この文書はまだ翻訳されていません。本文は英語のまま表示されています。

M0 report: grid and contract

Run stamp: 2026-09-01, first commit of this repository, rustc stable (rustup toolchain), cargo test --workspace 16 tests green, clippy clean. Every number below is pasted from that run or from the cited source, not retyped from the spec; M5 puts them under a check-self-counts scan.

What closed

ntsc-grid exists: exact rational sample rate (f_sc stored as 39,375,000/11 Hz, the grid at 472,500,000/11 Hz), Phase, both Geometry profiles with phase_at and line_len, the CompositeFrame types, and the CompositeSource/Stage traits from spec section 3.4. Phase is a pure function of geometry, per section 3.3: the only free parameter is phase_at_origin.

Residues: every section 3.2 claim confirmed

tests/residues.rs, 12 tests. Confirmed by test, not copied:

• NES line: 2728 samples, residue 4 (120 degrees), each line starting 4 later through phase_at itself.
• NES full frame: 714,736 samples, residue 4; three Even frames return the origin phase, one does not (the rendering-disabled three-frame pattern).
• NES short frame: 714,728 samples, residue 8; Even then OddShort returns the origin from all 12 starting phases, Even alone from none.
• Broadcast line: 2730, residue 6 (180 degrees); frame 1,433,250, residue 6; two frames (four fields) return the origin, one does not.
• The closed-form phase_at matches a brute-force prefix sum over every line of both profiles, and the accumulated totals match samples_per_frame.
• OddShort on the broadcast profile is refused by name at line_len and phase_at (should-panic tests).

Rates, exact

Rate	Exact	Approx
NES full frame	29,531,250 / 491,381 Hz	60.09848 Hz
NES short frame	8,437,500 / 140,393 Hz	60.09915 Hz
NES two-frame average (rendering on)	39,375,000 / 655,171 Hz	60.09881 Hz
Broadcast frame	30,000 / 1,001 Hz	29.97003 Hz
Broadcast field	60,000 / 1,001 Hz	59.94006 Hz

Spec discrepancy, for v0.3: section 3.2 states "Field rate: (6 x f_sc) / (4 x 341 x 262) = 60.0988 Hz for full frames". The formula's shape is right but its quoted value is not: it evaluates to 60.09848 Hz. 60.0988 is the two-frame average with the short frame alternating in, i.e. the rendering-enabled rate a player actually sees, and the figure the literature usually quotes. field_rates_exact pins all three NES rates separately so the two cannot be conflated again. The broadcast rates falling out as exactly 30,000/1,001 and 60,000/1,001 is itself a check that 315/88 and the geometry are wired correctly.

Data files and the transcription gate

data/nes-levels.toml: the terminated measurement table (lidnariq, 75 ohm, +/-2 IRE), the page's average emphasis attenuation factor 0.816328, and the wave-number facts, from nesdev wiki "NTSC video" revision 23864 (2026-06-01), transcribed from the raw wikitext. The wiki declares no site-wide licence (API rightsinfo empty, no footer notice); recorded in the file header and NOTICE.md.

Closing gate, run 2026-09-01: a second agent independently fetched the same revision and transcribed to the same schema without sight of the first copy. tools/diff-transcriptions.py compared 43 numeric values, 0 disagreements; a deliberate one-digit perturbation of one copy makes the gate exit 1, so the gate can tell. Both transcribers independently made the same single judgement call: the $3x high slot copies the $2x row, per the page's own levels[16] table and its "$20 and $30 are exactly the same" statement. Transcription B is kept as data/nes-levels-transcription-b.toml, labelled evidence, never read by any crate.

data/yuv-matrix.toml: the SMPTE 170M base matrix, reduction factors 0.492111 / 0.877283, the stated inverse coefficients, the 33-degree I/Q rotation and the demodulation constants. Secondary source: the ITU-R BT.1700 download 404'd on 2026-09-01, so this is transcribed from the same wiki revision, which cites BT.1700 pages 4, 16, 17. Marked in its header for confirmation against the primary at M2.

tests/data.rs, 4 tests, holds the transcriptions to themselves:

• Every (volts, printed IRE) pair cross-checks through 1 IRE = 7.14 mV with $1D as the 0 IRE reference; worst row is CBH at 29.69 computed vs 30 printed, inside the tolerance justified by the page's own +/-2 IRE.
• The page's 0.816328 average recomputes from its seven measured attenuation pairs to within 1e-4 (the duplicated $3x row excluded, or it would double-count).
• Every inverse-matrix coefficient re-derives from the base matrix and the reduction factors to within 1e-4 (actual agreement is within 5e-6).

MUTATE=1, the red proof

MUTATE=1 cargo test --workspace --no-fail-fast: a geometry two samples too long per line and a perturbed level voltage. Result: 11 tests red (all eight geometry-residue-and-rate tests; the IRE cross-check, the attenuation recomputation and the matrix re-derivation), 5 green by design and named in the test docstrings: the grid-rate identity (no geometry in it), the closed-form-vs-prefix-sum consistency (it checks the phase formula, not the constants), the two OddShort refusals, and the wave-number range check.

Carried forward

• Open questions from spec section 9, answered by the director 2026-09-01: blargg nes_ntsc pinned at 0.2.2 (last release), composite preset, recorded by hash at M1; CIEDE2000 under D65 for the RGB roundtrip at M2; Rung D (temporal comb) is wanted at M2.
• The yuv-matrix primary-source confirmation is an M2 obligation.
• The spec's full-frame field rate correction above belongs in v0.3.

Pulled at build time from ntsc-crt/docs/m0-report.md; the repository is the one copy.