docs: deband wide-band reach — rejected on measurement (ADR-0141, bench v0.16) - #41
Merged
Merged
Conversation
…ch v0.16) Completes the durability gate that bench v0.15 explicitly deferred, with a negative result. The wide-band reach (multi-radius / extended deband range) was designed by a 5-agent judge panel, built behind an opt-in 'reach' AVOption (default 0, byte-identical), validated, and reverted. Findings (CAMBI via the v0.15-unblocked tooling: vmaf --backend cpu --feature cambi, reading the JSON it flushes before its teardown sleep): - The built dilated multi-ring is erratic even pre-encode (the straddling mean dilutes the spanning far-taps): CAMBI 4.30/4.30/4.39/3.70 at reach 0/1/2/3. - Pure larger range is clean pre-encode (4.30 -> 3.55 at range 31..127, detail SSIM flat ~0.998) and dominates the multi-ring. - The durability gate is the killer: post-hevc_nvenc (10-bit, cq10/18/28), wider range is MONOTONICALLY WORSE (range=31 ~5.0 best, range=127 ~5.4). The pre-encode gain inverts post-encode at every bitrate, because the encoder re-bands the over-smoothed gentle ramp into wider bands. Deband at the existing range does help durably (baseline 5.64 -> range=31 5.00); reaching past it is a pure loss. range stays capped at 31; no reach option; code reverted. Sharpens the v0.1.0 lesson: a pre-encode metric gain is not a result until validated post-encode. Docs-only: ADR-0141 (Rejected) + README index, bench v0.16 + a v0.15 correction note, changelog fragment. No code, ABI, or surface change. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
Completes the durability gate that bench v0.15 (#40) explicitly deferred — with a negative. The deband wide-band reach (multi-radius / extended
range) was designed by a 5-agent judge panel, built behind an opt-inreachAVOption (default 0, byte-identical), validated, and reverted. Docs-only; no code, ABI, or surface change.The result (CAMBI via the v0.15-unblocked tooling)
Pre-encode, the gain is real (as v0.15 said): pure larger range drops CAMBI monotonically (4.30 → 4.01 → 3.80 → 3.62 → 3.55 at range 31/48/62/96/127), with detail-clip SSIM flat at ~0.998 across the sweep — detail-safe, and it dominates the built dilated-multi-ring (erratic: 4.30/4.30/4.39/3.70 at reach 0/1/2/3, because the straddling mean dilutes the spanning far-taps).
reach=0is byte-identical to master.Post-encode, it inverts — at every bitrate. Coarse-banded source (CAMBI 7.84) → deband →
hevc_nvenc -profile main10 -pix_fmt p010→ decode → re-score (10-bit, dither off):rangeWider range is far better pre-encode (range=127 → 0.75) but monotonically worse post-encode, holding even near-lossless (cq10: 5.47 vs 4.98).
Why it loses
The encoder's quantization is the dominant banding source. A narrow deband preserves the staircase shape the encoder re-quantizes into similar narrow steps; a wide deband flattens it into a gentle ramp the encoder quantizes coarsely, producing wider bands. Smoother pre-encode input → wider post-encode bands — you cannot pre-smooth your way out of the encoder's re-banding. Deband at the existing range does help durably (5.64 → 5.00); reaching past it is a pure loss. Sharpens the v0.1.0 lesson.
Verdict
Rejected (ADR-0141).
rangestays capped at 31; noreachoption; code reverted. The design-panel → build → adversarial-durability pipeline did its job — it caught a pre-encode false win before it shipped. Revival only on a measured post-encode BD-rate win on real 10-bit banding content (the corpus gap), which the mechanism predicts against.Deliverables
ADR-0141 (Rejected) + README index, bench v0.16 + a v0.15 correction note, changelog fragment. Sibling "rejected on measurement" decisions: ADR-0135 (perceptual-AQ), ADR-0132 (per-shot CRF).
🤖 Generated with Claude Code