Commit Graph
2 Commits
Author SHA1 Message Date
Marc-André Bombeck 75c43c74bc docs(mood): state the validation's real leak boundary, drop the unused weekday floor, correct two comments
Four honesty fixes with no behaviour change to the math.

The validation header claimed more than it proves. The ridge FIT is
out-of-sample and no training fold sees ahead — that holds. But feature
SELECTION and per-user standardisation both run over the whole window
before the split, a small documented optimism (measured at about one
part in sixty on the error). The header now says exactly what is and is
not leak-free, and says to nest the screening if the feature set ever
grows or the optimism is ever measured larger.

The per-weekday observation floor had no consumer: no weekday board is
built on this fit, the forecast only publishes seasonalUnlocked and
leaves the weekday view to the insights page, which has its own floors.
Removed rather than left as a number nothing reads, with a note to bring
it back alongside such a board.

Two comments were wrong and are now true: the route's parameter is
declared for the exported type, not to steer a method check the wrapper
resolves from the Next.js request with a GET fallback; and the label
surface imports the key parser from feature-keys.ts, not features.ts.

And the write-surface guard now states its own structural blind spot,
the way the session-surface guard does: the wipe path deletes this table
through a dynamically-resolved delegate the literal matcher cannot see,
which is why that deletion is correctly not counted as a writer.

The Coach staleness fixture moves above the forecast floor, because the
read path now refuses below it: a one-entry account correctly gets an
absence rather than the present forecast that case exercises.
2026-08-08 15:30:38 +02:00
Marc-André Bombeck 64c077f7ac feat(mood): build the day matrix and fit it without letting it see ahead
The target is the day's own pleasantness; the inputs are the other four
self-ratings, the numbers and closed vocabularies of the day's context,
and the figures the sleep, activity and recovery modules already hold.

Nothing is imputed. A day missing one of the selected inputs is a day
the fit does not see, and a feature nobody recorded on most days never
becomes a column in the first place — the coverage floor is what keeps
the complete-case rule from quietly emptying the matrix. A day with no
context row contributes no work-status column at all, rather than a zero
that would read as "worked a regular day".

Columns are standardised against the account's own window, so a person
whose stress never passes six does not have six read as the maximum.
Candidates are screened with the machinery the rest of the product
already runs — Pearson at the shared paired floor, Benjamini-Hochberg at
the shared FDR, the shared effect-size floor on the shrunk estimate —
because a sweep over sixty features at raw p < 0.05 finds three things
that are not there.

The fit is a ridge solved in closed form, validated on expanding windows
in date order. The training range ends exactly where the validation
range starts, asserted on the fold indices and on the day keys, not on
the resulting error: a leaking split passes an error check whenever the
leak is small, which is when nobody looks. The band is the central 80%
of the validation residuals, so it widens when the model is bad instead
of staying confidently narrow.

The window read is the linked resolver's own engine, extended to answer
many days in one query rather than growing a second opinion about what a
Tuesday looked like.
2026-08-08 11:25:34 +02:00