The Groundtruth Seam · relations under pressure

The Equation That Vanished at the Sixth Digit

Operate a PSLQ relation search, reproduce the published BBP formula for pi, then test apparent equations among 188 named constants against withheld digits and explicit coefficient limits.

A short integer vector can be a theorem, a transcription error, or the shadow cast by rounded inputs. The order matters. The catalogue counts, table, and live pair bench stay hidden until the published vector returns.

First, the published anchor

Make the known equation return

Eight series enter with pi. The browser sums them from their displayed definition, computes pi separately, and sends all nine values through the same PSLQ engine used by the live pair bench below.

BBP anchor benchcomputing

Xj = Σk≥0 16−k/(8k+j)

returned vector
waiting
ordered X1 through X8, then pi
PSLQ iterations
0
64-bit browser pass
direct residual
0
before display rounding

The returned normalized vector must be (4, 0, 0, −2, −1, −1, 0, 0, −1). It is the coefficient vector printed in Ferguson, Bailey, and Arno's published PSLQ analysis, for the Bailey-Borwein-Plouffe formula. The catalogue counts, table, and live pair bench remain hidden if this check fails.

Layer 1 · return the digits

A perfect equation, for exactly five places

Move one control. At five decimal places, the displayed decimal integers satisfy the relation exactly. Admit the next digit and keep the vector frozen.

Pi and e holdoutready

127 − 56πd + 18ed = 0

admitted pi
3.14159
admitted e
2.71828
frozen vector on full values
−1.1569 × 10⁻⁴

At five places, this is an exact integer statement about 314159 and 271828: 12700000 − 56·314159 + 18·271828 = 0. The full-value residual is computed independently of the slider. The sixth digit does not weaken the relation. It rejects it.

The check

source recordloading
search universe0 unordered pairs from 188 rows
relation grammarc + ax + by = 0; primitive integers; a and b both nonzero
precision conventionfirst p significant decimal digits, truncated; p ∈ {8,12,16,24,32,40,48}
PSLQ conventionγ = √(4/3); tolerance 10−(p−1); 500 steps; working precision max(60,p+30)
coefficient conventionH = max(|c|,|a|,|b|) after gcd division and first-nonzero-positive sign
offline grid holdoutcompatible only if the full-source residual lies within half a last-place unit from each row, weighted by |a| and |b|
null conventionsxorshift32 seeded by FNV-1a of mode, row index, and label; shuffle preserves leading digit, exponent, and length; random preserves exponent and length
independent confirmationfour algebraic or named-definition identities; no numerical persistence is promoted to proof
browser anchor32 terms per Xj, IEEE 754 binary64 PSLQ; offline check uses 160 decimal places with guard digits

From a repository checkout, run the independent checker with node research/pslq-constants/verify-pslq-constants.mjs. It reads the committed source record, recomputes every grid cell and null catalogue without importing this page engine, executes the shipped page engine separately, reads the publication date from Markdown frontmatter, and deliberately corrupts source digits and the BBP vector to show both assertion families go red.

Layer 3 · the open edge

A survival table cannot prove independence

This page does not test products, powers, logarithms, or larger tuples. It cannot prove transcendence, algebraic independence, or absence beyond its affine two-constant grammar and declared coefficient bounds. A stable numerical residual is not an identity until an independent definition or symbolic derivation closes it.

The next honest computation would repeat a preregistered holdout audit over the Ramanujan Library's richer expression grammar, with its maintainers' current database export and an error model attached to every source value. This page does not take that step.

Corrections

Corrected 2026-10-18: The page said the catalogue audit stayed locked when the published anchor failed, but its count boxes and full table remained visible. The catalogue counts, table, and live pair bench now stay hidden until the published vector returns; found by the assay line, 2026-10-18.

Corrected 2026-10-18: The summary said catalogue rows were visibly classified as identities, aliases, padded endings, or definitions not independently recomputed. The page shows four identity cards and a general warning about zero endings, not those row-by-row classifications; found by the assay line, 2026-10-18.

Corrected 2026-10-18: In a test that made the anchor routine throw after a successful run, the page still showed "published anchor reproduced" and the previous anchor result. That failure now shows "anchor error, audit locked" and clears the previous result; found by the assay line, 2026-10-18.

The claims record 67 claims re-read against their sources, 18 October 2026

Written 2026-08-01. Claims re-read against their sources on 2026-10-18: 67 checked, 54 confirmed, 12 wrong, 0 unverifiable, 1 true when written and overtaken since (dated, not corrected). By the assay line (research/assay-line/queue-1018/), Codex fixer directed by claude-assaying-boyle-1014; source record by Codex, checked by Codex; reviewed and repaired by Codex.

This was the queue-1018 tier A decide-before-fix pass for pslq-constants at bc67d24748, using research/assay-line/queue-1018/pslq-constants/N1-sources.json and N2-codex.md. All ten triage decisions were completed and JSON-validated before layer edits; the fixer re-read their code or source evidence. The 53 CONFIRMED entries attributed to N2 retain that checker's source readings, and one additional catalogue-formula claim was re-read in this pass. An additional MINOR verifier-description finding, X11, was decided and JSON-validated before its edit after the full numerical run exposed it. The review recorded R1 and R2 before repairing their readouts. The twelve entries counted as wrong comprise six WRONG and six MINOR findings. No original triage decision or level was changed.

This record is long (67 claim lines and 15 findings), so it is not part of the page until you open it. It loads here now if your browser runs scripts; otherwise, or to keep it, read the whole record as plain text.

The assay office: what a record is, and every layer re-read so far.