Research-0086 — Tiny-AI SOTA deep dive: is the lusoris approach state of the art?¶

• Status: Active — strategic input. Every quantitative claim is web-sourced with date-stamped citations; no fabricated PLCC/SROCC numbers. Where a number could not be verified the entry is tagged [UNVERIFIED].
• Author: research-subagent (Opus 4.7, 1 M context)
• Date: 2026-05-08
• Tags: tiny-ai, fr_regressor_v2_ensemble, NR-VQA, vmaf-tune, predict-then-verify
• Workstream parent: tiny-AI strategy ("the real cheese") — not yet ADR-bound, this digest exists to inform the next ADR round
• Pairs with: ADR-0303 (production-flip gate), ADR-0291 (probabilistic head), ADR-0309 (ensemble retrain), ADR-0310 (BVI-DVC ingestion), ADR-0042 (tiny-AI per-PR doc bar)

TL;DR¶

• Where we are clearly SOTA-adjacent (matched, not exceeded): our fr_regressor_v2_ensemble (codec-aware, ensemble-of-5, ~5–20 K params) occupies the same niche as Synamedia/Quortex pVMAF (PLCC 0.985 / SROCC 0.988 sequence-level, shallow MLP, encoder-loop features) and MainConcept VMAF-E (±2 VMAF, 10× faster, neural net, in-encode-loop) — three known industry analogues, all 2024–2025, none open-source under permissive licences. We are competitive; we are not uniquely SOTA on raw accuracy, and we have no published external benchmark that lets us claim a ranked win. Predicted-VMAF surrogates are now a crowded niche.
• Where we are clearly BEHIND: NR-VQA (no-reference). DOVER-Mobile (9.86 M params, PLCC 0.853 KoNViD / 0.867 LSVQ_test, 1.4 s CPU per video) and Q-Align (LMM-based, ICML 2024) are the deployable open-weight baselines on UGC. We have nr_metric_v1 (~19 K params, KoNViD-1k only) and no published external benchmark — at least an order of magnitude smaller but also untested on the canonical leaderboards. KonViD-150k is the right corpus to start; LSVQ (39 K videos, 5.5 M ratings) is the bigger one the field actually trains on.
• Where we are GENUINELY NOVEL (no clear prior art on the exact pattern): (a) the predict-then-verify loop with GOSPEL/RECALIBRATE/FALL_BACK verdicts as a deployment pattern — the closest analogue is two-pass per-shot encoding (Wu et al. 2022) but no one publishes the explicit three-state verdict semantic; (b) the ADR-0303 production-flip gate (mean-PLCC + max-min spread on LOSO ensemble as a deployment gate) — no external paper or vendor blog uses that two-criterion shape; (c) the combination Sigstore-keyless-signed ONNX + ONNX op allowlist + per-model PTQ accuracy budget as a standard — sigstore/model-transparency exists but no one ships a permissively-licensed VQA stack with this whole stack wired into CI; (d) codec-aware FR with a 16-slot ENCODER_VOCAB v3 one-hot — the field generally trains one model per codec rather than conditioning, which is empirically a wash but architecturally more reusable.
• Biggest single threat to our differentiation: MainConcept vScore / VMAF-E (Sept 2025) and Synamedia x264-pVMAF (Nov 2024 OSS, GPL-2.0) ship to the same audience. Once a GPL-2.0 reference exists, our BSD/permissive stack is the only durable moat — pVMAF can't be linked into proprietary pipelines that ours can.
• Most surprising finding (positive): the field has NOT solved community-uploaded VQ data. CHUG (2025, 856 videos, 211 K AMT ratings) is the only open-call UGC-HDR dataset of its kind and it relied on AMT for ratings, not a federated/online learner. The "community contributes encode results, model improves" pattern lusoris is sketching has no published precedent I could find — that direction is genuinely open.
• Most surprising finding (negative): SVT-AV1-PSY (the perceptual fork) was archived 2025-04-20; the maintainer is upstreaming work into mainline SVT-AV1 (juliobbv-p/svt-av1-hdr is the recommended continuation). Our saliency-aware ROI x264 path is more advanced than any open-weight AV1 saliency-ROI path that currently ships, because the AV1 ecosystem just lost its psy-tuned fork.

