# MET-AL — Consolidation
This document records the `integration`-branch consolidation of the seven MET-AL
experiments: a shared verification-math library, a single-file inliner, the
double-clickable `dist/*.html` builds, the per-app de-dup decisions, and the
data/logo trimming. It is a companion to the branded gallery at
[`index.html`](index.html) and the experiment log in
[`docs/EXPERIMENTS.md`](docs/EXPERIMENTS.md) /
[`docs/reports/EXPERIMENT_REVIEW.md`](docs/reports/EXPERIMENT_REVIEW.md).
Everything stays **dependency-free / offline**: vanilla HTML/CSS/JS ES modules,
no build tool, no `npm`/`pip install`, no CDN, no network. `node` is used only for
the inliner and checks.
---
## 1. The shared library — `lib/met-stats.mjs`
`lib/met-stats.mjs` is the **canonical, pure, DOM-free, dependency-free** MET-style
verification math, reconciled from the seven apps. It is the single source of truth
that the migrated apps import instead of carrying their own copies.
### Conventions
**2×2 contingency cell** (matches MET docs):
| symbol | meaning |
|--------|---------|
| `a` | hits (forecast yes, observed yes) |
| `b` | false alarms (forecast yes, observed no) |
| `c` | misses (forecast no, observed yes) |
| `d` | correct negatives (forecast no, observed no) |
| `n` | `a + b + c + d` |
**SL1L2 partial-sum line** (MET's scalar partial-sum line for continuous stats)
carries the **means** `{fbar, obar, ffbar, oobar, fobar}` over its `n` pairs, plus
`n` — exactly as MET writes them. Storing means+`n` (not raw sums) matches MET
`.stat` output and lets the metric be re-weighted by `n` on aggregation.
**The fidelity rule (the round-1 lesson):** aggregating across regions / cases /
time means **SUM the raw counts (categorical) or the SL1L2 partial sums
(continuous) FIRST, then DERIVE the metric** from the combined totals —
"**ratio-of-sums**". Averaging per-group derived statistics ("mean-of-ratios") is
WRONG; `wrongMeanOfRatios()` exists purely to demonstrate the gap on screen.
**NaN-safety:** every metric guards a zero (or ~zero) denominator by returning
`NaN` (never throws, never returns `Infinity`). UIs render `NaN` as a visible
em-dash via `fmt()`.
### API surface
**Categorical (2×2 → number),** each takes a `{a,b,c,d}` cell:
`baseRate`, `fcstRate`, `podRate`, `far`, `pofd`, `sr`, `csi`, `fbias`, `gss`
(GSS/ETS), `hss`, `pss` (PSS/HK), `orss`.
`categoricalFromCTC(cell)` returns the **full bundle** (`a,b,c,d,n` + all of the
above) so the values can never drift apart.
**Continuous from an SL1L2 line** `{fbar,obar,ffbar,oobar,fobar,n}`:
`me`, `mse`, `rmse`, `bcrmseFromSL1L2`, `pearsonFromSL1L2`,
`continuousFromSL1L2(ps)` (full bundle; **MAE is NaN** here — not recoverable from
SL1L2). `sl1l2FromPairs(f, o)` builds a line from matched arrays.
**Continuous direct-from-pairs** (independent path; supplies MAE):
`mePairs`, `maePairs`, `msePairs`, `rmsePairs`, `bcrmsePairs`, `pearsonPairs`,
`continuousFromPairs(f, o)` (full bundle incl. MAE).
**Aggregation (ratio-of-sums, the correct path):**
`sumCTC(list)`, `sumSL1L2(list)` (n-weighted pool of the means),
`aggregateCategorical(list, metricFn)`, `aggregateContinuous(list, metricFn)`,
and the teaching demonstrator `wrongMeanOfRatios(list, metricFn)`.
**Formatting:** `fmt(x, digits=3)` → `NaN`/null ⇒ `—`, `±Infinity` ⇒ `∞`,
else `toFixed(digits)`.
### Self-test
`lib/selftest.mjs` is a standalone, dependency-free harness:
```
node lib/selftest.mjs # -> "met-stats self-check: 129/129 pass"
```
It checks the categorical and continuous formulas against hand-worked cells,
the SL1L2 ⇄ pairs equivalence, NaN guards on zero denominators, and the
ratio-of-sums ≠ mean-of-ratios invariant.
---
## 2. The single-file inliner — `tools/inline.mjs`
`tools/inline.mjs` turns a served ES-module app into **one self-contained
`.html`** that runs by double-click (`file://`) offline. It solves the round-1
cross-cutting finding: Chromium/Safari block ES-module imports over `file://`
(`origin: null` CORS), so the served `apps//index.html` only work over HTTP.
### Usage
```
node tools/inline.mjs apps//index.html dist/.html
```
It prints the byte size + module count, an offline-data-fallback note, and ends
with `inline: OK — output is self-contained` (exit 0) or lists self-containment
problems (exit 1).
### How it works
1. From each ``. Inline module blocks keep their code but get their
relative imports rewritten too.
4. Inline `` as ``.
5. Inline local `` (e.g. the logo) as `data:image/…;base64`.
**Runtime data fetches** (`fetch('./data/x.json')`) are **left alone** — under
`file://` they fail, and the app is expected to fall back to an **inlined `.js`/`.mjs`
data module** that is already in the module graph (and thus inlined). Apps without
such a fallback are flagged on stderr. As part of this consolidation, the apps that
lacked one (`spatial-maps`, `stat-interaction`) gained a `*-inline` ES-module data
fallback wired into a `fetch → fallback` path, so the single-file builds genuinely
run offline.
### Built-in verification (every build)
The inliner self-checks the emitted HTML: **no** `
