PedalScope · Part 2 · Understanding the charts

Summaries

The charts, in words. Every Fingerprint (and every comparison) can carry a short plain-language description of what the numbers say: how it clips, whether it cleans up, what chords will do. The summary describes the whole acquisition, so it renders in exactly two places — the Fingerprint’s overview page and the pedal overview; the individual chart pages link to it rather than repeating it. This page explains where the words come from — and, just as important, where they are not allowed to come from.

Where the words come from

Every sentence in a summary is generated from numbers the measurement pipeline computed — a knee level, an even-to-odd harmonic balance, a cleanup point. The app first extracts those features, then turns them into prose: by default with fixed templates (“generated from measurements”), or — if you opt in on a machine with Apple Intelligence — with an on-device language model (“AI-generated description · beta”). The AI wording is labeled beta because its fluency is still being tuned; its grounding is enforced either way. The language model never analyzes audio and knows nothing about your pedal — it only rewords the evidence it is handed, and a validator rejects any draft that invents a number or claims circuit details. The Evidence disclosure under each summary lists exactly the values the words were built from.

Three layers, and each is honest about its job:

  1. Features — deterministic numbers from the measurement products (harmonic distortion, compression, transfer curve, chord IMD, gain map). Whatever wasn’t measured simply isn’t there: a record with no two-tone measurement gets a summary that says nothing about chords.

  2. Perceptual metrics — brightness (“sharpness”) and texture (“roughness”) computed on the same loudness-matched renders the Hear-it panels play. These are simplified proxy implementations of published psychoacoustic models — good for comparisons, not certified absolute values, which is why their evidence rows say “(proxy)”.

  3. Wording — a versioned lexicon maps feature ranges to the vocabulary the description may use. The default tier is Measurement templates only: deterministic wording, no language model involved. The AI-worded (beta) tier, chosen in Settings (⌘,), adds fluency, never information; if Apple Intelligence is unavailable or a draft fails validation, the template tier renders instead and the chip under the summary says which tier actually worded it.

Settings holds all three controls: the Show summaries toggle (hide them entirely if you’d rather keep the charts to themselves), the wording tier picker described above, and — beneath them — the grounding footnote restating the contract in one line: whichever tier words a summary, every number must trace to a measurement, and the language model only ever rewords evidence it is handed.

For the curious: the validator

After the language model drafts a summary, plain code extracts every number from the text and checks it against the evidence packet (rounding is tolerated; fabrication is not). A second check rejects circuit-topology claims — component names, chip types, clipping-element chemistry — because a black-box measurement cannot license them. A rejected draft gets exactly one retry with the violations spelled out; if it fails again, the deterministic template renders instead. The same validator runs in the project’s test suite against deliberately corrupted drafts.

Reading a comparison

Comparison summaries are built from both records’ evidence plus their computed differences, so their claims are relational — “clips about nine dB later”, “carries more even-harmonic content” — which restates the measurements directly. The last line, After loudness matching, is the one to trust most: it lists only the differences that should survive once both devices are matched to equal loudness, because louder-sounds-better is the oldest trick in gear comparison.

What the caveats mean

Summaries hedge on purpose. A feature extracted from a short live-mode sweep, a hysteresis reading with no same-setting Harmonic Distortion measurement to phase-compensate it, or a proxy metric all carry reduced confidence — the wording softens (“appears to”, “may”) and a caveat line says why. A caveat is not the app being coy; it is the measurement telling you how far to trust the words.

The words are heuristics

The mapping from numbers to vocabulary is editorial judgment informed by circuit theory and psychoacoustics — not listener-validated ground truth. Nobody has run a listening panel to confirm that this knee sharpness reads as “abrupt” to most players. The lexicon is versioned (every summary records the version that worded it) so the mapping can be calibrated against real listening data someday without silently changing old descriptions’ meaning.

Common misreadings