Topic-by-topic¶

1. Tiny full-reference VMAF prediction¶

What the field is doing (2024–2026):

• Synamedia / Quortex pVMAF — shallow MLP (1–2 hidden layers, neural features include QPs + pre-analysis stats + PSNR), claimed PLCC 0.985 / SROCC 0.988 sequence-level against full VMAF, MAE 1.34, RMSE 2.20; CPU overhead ~0.06 % during FHD medium-preset encode. x264-pVMAF released as GPL-2.0 OSS at https://github.com/quortex/x264-pVMAF (Nov 2024); 35× faster than VMAF, frame-level SROCC 0.991, FHD-only, inference C-code, no training code, no weights detail published.
• MainConcept VMAF-E (vScore suite) — neural net, ±2 VMAF accuracy, up to 10× faster than VMAF, integrated into Codec SDK 16.0 (Sept 2025), in-coding-order so true real-time. Closed source.
• Huawei PyTorch VMAF re-implementation (Sept 2023, arXiv 2310.15578) — full-pipeline reproduction (VIF, ADM, motion + SVR-RBF) with discrepancy ≲ 10⁻² VMAF units, purpose-built for gradient-based optimisation of preprocessing filters. Not a tiny student. Code promised, not yet released as of latest version.
• Ren et al. — per-title CRF DNN (Expert Systems with Applications 2023, doi 10.1016/j.eswa.2023.120469) — DNN regressor that predicts CRF for a target VMAF from per-segment features; targets VP9; reports ~1 VMAF MAE on the test split.
• Constant Target Quality (MainConcept blog 2023) — same family, proprietary.
• Direct VMAF-distillation papers: explicit search for "VMAF distillation" returned no published paper beyond preprocessing- network work (Antsiferova et al., ResearchGate 376439157, 2023, who use a trained VMAF approximation as a proxy teacher for a U-Net preprocessor — the inverse of what we do). I.e. no public paper distills VMAF into a tiny student as a metric in itself. The closest is pVMAF's blog post, which is a marketing surface, not a paper.

Where we stand:

Our fr_regressor_v2_ensemble (5 seeds, 6 canonical libvmaf features + 18-D codec block, mean LOSO PLCC ≥ 0.95 per ADR-0303 gate, ~5–20 K params per seed) is structurally identical to pVMAF: shallow MLP, encoder- adjacent features, full-VMAF teacher. The numerical comparison is roughly:

Gap / opportunity:

1. We are slightly behind on raw correlation versus pVMAF's claimed 0.985 — but pVMAF's 0.985 is in-distribution sequence-level on proprietary data, while ours is LOSO mean (out-of-source-content), a strictly harder gate. A like-for-like comparison would require us to publish a sequence-level in-distribution number; per ADR-0303 we deliberately gate on the harder LOSO number, which is correct for a generalisation claim and underclaims compared to pVMAF.
2. We should publish a head-to-head digest vs x264-pVMAF on the Netflix Public Drop (or BVI-DVC) corpus. x264-pVMAF is GPL-2.0 and we cannot link it into our BSD stack, but we can run it as a side-by-side benchmark. This would be the single highest-leverage external-credibility deliverable for the FR predictor track.
3. Net: our FR predictor is at parity with the published industry SOTA on accuracy and ahead on three procedural dimensions that no competitor publishes — LOSO methodology (ADR-0309), production- flip gate (ADR-0303), and probabilistic head with conformal prediction (ADR-0291). These are decision-quality moats, not number moats.

2. No-reference video quality (NR-VQA)¶

What the field is doing:

