EXPLORATORY · RECONSTRUCTED evidence: negative validation

The Cup Hides the Clock

Many hidden inventory–rate combinations can produce nearly the same final cup. Timed measurements reveal what the endpoint conceals.

Generated from the committed Paper A result bundle.

Static equivalents: text summary · summary image · download the data (JSON). Source: github.com/trbrewer/puckworks.

Loading the generated data…

Summary (JavaScript is off)

This page is interactive with JavaScript enabled. The complete result — every number and every caveat — is in the generated static text summary and the summary image. In short: fitting an extraction model to whole-cup endpoints leaves its hidden inventory and release rate practically non-identifiable, yet the endpoint still predicts reasonably and barely beats a level-only comparator — so endpoint accuracy is not evidence that the mechanism transferred. Timed fractions and an independent inventory measurement recover what the cup discards.

Two plots. Left: fit error versus release rate forms a flat valley, so the whole cup
        barely constrains the rate. Right: scored against timed fractions the error has a sharp trough
        (the rate is identified), while scored against the whole cup it is nearly flat (the rate is
        not identified). The exact numbers are in the static text summary linked above.

Fitting an extraction model to a whole-cup endpoint leaves two hidden things — how much soluble coffee is available (inventory) and how fast it releases (rate) — practically non-identifiable: the fit-error valley is nearly flat along a curve where the two trade off. On the caffeine objective the condition number is · and the inverse-curvature coupling is · (this is objective-surface geometry, not a statistical parameter correlation).

1 · Same cup, different hidden parameters

Move the release rate. At each rate the best-fit inventory adjusts so the final cup stays nearly the same — the fit barely changes across most of the tested range (· of it is near-optimal, and the upper end is right-censored by the tested domain).

rate · · best-fit inventory · · fit · · ·

steps through the near-optimal rate–inventory pairs; the endpoint stays similar.

Fit error (SSE) versus release rate. The shaded band is within 10% of the best fit (the near-optimal set); the dot marks your selection.
Show the numbers behind this plot (text table)

2 · Keep or throw away the clock

A final cup collapses the whole time sequence into one endpoint. Timed fractions retain information about how quickly material appeared.

What the objective is scored against

·

Fit error versus rate. Solid line = timed-fraction scoring (a sharp trough); dashed line = whole-cup scoring (nearly flat). Fraction agreement here is in-sample verification on the model's own calibration data, not independent validation.
Show the numbers behind this plot (text table)

3 · Prediction is not identification

Across held-out grinds the mechanistic model predicts the cup about as well as simply carrying the fitted concentration level — it adds little skill over that null.

Endpoint prediction
mechanistic MAPE · vs level-only comparator ·
Parameter identifiability
inventory and rate are practically non-identifiable at the endpoint (condition number ·)
Incremental mechanistic skill
worse than the null on · of · held-out points

·

Three different things — never one red/green score. A modest held-out error is not mechanistic validation.

4 · What measurement would help?

Better measurement design, not a more confident endpoint fit, is what distinguishes hidden mechanisms.

What this does not prove