ADR-0481: ADM p-norm Parameter Hardcoded at 3.0 — Deferral Decision¶

Status: Accepted
Date: 2026-05-16
Deciders: lusoris, Claude (Anthropic)
Tags: adm, predict, ai, testing

Context¶

The ADM (Detail Loss Measure) extractor computes per-scale accumulation scores with an L-p norm root step using a fixed exponent of 1/3.0:

float num_scale_h = powf(f_accum_h, 1.0f / 3.0f) + ...;

AdmState carries an adm_p_norm field (defaulting to 3.0f) and an adm_p_norm option declaration, but this field is not wired into adm_cm_partial_* or the accumulation loop. A //TODO comment in integer_adm.c:2492 (pre-dating the fork) noted the parameterised form for future integration.

The Python training harness (quality_runner.py) also hardcodes adm_p_norm=3.0f during training and scoring. Changing this value produces ADM scores outside the training distribution of the shipped vmaf_v0.6.1 family of models, violating the Netflix-golden gate.

Decision¶

The adm_p_norm integration is formally deferred until a retrained model that varies p-norm over the training corpus is produced and registered. The //TODO comment is replaced with a DEFERRED block citing this ADR. The adm_p_norm option on AdmState remains in the struct (it is part of the public option schema and removing it would be an ABI break) but continues to have no effect on computed scores.

Alternatives considered¶

Option	Pros	Cons	Why not chosen
Wire `adm_p_norm` now	Enables p-norm tuning research	Breaks Netflix-golden gate; requires model retrain; no corpus currently supports it	Correctness risk outweighs benefit
Remove `adm_p_norm` option entirely	Reduces dead code	ABI break (field is in `VmafOption`); downstream code may set the option via dict and rely on the error-only path	Not justified yet
Leave as-is with //TODO	Avoids the decision	`//TODO` without a rationale is ambiguous; audits flag it as unresolved open work	Does not close the audit finding

Consequences¶

Positive: the audit finding from .workingdir/audit-todo-fixme-2026-05-16.md item #6 is closed; the deferred state is formally documented with a clear re-open trigger.
Negative: adm_p_norm remains dead code until a retrained model ships.
Neutral / follow-ups: if a future tiny-AI or retrained ADM model varies p-norm, the wiring in adm_cm_partial_* should be revisited in the same PR as the model registration, citing this ADR as the superseded deferral.

References¶

core/src/feature/integer_adm.c — DEFERRED comment at adm_cm_partial_* p-norm step
TODO/FIXME audit: .workingdir/audit-todo-fixme-2026-05-16.md item #6
python/vmaf/core/quality_runner.py — adm_p_norm=3.0f hardcoded in Python harness