• DOVER (ICCV 2023) — two-branch (aesthetic + technical), CNN backbone; DOVER-Mobile uses convnext_v2_femto (inflated), 9.86 M params, GFLOPs 52.3, CPU 1.4 s/video. Reported PLCC: KoNViD-1k 0.853, LIVE-VQC 0.835, LSVQ_test 0.867, LSVQ_1080p 0.802. Repo https://github.com/VQAssessment/DOVER. (DOVER full: 200 + MB, ~280 GFLOPs, CPU 3.6 s/video; PLCC ~0.88 / 0.85 / 0.89 / 0.83 on the same four sets.)
• FAST-VQA / FasterVQA (ECCV 2022 / TPAMI 2023) — fragment sampling via Grid Mini-patch Sampling; 99.5 % FLOPs reduction at 1080p vs the preceding SOTA, fragment-attention transformer. Repo https://github.com/VQAssessment/FAST-VQA-and-FasterVQA. Still cited as a strong tiny-VQA baseline in 2024 challenge tables.
• Q-Align (ICML 2024) — LMM-based, "score from text-defined levels". SOTA on IQA / VQA / IAA simultaneously; multi-billion-param backbone. Not "tiny" by any definition.
• MaxVQA / Maxwell (ACMMM 2023, MM database, 4 543 in-the-wild videos, 13 quality dimensions) — language-prompted, builds on CLIP.
• MVQA (April 2025, arXiv 2504.16003) — Mamba state-space model, Unified Sampling, "comparable to SOTA at 2× speed and 1/5 GPU memory".
• LSVQ — LIVE Large-Scale Social Video Quality, 39 000 UGC videos, 5.5 M ratings, the canonical large NR-VQA training corpus (Hugging Face: teowu/LSVQ-videos). Train split 28 056 videos.
• CHUG (Oct 2025) — open-call UGC-HDR uploads, 856 sources, 211 848 AMT ratings, 700 + raters. Public dataset (licence terms not yet specified in the paper).
• NTIRE / AIS challenges 2024–2025 — DOVER fine-tuned on YT-UGC is one of the SOTA baselines.

Where we stand:

We ship nr_metric_v1 — a tiny MobileNet-style baseline (~19 K params, 224×224 grayscale single-frame, KoNViD-1k only training, dynamic-PTQ INT8 sidecar, BSD-licensed). It has not been benchmarked against DOVER / FAST-VQA / Q-Align on KoNViD / LIVE-VQC / LSVQ. We do not have a published PLCC vs MOS for our NR head.

In our defence:

• We are ~500× smaller than DOVER-Mobile.
• We run through libvmaf's ONNX Runtime surface, so the same artefact serves CPU, CUDA, SYCL, Vulkan and HIP execution providers — runtime is whatever the host has, with no separate model build per backend.
• We use 224×224 grayscale single-frame — DOVER uses fragments + temporal. These are different design points; the right comparison is NR-VQA deployment cost (params + FLOPs) at iso-PLCC.

Gap / opportunity:

1. The KonViD-150k Phase-2 corpus expansion is necessary but probably not sufficient. The field's actual large corpus is LSVQ (39 K videos, 5.5 M ratings), not KonViD-150k. We should also ingest LSVQ; it's CC-BY (per Hugging Face card, teowu/LSVQ-videos) and ~3.7× the ratings of what we're planning.
2. We should run DOVER-Mobile as an open-weight baseline in our benchmark suite. Hugging Face hosts the weights; the architecture is convnext_v2_femto-inflated, ONNX-exportable in principle. This gives us a reproducible PLCC number to beat.
3. Honest framing: our NR head today is a placeholder. The nr_metric_v1 model is a baseline-of-record from 2025; it is not a serious answer to DOVER. The user's framing of NR-VQA as "in flight via KonViD-150k" is correct — but the destination has to be "PLCC ≥ 0.80 on KoNViD-1k and LSVQ_test, at < 1 M params, under 100 ms CPU/frame", not just "trained on KonViD-150k".

3. Codec-aware quality prediction¶

What the field is doing:

• Per-codec is the dominant pattern. Synamedia/Quortex publishes separate proprietary models for H.264/AVC and H.265/HEVC; the open-source x264-pVMAF is only x264 (FHD progressive 4:2:0 medium preset, no --tune). MainConcept VMAF-E does not document codec conditioning. Per-title-CRF DNN (Ren et al. 2023) trains for a single codec at a time.
• Codec embedding / one-hot conditioning as a deliberate architectural choice is not standard in published VQA work. The closest analogue is the AIS 2024 / NTIRE 2024 UGC tracks where a single model is evaluated across codecs without explicit conditioning, treating codec as part of the in-the-wild distribution.
• Content-adaptive Encoder Preset Prediction (Telecommunications Research Centre Vienna 2022) — content-side codec conditioning, reports relative bitrate decrease 17.8 % (x264) and 7.9 % (x265).

