn and re-run. The per-composition differentials in the joint record were the per-author values, not the joint ones, verified by recomputation. The reading built on them was withdrawn. Re-run correctly, the joint record is flat at entry after its founding period.
Two clocks, over 102 orders of entry. The apparent jumps at the founding of the record are position effects and appear in every order. Exactly one letter holds its place independently of when it arrived: 12 September 1821, Jefferson, in 98 of 100 random orders and under both true and reversed order. Beyond that one, named letters are at chance.
A prediction FAILED and an effect was KILLED. The prediction was that after founding the record would play its least-practiced direction more than chance. It plays it less. The interference effect that was proposed alongside it showed no departure at all and was killed. Published as they fell.
Ridge-independence FAILED. Across modes, orders and ridge values, the declared pair came out at 0.77 rather than independent. A second prediction was mis-specified by its author, an impulse series fed to a rule meant for trajectories, and also failed. Letter-level effects sat at chance. One order-independent outlier survived at every ridge value, 21 January 1812.
The first self-read: one half KILLED, one half held. The absorption reading's first form died. Its distribution was indistinguishable from a Gaussian null at KS p 0.85, and the premise underneath it was simply wrong: the null was measured at 0.185 where the author had asserted 1. That was logged as a guess stated before verification. The second half held, with mode tracking at overlap 1.000 after founding.
Re-specified, the surviving reading is narrower than hoped. About 18 percent of acts sit outside the noise band, wider than noise rather than shifted, and the median act is indistinguishable from noise. The same letter appears at both extremes, which means these are states an act passes through and not properties an act has. A stated limit: the variance is not matched per entry step. The second reading holds exactly: six harmonics carry half the field's energy and twelve carry ninety percent, constant across 1,054 steps after founding.
2026-09-04¶
Which documents carry a record. All 164 letters, each removed in turn against a floor of 300 same-size random removals the record generates itself. Eight predictions were frozen before the run. Six held, two split.
The raw distance is largely a measure of length, correlating with act count at 0.85. The percentile against each letter's own floor is not, at 0.17, which is the reason the floor exists. Thirteen letters sit at or above the 95th percentile of their own floor.
The declared kill condition passed: the geometry fingerprints match those published from a separately written implementation, byte for byte, both for the whole record and for the record with one letter removed.
A tolerance was wrong, not the code. One prediction asked two figures to agree within five points. They differed by six. Both are 300-draw estimates of the same quantity, so their difference carries a standard error of four points by construction, and a five-point band fails correct code about one run in five. The band was widened with that arithmetic stated, before the run rather than after.
And a guess about the record that the record refused. The letter furthest outside its own floor was predicted to be a short one, different from the largest mover. It is the same letter.
Portability, measured rather than asserted. Two ways of summing the same covariance differ by at most 6.4e-16 across all 164 letters, while the smallest gap any decision sits from its boundary is 4.1e-9. Every rank and every percentile therefore holds by a factor of about 6.5 million and cannot change with the machine.
2026-09-05¶
Which act changed an agent: no act is established. Run on two records.
On the first, the instrument located a candidate and could not certify it. The reading that tests position put it at 0.111 against a required 0.004.
On the second, a record made by someone else, the answer moved when a declared setting moved: the dated act jumped from 220 to 3,617 and the verdict flipped twice. That is an artifact of the setting, not a property of the record.
The floor, measured on a change planted deliberately. A doubling in the spread of the work is found and dated to within a dozen acts out of four hundred. Half again is missed. Smaller than a doubling is not seen at all. Pure noise is correctly silent.
Three corrections to the instrument, all found before the result was believed. A verdict resting on a category label rather than a magnitude scored a coin flip as a finding, and on pure noise it called three of six columns survivors on a field with nothing in it. A discharge detector could not see a discharge in this kind of series at all, because the series decays and the detector looks for a spike after the flattest stretch, which is the end. And a first attempt at dating read the decay rather than the record, scoring pure noise above a real planted change.
A control that was not a control. Public reference files were used as order-free records on the assumption that their row order carries no meaning. It does. One file dates a change at 0.010 as shipped and 0.960 once its rows are shuffled. The control is now the same record shuffled, which is order-free by construction, and under it nothing is dated.
A published corpus was found to be mostly empty. One of the records published here is 4,262 rows by six columns, and four of those columns cannot vary: every record in it is an observation the agent received rather than an act it produced, and the other three constants follow from that. Its covariance has rank two of six. Nothing in the build noticed, because every check asked whether the published parts agree with each other and none asked whether a published column carries information. A gate now refuses any column that cannot vary unless the reason it cannot is declared. No result was withdrawn: a fingerprint over a thin field is still a sound fingerprint. What a dead column costs is the geometry.