Method, minus the lab coat

Resemblance is searchable.
Explanation is not.

A result answers one narrow question: after declared transformations, did two vectors move together over shared coordinates? It does not say why—and it certainly does not say “because.”

Follow one query

THE WHOLE SYSTEM IN ONE LINE

cheap approximate retrieval finds what to inspect; exact local statistics decide what to show.

Align

Infer a coordinate schema, preserve missingness, and compare only compatible axes.

Fingerprint

Transform, PAA-compress, and project the query into eight deterministic band keys.

Retrieve

Fetch eight compact static shards and let matching bands vote for at most 5,000 candidates.

Rerank

Fetch candidate values, then compute exact lag, overlap, correlation, and fragility locally.

00 / before the clever bit

Coordinates first. Numbers second.

A vector is meaningful only over a named coordinate space. Annual years compare with annual years; monthly dates compare with monthly dates; geological age bins compare only when the bin schema and dating precision are compatible. Dinosaur occurrence bins never quietly line up with modern GDP years.

Unknown observations remain missing; they never become zero.
Windows below the release coverage threshold are rejected.
Small interpolated gaps are recorded in public quality flags.
Resampling follows stock, flow, rate, or source semantics.

PAIRWISE-VALID OVERLAP

O = \{i : x_i \text{ and } y_i \text{ are both valid}\}

z_i = \frac{x_i-\bar{x}_O}{s_{x,O}} \qquad \operatorname{coverage}=\frac{|O|}{|W|}

Every result shows |O|. High correlation with six surviving points should look suspicious because it is suspicious.

01 / shape retrieval

Turn one vector into eight cache keys.

The selected representation—level, difference, rank, detrended, seasonal residual, or robust shock—is PAA-resampled to 128 values, normalized, and projected against a deterministic 128-hyperplane matrix. Signs become a 128-bit SimHash split into 8 × 16-bit bands.

INTERACTIVE 01

Compress a shape, then fingerprint it

PAA buckets in this diagram

Probe the least-certain bit

z-normalized input bucket means

\operatorname{PAA}_j = \frac{1}{|B_j|} \sum_{i \in B_j} z_i \qquad b_k = \mathbb{1}\!\left[h_k^\top p \ge 0\right]

band-000111

band-011110

band-020000

band-030010

band-040110

band-051111

band-060000

band-070000

The diagram samples 32 bits so they fit. Production uses 128 deterministic hyperplanes and eight 16-bit bands; the projection seed and bit order are versioned corpus contracts.

Band votesmore matches → fetch first

Ceremonial umbrella meetings7/8

Ceratopsian fossil bins1/8

Very confident turnip index0/8

Moon support tickets0/8

02 / static retrieval

Eight GETs. No query server.

Each 16-bit band key is split into a high and low byte. The high byte selects one immutable R2 object; the low byte selects one posting list inside it. Matching postings vote for a series and rolling-window offset.

GET/index/v17/month/60/diff_z/band-03/a7.bin

If too few candidates answer, the client may flip low-margin bits and fetch up to eight bounded neighboring shards. Voting stops at 5,000 candidate IDs before any value blocks are requested.

SLSHSTATIC SHARD v1LITTLE ENDIAN

magic + version6 B

band + high byte2 B

257 posting offsets1,028 B

sorted postingsseries_id:u32 + offset:u16

03 / exact rerank

Approximate in. Exact out.

After retrieval, the browser fetches only compact value and metadata blocks. Rust/WASM recomputes Pearson, Spearman, allowed lags, shock correlation, and leave-one-out influence over pairwise-valid overlap. React never performs the numerical scan on its rendering thread.

INTERACTIVE 02

Slide the lag; watch exact correlation answer

Candidate lag0

Exact Pearson r−0.433

Pairwise overlap18

Worst leave-one-out Δ0.073

query candidate at lag 0

ALIGNED PAIRS

r_\ell = \operatorname{corr}\!\left(\{x_t, y_{t+\ell}\}_{t \in O_\ell}\right) \qquad \Delta_{\mathrm{LOO}} = \max_i |r-r_{-i}|

Retrieval only nominates candidates. This number is recomputed from the fetched values, using only valid aligned observations. Auto lag is a freedom—and therefore appears on the receipt.

04 / ranking

Strong is not automatically interesting.

Pure absolute correlation remains available, but the default ranking also rewards usable overlap, stable shape, data quality, semantic distance, and retrieval agreement. Mirrors and near-duplicates are penalized before they can masquerade as delightful coincidences.

“Weirdness” begins as taxonomy distance—not a hidden language model. Moving the slider suppresses the same source family and nearby subject branches while retaining the exact coefficient.

S = |r|\, q_{\mathrm{overlap}}\, q_{\mathrm{stability}}\, q_{\mathrm{quality}}\, q_{\mathrm{weird}}\, q_{\mathrm{votes}} - p_{\mathrm{duplicate}} - p_{\mathrm{lag}}

coefficientevidence qualitysemantic distancepenalties

Conceptual score; release manifests version the actual weights.

05 / statistical honesty

The search itself creates evidence debt.

Search enough candidates, lags, transforms, windows, and smoothing choices and chance will eventually offer a spectacular number. A single-test p-value forgets how the result was selected. We do not.

INTERACTIVE 03

Spend researcher degrees of freedom

NOT A P-VALUE

Candidate universe 100,000Auto-lag radius ±6Transforms tried 3Smoothing choices 1Overlap n 48

ILLUSTRATIVE RECEIPT3,900,000

nominal opportunities to find something impressive

Candidates: 100,000
Lag choices: 13
Transforms: 3
Smoothing: 1

Independence heuristic for the strongest accidental |r|: 0.686

M = N_{\mathrm{series}}N_{\mathrm{lags}}N_{\mathrm{transforms}}N_{\mathrm{smooth}} \qquad \widehat p_{\max}(t)=\frac{1}{S}\sum_{s=1}^{S}\mathbb{1}\!\left[\max_{m\in\mathcal M_s}|r_{s,m}|\ge t\right]

The meter is deliberately a rough independent-test intuition. Real series and nearby windows are dependent, so a calibrated public release will use block-shuffled or phase-randomized surrogates searched through the same pipeline. The empirical maximum—not this shortcut—will drive its warning.

Empirical null

A calibrated public release will bucket surrogates by length, frequency, transform, autocorrelation, and universe size, then search them through the same retrieval and rerank pipeline. Until that gate passes, fixture results say the null is unavailable.

Fragility

Every top result is rerun with each aligned observation removed. The badge says stable, sensitive to one point, or mostly one shock—and can name the influential coordinate.

Dredging receipt

Transform, chosen lag, smoothing, overlap, candidate universe, retrieval path, and freedoms used are shown beside the result, not buried in a methodology appendix.

Try the local engine Browse precalculated absurdities