Where we stand:

ENCODER_VOCAB v3 (16-slot one-hot, ADR-0078 / ADR-0323) lets one model serve every codec the harness adapter knows about — x264, x265, libaom, SVT-AV1, libvpx-vp9, vvenc, NVENC variants, QSV variants, VideoToolbox variants. This is architecturally cleaner than the "one model per codec" industry default but empirically a wash on accuracy: fr_regressor_v3 reports LOSO mean PLCC ≈ 0.9975 (registry note), which is in the same band as pVMAF's per-codec 0.985 — the codec-conditioning neither helps nor hurts within measurement noise.

Gap / opportunity:

1. The 16-slot vocabulary is future-proofing for codec churn (VVC, AV2, hardware encoders). The cost is one extra hidden-layer slot per codec; the benefit is "every new codec adapter ships with the same model checkpoint". This is a correct architectural call; we should publish it as a positional paper or short tech-note — no industry counterpart is on record claiming the same trade-off.
2. Watch-out: per-codec models can specialise on per-codec quirks (e.g. x265's psy-rd, libaom's tile-rows) in ways a one-hot model can't. We should run per-codec slice metrics on the BVI-DVC evaluation set to confirm the multi-codec model does not hide per-codec underperformance behind an aggregate PLCC.

4. Per-shot adaptive encoding with predicted quality¶

What the field is doing:

• Netflix Per-Title (2018) and Per-Shot (2019) — the founding blog posts; Netflix's own pipeline still runs full VMAF, just on shot-level segments rather than the whole title.
• Wu et al. 2022 (arXiv 2208.10739) — quality-constant per-shot encoding by two-pass learning. RF-parameter prediction, claims 98.88 % accuracy of compressed VMAF being within ±1 of target, 1.55× encoding complexity overhead. The closest published analogue to our predict-then-verify loop. Two-pass: fast preprocessing → feature extraction → DNN-predicted RF → encode.
• Constructing Per-Shot Bitrate Ladders using VIF (arXiv 2408.01932, 2024) — 145-feature VIF set, Extra-Trees regressor, BVT-100 4K corpus, 6 resolutions × 16 CRFs, ~20 % bitrate savings vs fixed ladder, mean BD-VMAF gain 4.3–4.5. No reference encoding required at inference.
• ab-av1, av1an, NETINT capped CRF, MainConcept Constant Target Quality — production tooling, all use full VMAF in a binary search rather than predicting CRF directly.
• rav1e per-frame quality / x265 per-frame VBV / SVT-AV1 adaptive temporal filtering — encoder-internal quality knobs, distinct from external predicted-VMAF tooling.

Where we stand:

vmaf-tune predict + vmaf-tune fast (per-shot path) implements the third published variant of this idea, with two distinguishing features:

• Predict-then-verify loop with explicit verdicts — GOSPEL / RECALIBRATE / FALL_BACK per ADR-0067 / ADR-0276 / predictor_validate.py. Wu et al. 2022 has a two-pass loop but does not publish a verdict semantic; the field does not have a language for "the predictor was confident-and-right vs confident-and-wrong vs uncertain-don't-trust-it".
• Per-codec MLP shipped alongside the whole-fork ensemble — the per-codec predictor_<codec>.onnx files in our model registry are separate from fr_regressor_v2_ensemble. The ensemble is the cross-validation gate for the whole stack; the per-codec MLPs are the runtime fast path. No published competitor has this two-tier shape.

Gap / opportunity:

1. The GOSPEL/RECALIBRATE/FALL_BACK verdict semantic is a publishable abstraction — write it up as a short arXiv preprint or a fork-public blog post. We already have ADR-0276, but the field does not have a named convention for this. First-mover on terminology has compounding citation value.
2. Compare VIFF9 (Wu/Bovik 2024) vs our ShotFeatures — they use 145 features, we use ~16. Ablation: does dropping our 16 to 8 meaningfully change LOSO PLCC, and would 145 close the gap to their 0.762 PLCC on cross-over bitrate? (This is a benchmark we could run; their corpus BVT-100 is publicly available.)
3. The Wu et al. 2022 2-pass loop's 98.88 % @ ±1 VMAF is a stronger target than anything we publish for the per-shot path. We should regenerate this number on our predictor + verify pipeline.

