The bias published beside the point

“The widening rule” fitted the band’s shape and left its level alone, and its open: block named what remained: the log error has a growing positive mean, and an interval centered on a biased point inherits that bias whatever its width. #423 asked whether to correct it and on what. crates/project/tests/the_bias_that_belongs_to_the_years.rs measured the four remedies it named and every one of them failed, for reasons recorded there and summarized in that block — the bias is a year effect with no gradient on anything a forecast can see at its origin, so there is nothing to correct it on.

What that file also found is the reason this is a decision rather than a note. The 5.8% the earlier work quoted is the mean district’s. The feed publishes a statewide total from the same projections, and it is a different number:

horizon   mean district, pre-closure   total, pre-closure   mean district, all   total, all
1                 -0.0071                  -0.0052              -0.0071            -0.0052
3                 -0.0007                  -0.0050              -0.0007            -0.0050
4                 +0.0027                  -0.0018              +0.0070            +0.0029
5                 +0.0075                  +0.0006              +0.0171            +0.0107
7                 +0.0179                  +0.0061              +0.0294            +0.0172
9                 +0.0272                  +0.0171              +0.0434            +0.0278
10                     —                        —               +0.0562            +0.0322
13                     —                        —               +0.0734            +0.0569

The two lines cross zero in different years. Before the closure the mean district has turned positive by four years and is 0.75% high at five; the total is still 0.18% low at four and 0.06% high at five — inside a tenth of a point of unbiased at the horizon where the district figure is already drifting. A correction that serves one is wrong for the other, and a reader holding one of these numbers cannot tell which one they have.

Publish both, at every horizon, and correct neither. A bias array on the feed’s projection block, one row per horizon from one to thirteen, carrying four figures: the mean district’s and the total’s, each across the pandemic closure and restricted to targets before it.

project::backtest::bias_profile computes it, on the same scored forecasts profile scores the band against — one projection path, so a band measured against one set of forecasts and a bias measured against another cannot happen. The harness that produced the table above lived in a test file and is hoisted for the reason an-example-is-reachable-by-no-gate records: a number no artifact can reach is a number that moves when the method moves and says nothing about it.

Four figures per horizon rather than two. The mean-district/total split is the finding, and the pre-closure restriction is not a sample of the same question — it is a different one. A forecast whose target is FY2021 or later is scored across a statewide level break that no method fitted before it could see. Both populations are published and neither is labeled the corrected one.

Absent rather than zero at the deep horizons. No origin reaches ten years without crossing the closure, so the pre-closure pair is null from ten to thirteen. The panel cannot say what this method does at ten years in a decade that did not contain a school closure, and the alternative — printing the crossing figure under the pre-closure label — is the one thing that would make the two columns incomparable.

Log errors, not percentages. The field is ln(point / actual), and a consumer wanting a percentage takes exp(bias) - 1. Log errors average without the asymmetry that makes a mean of ratios depend on which way round the ratio is written, and the total’s figure is a log of a ratio of sums, which has no percentage form that survives being averaged with the others.

The feed contract breaks to 46.0.0. One new array on the projection block. Breaking rather than additive by the rule the-shrunk-rate set: a consumer reproducing the projection block should be told that what the block says about its own reliability has changed, and the version is where it finds out.

Eight places, not four. num rounds to four, which renders the pivotal figure — the total’s +0.00057 at five years before the closure — as 0.0006, a value a reader cannot distinguish from zero and a consumer cannot reproduce. The bias rows go through share/opt_share, which is the trap the-shrunk-rate recorded and the second field to fall into it.

The FY2032 leg states its own, which is #431’s fourth question answered. scenario/guarantee-phase-out carries both the FY2032 band and this backtest, so the bias is stated in the property that qualifies the projection rather than on a node of its own. Its leg is six years, an interior point of a curve whose other pins are its ends — pinned anyway, because the alternative was a reader standing at the FY2032 band with figures for one, nine and thirteen years and none for six. This record stated those two numbers in prose before they were pinned, which is exactly the defect an-example-is-reachable-by-no-gate names.

/method states it beside the point rather than in the band. The forecast card’s footnote carried the 5.8% as a sentence; it now draws all four lines against zero, and the two panels are the two populations. The band itself does not move, and the center does not move.

Twelve pins, in ratio and not share. Two curves at three figures a line is twelve, and the document carries ten of them plus the two the FY2032 leg needs: horizons one to three are the same forecasts in both populations, so the two curves name one pair of first-point figures between them. The unit is Unit::Ratio — a log error is dimensionless and is not a fraction of one — and that is also the only unit available, because Unit::Share is refused below two points and most of these are under it. ratio renders bare, so the prose that states them carries no % and no scale word, which is the same discipline as Log errors, not percentages above, enforced by a gate rather than by care.

The two negative figures are pinned as magnitudes. crates/figures.json cannot bind a negative — its numeral regex captures no sign — so the one-year pair takes -negative in its key, .abs() in its compute, and a U+2212 in the prose beside it. That is a limitation of the binding mechanism and not a judgment about the figures, and it is recorded here because the sign is the whole of what this record is about.

Four of the twelve name a value another pin already names. Every line here runs monotonically away from zero, so each one’s worst departure is its last point. That is the endpoint rule answering rather than idling: a line that turned back would make those four different numbers, and the pin that looks redundant today is the one that would catch it.

The per-district fan carries it now, by #461. This record left the district route out: /district/<irn> drew a projection fan per district and said nothing about bias, and #445 named that as a third delivery. #445 closed on its two halves and handed the third to #458, which #461 delivered.

The fan states the mean district’s figure and never the total’s, for the reason this record gives for publishing both: a district’s aid is a district-level quantity, and the two carry different biases. It states both populations at the fan’s own horizon rather than choosing between them, and says in so many words that the figure is a mean over districts and not this district’s own bias. Where the band collapses — a district whose aid enrollment does not enter — there is nothing for an enrollment-forecast bias to be about, and no sentence is printed.

#458 framed the work as a horizon mismatch, a ten-year fan against a pre-closure panel that stops at nine. The fan is a six-year leg, base_year + 6, and always was, so the mismatch did not exist; the correction is recorded on #458. Nothing here moved the fan, its band or its center.

Correct the point. The question #423 asked, and every version of it failed on measurement — the file that measured them is the context above. The shortest statement of why: sorted on the rate to the origin or on the district’s size, every quarter of districts is forecast high by about the same amount, so there is nothing a forecast could condition a correction on.

Publish one figure per horizon. The obvious simplification, and it is the one thing the measurement forbids: the two published quantities differ in sign at four years and by an order of magnitude at five. Whichever one was published, some reader would apply it to the other.

Publish only the pre-closure figures, as the “clean” ones. Rejected. They stop at nine years and the feed forecasts to ten, so the horizon the feed actually publishes would have no figure at all. The closure is in the record and a projection made today runs through it.

Publish percentages rather than log errors. Friendlier and lossy; see the decision. A percentage per horizon is a derived view and the page computes it.

Leave it in the test file and quote the numbers. What was already happening, and what the first half of this work was undoing on the coverage side: three artifacts quoted figures a test produced and nothing recomputed them.