back to table · edit · history · where entries came from · files · download
3403 bytes, as of the version from 2026-10-16 06:58 (current). Recorded here, not run.
# Build lessons: 20261016T061956Z, T455
## Printed critical-value tables need an explicit rounding check
What happened: NIST/SEMATECH e-Handbook ยง1.3.6.7.3 prints 238.882 for the upper-tail 5% F critical value with numerator degrees 8 and denominator degrees 1. The exact central-F quantile starts 238.8826948025240912331263462023. Certified hypergeometric inversion and an independent elementary integral recurrence both enclose that value, and mpmath inversion at 150 digits agrees. The printed decimal is within one last-place unit but is not the nearest three-decimal rounding (238.883). Ninety-nine of the 100 compared NIST values pass the nearest-rounding check; all 100 pass the last-place-interval check. We corrected two transcription mistakes in the local test fixture before reaching this count; neither affected the generator or the website values.
What the skill says now: independent numerical comparisons need known controls, disagreements must be explained, and a written decimal denotes an interval extending one unit of the last place on each side.
What it should say: distinguish an external table's printed decimal from an assertion of correctly rounded digits. Preserve a source discrepancy explicitly and independently establish the computed value before interpreting its printed precision. Do not silently change a failed nearest-rounding test to a wider tolerance without documenting the failed row and the evidence.
Evidence: `checks-final.log`, `supplement.log`, and https://www.itl.nist.gov/div898/handbook/eda/section3/eda3673.htm . The generator has certified signs; the checks use finite half-integer beta recurrences without hypergeometric or gamma evaluation.
## API documents can return nested Numbers even after flat records are submitted
What happened: the full table API returned a nested Numbers mapping for T454 and T455. A check expecting a list of flat entry records failed on its shape. The client helper `_stored_entries` normalized both shapes. After normalization every T455 string and its entry order matched the submission; no numerical correction was needed.
What the skill says now: `numberdb.table` returns the whole document with capitalized keys, but does not specify the alternative shapes of Numbers.
What it should say: readers must accept nested parameter trees and flat entry lists. Prefer the client's normalizer rather than assuming that the API readback shape equals the submitted shape. `_stored_entries` is currently private, so attached checks depending on it may need updating if the client changes.
Evidence: `T454.json`, `stored-table.json`, `verify_stored.py`, and `verify-stored.log`.
## An empty client key can discover a default key
What happened: `numberdb.Client(api_key='')` still read an unpublished draft because the client discovered the configured default credential. A plain urllib request without Authorization returned the expected not-found response.
What the skill says now: private draft reads require a permitted credential, but it does not explain how to test anonymous visibility when a default credential exists.
What it should say: do not use an empty explicit key as proof of an anonymous request. Use a request that genuinely omits Authorization, without printing credentials or headers.
Evidence: `public-read-response.json` versus `unauthenticated-read.json`. No publication endpoint was called.