• Status: proposed 2026-10-15. What it decides was ruled before it was written: the maintainer ruled the decision below on 2026-10-15, on #496. The text is not yet accepted, and nothing in it is implemented. #496 stays open: this record answers its question, and the issue closes when the code lands
• Shipped: unreleased
• Date: 2026-10-15
• Opens: nothing on landing. No code, no SCHEMA_VERSION, no OUTPUT_VERSION. The code that implements it is a break, said under Three things this carries
• Amends: ADR 0028 §4, which left a direction for samples to an amendment that says which half wins and why; the amendment is there, in its own words. ADR 0006 §5 and §6 carry dated notes pointing here
• Assumes: ADR 0006 §5 (the floor is the baseline's, never the current run's), §7 (the aggregate's interval) and §12 (agreement counts an errored sample against itself)
• Touches: nothing in CLAUDE.md's fixed section
• Number: 0050. Swept on 2026-10-15, before a line was written, across origin/main (2549e97), every local and remote branch and tag (81 refs, after a fetch with --prune), the worktrees of this repository and the open pull requests (none). The highest number taken anywhere is 0049

Context

digline#496.

What the code does. _noise, in src/digline/core/compare.py, builds the interval [sample_min, sample_max] of the baseline's individual samples, and Score, in core/types.py, refuses to exist unless those two are the minimum and the maximum of its rounded samples. The value tested against that interval is the current run's score, which for a case's check is the mean of its judged samples, not of all K. Noise.covers is inclusive at both ends, at storage precision. compare() reaches the floor last: a verdict present on one side only, an error, a flipped outcome, a moved denominator and the declared tolerance are all decided before it.

First consequence: the floor can only widen the invisible zone. unchanged is within the tolerance or within the floor, a disjunction tried in that order. The floor can only add movements to unchanged, never take one away. It is not a filter; it is an enlargement.

Second consequence: the two sides are not the same quantity. On one side, the observed range of single draws. On the other, a mean. The range of draws does not shrink as K grows — more draws can only find a wider minimum and maximum — while the spread of their mean shrinks with K. So as K rises the floor widens while the quantity it judges becomes steadier. A noise floor should narrow as the measurement improves. This one does the opposite, by construction, and no particular suite is needed to see it.

Third consequence: for a binary check the floor is either nothing or everything. When a check's samples are 0 or 1 — which is not so when Score.sample_means is true, where each sample is already a mean — sample_min, sample_max is (0,0), (1,1) or (0,1). A unanimous baseline gives a point, with no effect. A baseline that is not unanimous gives [0, 1], which covers every possible score. For a check whose declared tolerance is 0, the whole side of the threshold the baseline sits on becomes invisible: every movement that does not flip the outcome is reported unchanged, with no tolerance to appeal to. A check written with tolerance=0 to say this must not move at all is the check the floor silences.

Fourth: for a binary check, below unanimity the floor concedes every same-side movement there is. Unanimity is a property of the outcomes, not of the scores: a binary check's samples are 0 and 1, so a baseline divided on outcomes has the interval [0, 1], which covers every score, and the floor then excuses the largest same-side movement the configuration permits. Measured at K=5 and a threshold of 0.5, the largest same-side movement and the largest the floor excuses are the same number in every case below unanimity: 0.5 at min_agreement 2/5 and below, 0.4 at 3/5, 0.2 at 4/5, and nothing at unanimity, where a baseline divided on outcomes cannot exist.

So the question is not how much the floor adds. It is how large a same-side movement the configuration permits, and that is set by min_agreement, which has no default: a suite that samples must declare it or it is refused, because "the answer would default to 'any'". min_agreement is measured over all samples, errors included, so it errors the verdicts whose agreement falls below it — those nearest the threshold, and those with too many samples that could not be judged — and that is what bounds the movement.

What makes the floor unreachable for such a check, then, is a declared tolerance at or above that number, and nothing else. One suite can run for months with the floor affecting no comparison, and another, differing only in two numbers somebody chose by hand, can have every same-side movement excused by it.

For a check that is not binary this says nothing. A baseline divided on outcomes there has the interval of the scores it saw, not [0, 1], and the floor concedes only what that interval covers: a rubric baseline whose samples span 0.45 to 0.6 does not excuse a same-side movement of 0.33. What holds for every check, binary or not, is the second consequence.

Out of scope, and named. A run-level aggregate's floor. with_noise_interval in core/aggregate.py takes its interval from aggregates computed index by index while the recorded score comes from already-combined verdicts (ADR 0006 §7), so there the compared value is not a mean at all, and a committed baseline has been shown whose recorded score sits outside its own interval. Nothing here decides that, and it is not the same defect.

What this ADR does not claim. How often each case occurs in practice, and what min_agreement real suites declare. Neither is measured here.

The decision

Ruled 2026-10-15. The measured floor stays, and two things change.

1. It compares means with means. The quantity on the baseline's side becomes the dispersion of the mean of the judged samples, estimated from those samples, instead of the range of the samples themselves. The floor then narrows as K grows, which is what a floor is for. This is the direction ADR 0028 §4 left open, and the amendment there says which half wins and why.
2. It applies only where an assertion declares it. Today every assertion is given a widened unchanged zone whether or not its author asked for one. After this, an assertion that declares no floor is decided by its tolerance alone.

The answer ADR 0006 is owed. ADR 0006 refused variance and standard deviation because "it assumes a distribution, and five samples are not enough to earn the assumption". That is right, and it is not an argument for the floor as built: the range of five draws is itself an assumption, that it bounds where their mean can be, and it is a worse one, because it widens with K. So the choice is not between assuming and not assuming. It is between an assumption that is declared, with a width written down, and one that is silent and runs backwards. What makes a poor estimate acceptable is that it is asked for: whoever declares a floor at K=3 is declaring that they accept an estimate made from three numbers, and digline is not imposing it on anyone.

Three things this carries.

• A break. A check with tolerance=0 and a baseline that is not unanimous will begin reporting movements it used to call unchanged. That is the point of the change, not a side effect, and it needs a release note and a version decision by RELEASING.md's rule.
• A new declared quantity. A dispersion-based floor needs a width — one standard error, two — and that width is declared, never inferred. Whoever writes it owns the number.
• Binary checks stay the tolerance's business. For a binary check the dispersion of the mean is sqrt(p(1−p)/K), about 0.179 at p=0.8 and K=5: still wide. The documentation should say that a binary check is served by its declared tolerance, rather than leaving each author to find it out.

Not decided here. The migration: whether a suite keeps the floor by default for one release with a warning, or loses it at once. That is a release question.

Test plan

The fourth point's figures come from an enumeration, and a figure in a public text regenerates from an artifact that runs in CI: tests/test_noise_floor_concedes_every_same_side_move.py. For a binary check at K=5 and a threshold of 0.5, it folds every count of passes, fails and samples that could not be judged with combine_samples at each min_agreement five samples can produce, and puts every same-side pair through compare() at tolerance 0. It holds the coincidence — the widest same-side movement and the widest the floor excuses are the same number — at each min_agreement, that at unanimity no baseline is divided and nothing is excused, and the five figures above.