lesson-proposal.md

back to table · edit · history · where entries came from · files · download

2831 bytes, as of the version from 2026-11-18 01:45 (current). Recorded here, not run.

## Requesting an Arb integration tolerance does not guarantee the returned width

What happened: long transformed and direct quadrature panels for T741 returned finite Arb enclosures with only 29–40 decimal digits although abs_tol and rel_tol requested about 120. Fixed width-four panels restored about 120 digits without changing the integrand, precision or mathematical definition. The first output serializer itself shortened decimals to supported precision, rather than enforcing the table's requested 100 digits.

What the skill says now: check finite, non-vacuous enclosures and measure retained precision; extra working precision is not an error bound.

What it should say: a requested quadrature tolerance is a goal, not a certificate that the returned enclosure achieves it. Check the actual enclosure accuracy before trusting an offline dry run. Subdivide troublesome transformed intervals before merely increasing precision or lowering the requested digit count.

Evidence: run-numerics.log and diagnose.log show n=12 results with 98–130 accuracy bits and returned bounds around 1e-33 through 1e-42. run-numerics-v2.log records the width-four-panel replacement; generate.py explicitly rejects insufficient accuracy. These are contributor-side Sage/Arb observations, not production-server behavior. No live skill change is proposed without a regression test and human promotion.

## Preserve enclosure width when rounding a ball's printed center

What happened: numberdb._write._ball_text formats a center at the requested significant digits and the original radius at three digits, without adding the center's rounding error. For a 120-digit Arb result exported with a 110-digit center, the discarded center digits can exceed the printed radius by many orders of magnitude. This run caught the issue in offline checkpoint construction before upload. It adds one exact decimal ulp of both rounded fields to the radius. The table entries use the normal decimal encoding instead.

What the skill says now: never round endpoints until they agree or turn a finite-width interval into an exact singleton; all enclosures must include propagated uncertainty.

What it should say: rounding a ball's center introduces error too. A printed center-plus-radius enclosure must absorb rounding of both fields, preferably by directed endpoint rounding or a conservative exact ulp expansion. A client accepting or emitting that representation does not prove inclusion.

Evidence: clients/python/numberdb/_write.py, _ball_text; original enclosures in run-numerics-v2.log, exact outer-enclosure transformation in continue_numerics.py, and expanded records in values.json. No client or skill edit was made in this immutable snapshot. A promoted fix needs a regression demonstrating containment of the original ball after serialization.