Real bit depth

The app gives a binary Upscaled flag. A negative flag means that neither of the signatures below was established; it is not a certificate of recording history. Detection results are binary. Unavailable checks are explained in the detection detail; file read failures remain explicit errors.

Exact unused bits#

Every decoded integer sample, on every channel and through the end of the file, contributes to a bitwise OR. Trailing zero bits identify unused precision exactly. For example, widening 16-bit PCM to 24 bits by multiplication by 256 leaves eight zero low bits. Negative samples are handled as two's-complement integers.

The existing silence rule is deliberately preserved: an entirely zero signal reports one effective bit and is flagged when the declared width is greater. One bit is the display convention, not evidence of a one-bit recording. An empty or broken stream is a decoding error, not digital silence.

Integer values must remain integers for this calculation, including 32-bit PCM. Float samples still feed spectral analysis, but must not be used to reconstruct the integer low bits. Floating-point sources have no integer depth measurement.

A grid surviving low-level export noise#

An export can add a small residual to a widened signal. Then the OR correctly reports all occupied bits, including that residual, but misses the lower-depth grid. For a candidate step S = 2^(declared_bits - candidate_bits), the second check measures the distance to the nearest multiple of S using integer arithmetic.

Requirements are deliberately conservative:

• Candidate source widths are at least 8 bits; the gap is at least 6 bits.
• Every sample must be within S/16 of the grid. A single outside sample, including in the last packet, permanently rejects that candidate.
• At least three disjoint windows of 4096 samples per nonzero channel must each span at least 64 grid steps and populate at least 32 of the 64 coarse-code residues. Few-level signals, DC plus tiny noise, tiny signals, sparse impulses, and silence cannot supply this evidence.
• Each channel is measured separately; the file takes the largest required depth. A native channel cannot be outvoted by padded or silent channels.
• All samples are checked even after enough rich windows have been found.

These are engineering thresholds, not published confidence levels. The radius is far smaller than half a grid step: with a 16-bit grid in 24-bit PCM, accepted residues cover only 33 of 256 possible integers. Real music is not independent uniform noise, so this ratio is not a false-positive probability.

The supplied Audacity export contains 16,558,080 samples. Independent FFmpeg PCM measurement found a residual between -10 and +10 destination LSB, RMS 2.050271, around a 16-bit grid spanning 64,972 distinct codes. All 24 bits are occupied. 142 samples exceed +/-8, which is why a strict eight-LSB bound would miss it. The chosen +/-16 bound covers this file; the unit/integration corpus tests both positive constructions and counterexamples rather than using this file alone.

The app displays this result as ≈16-bit, keeps exact occupied depth in bit_depth_evidence.stored_bits, and explains the residual in the detection detail. Saved JSON retains both values. CSV exports also keep bit_depth_method and stored_bits, so estimated depth can be distinguished without changing the numeric real_bit_depth column. Old JSON reports without the new evidence field still load.

Limits and rejected approaches#

An intentionally generated grid-like signal, or 16-bit-grid music mixed with a genuine signal weaker than the accepted residual, can be indistinguishable from an export with dither. Hence the grid result is an estimate, not an assertion that the residual must be dither. Stronger noise, gain changes, resampling, or other processing can erase the grid and lead to a negative result.

Do not infer source bit depth from a -98 dBFS noise floor. The familiar 6.02*N + 1.76 dB equation describes an ideal quantizer driven by a full-scale sine wave under specific noise/bandwidth assumptions. Acoustic noise, gain, dither, FFT bandwidth, and window normalization defeat a universal threshold. A noisy native 24-bit recording is not thereby upscaled. A low spectral floor also does not certify a native master. The PDF's early exit after 100,000 samples is rejected because a padded intro says nothing about the rest.

Sources and reproducibility#

Validation on 2026-09-23: all 296 Rust tests (including 8 new cross-module upscaling tests and the expanded integer-decoding tests), TypeScript checking, frontend production build, 12 frontend regression tests, and Clippy across all targets passed. A release-mode probe through the app's analysis code returned:

These results validate the supplied cases and synthetic corpus, not a universal error rate across all recordings. Existing displayed or imported results must be reanalyzed with a freshly built app; an installed older binary is unchanged.

The supplied 13-track No More Tears album (24-bit / 96 kHz) was also checked through the complete file-analysis pipeline. Every track reports 24 effective bits, Clean, no detection flags, and no read error. Unavailable checks remain explicit in the detail and do not change the binary summary. Saved reports with unsupported detection statuses require reanalysis; obsolete results are not imported or converted.

• FLAC specification, RFC 9639, wasted bits: exact low-bit padding and restoration; FLAC supports 4–32 bits.
• Audacity Dither manual: float processing and export dither. Its illustrative +/-3 figure does not bound the supplied shaped export, measured at +/-10.
• Analog Devices: data-converter noise budget: conditions behind SNR and ENOB, distinct from storage width.
• Lipshitz, Wannamaker and Vanderkooy, Quantization and Dither: A Theoretical Survey (1992): dither properties and their statistical assumptions.

Implementation is derived from integer arithmetic and these principles, not another detector's source. Run cargo test -p flaccompagnon-core --offline for the synthetic regression corpus, and the following read-only probe for a real file:

cargo run --release -p flaccompagnon-core --example bitdepth -- /path/to/track.flac