TELESCOPE

Integrated, short-term and momentary LUFS, LRA and true peak — also per channel. Against ITU-R BS.1770 gating, EBU Tech 3342 for LRA and the literal FIR table of BS.1770 Annex 2 for true peak.

How much headroom is left: PSR over the last 3 s, PLR since reset (AES TD1004), a 61-bin short-term histogram and clip events on a 10-minute timeline. Silence stays out of the histogram: only short-term values above −70 LUFS count.

FFT from 1 024 to 32 768 points with Hann, Blackman-Harris or Kaiser, and free, ⅓-octave (ISO 266) or Bark bands — every bin in dBFS referred to a full-scale sine, with an exact coherent-gain correction.

Level over time: 512 log rows from 20 Hz to 20 kHz, 10 / 30 / 60 s of history and level as colour — the instantaneous power, never an average.

The same data in depth: level as height, time receding from now. A 2.5D oblique projection drawn in software — no GPU — with an occlusion proven identical to the painter's algorithm.

The spectrum by note: constant-Q at 24 bins per octave from A0, a chromagram and a key estimated against the Krumhansl & Kessler profiles — never shown without its confidence and its share of time.

The same constant-Q coiled, one turn per octave: equal notes line up on one radius, so the octave structure of what is playing reads at a glance.

Correlation, width, balance and mono loss over 100 / 300 / 1000 ms, as Lissajous, POLAR SAMPLE or POLAR LEVEL: a half-circle with one ray per degree, where what is out of phase folds onto the base and prints its percentage.

SCOPE's figures per ⅓-octave band, so mono bass and open highs stop averaging into one number. Its band-wide figure, computed from the STFT bins, matches the time-domain meter to 0.00076 in correlation.

The sonogram with phase as colour — red out of phase, green wide, white mono — and level as brightness. A cell with no energy is black, not dim red.

Energy by pan direction × frequency on a 64 × 96 grid, with decay and a trail, in a 2.5D view. The label under its axis cannot be switched off: energy by pan direction · not localisation.

Your mix against a reference track you load, each normalised by its own integrated LUFS: tilt at equal loudness, so the delta sums to zero — checked band by band to 0.006 dB. Drag on the reference’s waveform to compare against just one section of it; the section is saved with the session.

21 deterministic rules over a per-second history: within range, how it will feel, where it translates, and what to check and where — every line with its number and its rule. Until the window has had audio above −70 LUFS it says “waiting for audio”, and nothing else. It reflows at 125 % and 150 % in all six languages, with its accents and a MODE button that names the whole mode. Measurement, not taste.

The fix that matters most in 0.2. In 0.1.0 the editor re-baked its full-window background on every lens frame, 30 times a second — 12–80 ms each, more than a second of the main thread per second at size M on a Retina screen — and 11 of the 13 lenses kept redrawing the same silence with the transport stopped. Now the background is baked once and a lens repaints only when something visible changes. At 125 % and 150 % it also lands on whole pixels instead of being resampled every frame.

It reads L and R, pushes them to a lock-free bus and writes nothing back: 0 samples of latency, 0 tail. Deterministic noise through all 13 lenses, with bypass on and off, in blocks of 1 / 7 / 64 / 4096 samples and from a mono source — 0 mismatches.

The literal 4 × 12 polyphase FIR of BS.1770 Annex 2, and a K-weighting recalculated per sample rate, checked against the published 48 kHz table to 8.9·10⁻¹⁶. The known 4× under-read is written down, not rounded away: EBU test 3341-17 reads −6.3160 dBTP against a real −6.0.

Drop a track and it is measured offline with the same classes: integrated, LRA, true peak, the 30 measured ⅓-octave bands and the per-second history come out equal to the bit — and equal to the bit to 0.1.0, checked against 13 golden dumps of its engine. 60 s of audio in 0.63–1.82 s on an M4 under load (ten runs).

FIELD and POLAR LEVEL show energy by pan direction: no HRTF, no ITD, no azimuth. Recovering where a sound was from a finished stereo mix has no unique solution, and TELESCOPE does not pretend it has one.

VERDICT is a threshold table with the source of every number — no model, no telemetry, nothing sent anywhere: the same signal gives the same report, word for word. Its per-device checks are generic, with no speaker curve inside. No findings is not the same as finished.

A light theme next to the dark one, chosen in the lens strip (THEME) and remembered per user. In the light theme SPECTROGRAM, WATERFALL, STEREO SPECTROGRAM and FIELD keep their dark screen and palette, so the data reads the same. Every axis label is at 4.5:1 contrast or better, with an 11 px floor: 0 of 1 272 below it, in both themes.

A command-line tool on the same analysis engine, in the source code: JSON Lines in — a file, and optionally a section in seconds — and one JSON line out per request, the same bytes on macOS and on Windows. It is not in the installer: it is there for anyone who builds TELESCOPE from source.

The first time the window opens with 0.2 on a Mac, a card inside it introduces EYEPIECE, OVNI’s free Mac app that reads your project and turns it into a snapshot you can paste into any AI. It blocks nothing and shows once. Its button opens EYEPIECE if you have it, or ovniaudio.com/eyepiece in your browser — only if you click it. On Windows there is no card.