lessons-20261117T035346Z-build.md

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

4982 bytes, as of the version from 2026-11-17 04:25 (current). Recorded here, not run.

# Contributor lesson proposals from T732

These are artifact proposals, not changes to the skill or source snapshot.
Policy read: ../20261117T035346Z-build-skills/manifest.json.

## Enclose printing error when serializing a ball

What happened: this snapshot client's `_ball_text` prints the midpoint to the requested significant digits and independently rounds the radius to three digits. When the midpoint is shortened substantially, the printed center plus/minus printed radius need not contain the original ball. The T732 quadrature diagnostics exposed this. Ordinary decimal serialization was used for final table values; saved diagnostic checkpoints were recovered by adding a full ulp of the printed center and of the printed radius. An exactly printed zero midpoint has zero midpoint rounding error under this significant-digit formatter.

What the skill says now: a real string with explicit radius denotes that exact center and radius, and precision must be honest.

What it should say: serialization of a bounded number must itself be outward. After rounding a midpoint, add its rounding displacement to the radius, then round the radius upward. Until the client is repaired and tested, use its ordinary decimal interval encoding or preserve exact rational endpoints instead of assuming `form='ball'` always preserves the enclosure.

Evidence: clients/python/numberdb/_write.py `_ball_text`; builder-check2.jsonl diagnostics, quadrature-checkpoints.json, numbertext.py recovery and builder-check5.jsonl in this run. Proposed regression: serialize a narrow ball with midpoint carrying many more digits than requested, decode it and require containment of both original endpoints.

## Preserve an exact zero component through the complex client path

What happened: a rigorously proved real component zero became `[0, 0]` in the client's complex text. Its precision counter then treated that exact interval as a short unrecognized component and rejected a 100-digit imaginary component. Using public `ComplexInterval(0, RealInterval(lower,upper))` preserves exact rational zero and the rigorous imaginary enclosure without altering the client or returning a bare string.

What the skill says now: omit a component only when it is exactly zero; exact and enclosed components can coexist.

What it should say: when a client version loses an exact component's semantics during serialization, preserve the component using the public exact-plus-interval representation and check the written output. Do not replace a proved zero by padded decimal zeros or weaken the numerical claim to satisfy a formatter.

Evidence: final-validation.log, generate.py and export.py; final 262-entry reproduction and stored-check-results.json. Proposed regression: a complex value with exact zero real part and a narrow nonzero imaginary interval must retain the latter's supported digits.

## Use stable complex-ball polynomial evaluation

What happened: long Horner evaluation of a q-primitive produced a finite ball with radius about 10^53 at level 15 despite a 480-bit working field. Rectangular ball wrapping amplified uncertainty. Separately bounded powers gave narrow balls throughout the selection. Independent quadrature uses rectangular splitting.

What the skill says now: reject vacuous balls and investigate cancellation, truncation and precision loss.

What it should say: for a long complex polynomial, monitor enclosure widths as well as midpoints; Horner evaluation can be unstable for rectangular enclosures even when |q|<1. Separately bounded powers or suitable block evaluation can preserve the required bounds.

Evidence: debug.log, computation4.jsonl (failed width), computation5.jsonl (corrected evaluation), final-validation.log and heegner.py. Proposed regression: a representative q near the unit disk boundary must give a finite narrow enclosure and agree with independently integrated coefficients.

## Programs requires language and code records

What happened: a Programs mapping from a language name to a plain string passed the server audit but prevented preview rendering. A labelled record with `language: Sage` and `code: ...` rendered correctly. The short example also needed an explicit base ring in `EllipticCurve(QQ, coefficients)` to avoid this modular Sage startup sequence's PolynomialSequence circular import; it was run successfully before the final upload.

What the skill says now: Programs should show a standard invocation for another value; its record schema is not illustrated.

What it should say: each labelled Programs item contains language and code fields. Run the exact displayed code in the supported runtime and inspect its rendered block; a clean audit need not establish either property.

Evidence: program-example2.log, program-example.py, program-fix-response.json, render-program.html and render-results.json in this run. Proposed regression: the validator should reject a plain-string program record before it reaches the renderer.