5. Saliency-aware ROI encoding¶

What the field is doing:

• DUTS-TR (2017) — image-level saliency, used by us for saliency_student_v1 (ADR-0270). Not video-temporal.
• ViNet-S / ViNet-A (ICASSP 2025) — DHF1K-trained video saliency, ViNet-S 36 MB / 1000+ FPS / U-Net decoder, ViNet-A 148 MB with spatio-temporal action localisation. Ensemble achieves SOTA on three visual + six audio-visual saliency datasets.
• SalFoM (Springer 2024) — UnMasked Teacher backbone + spatio-temporal locality decoder; SOTA on DHF1K + Hollywood-2 + UCF Sports.
• DHF1K (CVPR 2018, PAMI 2019) — 1 K eye-tracked video sequences, the canonical video saliency benchmark.
• Encoder-integrated saliency:
• x264 + saliency map (msu-video-group/x264_saliency_mod) — fork of x264 that takes external saliency maps as input. Active.
• x265 visual-attention-guided AQ (KTH thesis 2023, integrated into x265, A/B tested on SVT Play). HEVC.
• AV1: no published saliency-ROI integration in mainline SVT-AV1; SVT-AV1-PSY (the perceptual fork) was archived 2025-04-20, with juliobbv-p/svt-av1-hdr the recommended continuation. The AV1 ecosystem just lost its psy-tuned fork.
• Saliency-guided pre-processing (Stanford CS231n 2022) — preprocessing rather than encoder integration.

Where we stand:

saliency_student_v1 (~113 K params, fork-trained on DUTS-TR, ADR-0270) drives the QP-offset map for x264 in our vmaf-tune saliency path. This is architecturally on par with the x265-visual-attention KTH thesis — same shape (saliency model → QP-offset → encoder), just on x264. We are static-frame saliency, not temporal; ViNet-S / SalFoM exist and are bigger but get temporal coherence we lack.

Gap / opportunity:

1. Switch to ViNet-S as the saliency teacher (or run a temporal distillation from ViNet-S into a fork-trained tiny temporal student). DUTS-TR-trained static saliency systematically flickers on real video — every published video-saliency method, including ViNet, beats DUTS-TR-only baselines on temporal-coherence benchmarks. ViNet-S is 36 MB; a distilled fork-trained student can be much smaller. Repo: https://github.com/ViNet-Saliency/vinet_v2.
2. Extend saliency-ROI to x265 / SVT-AV1. The x264 path is solid; x265 has --zones we could drive directly (the KTH thesis already wired this). SVT-AV1 has no public saliency-ROI integration since the PSY fork archived; a fork-public AV1 saliency-ROI path would be a genuine first.
3. The user's claim "our saliency path is competitive" is correct for x264 today and uniquely correct for AV1 if we ship the AV1 integration in 2026.

6. Online / sidecar learning for video quality¶

What the field is doing:

• Continual-learning literature (CORE 2024, SuRe 2025, Manifold Expansion Replay 2023) — replay buffers, EWC, surprise-driven prioritised replay. No published video-quality application found in the search.
• Federated learning for video streaming (Springer Nature 2025 chapter "Federated Learning for Scalable Video Streaming") — focuses on bitrate adaptation / personalised content recommendation, not on training quality predictors federatedly.
• Personalised quality models — there is academic work on personalised quality preference (per-user MOS regression) but no shipped production system found.
• Catastrophic forgetting in tiny models — explicit research area; Schick et al. 2024 (ResearchGate 381882542) shows model size affects forgetting in nuanced ways, smaller models forget less on simple tasks but more on long task chains.

Where we stand:

The local-sidecar bias-correction scaffold (today's commit) is ahead of any published prior art for VQ specifically. The pattern "shipped predictor + sidecar additive-bias term that learns from a local user's encode results" exists nowhere I can find in the VQ literature. The closest analogues are personalised-recommendation systems (Spotify, YouTube) that update user-specific embeddings on the device — but those are recommendation, not quality prediction.

Gap / opportunity:

1. We should document the sidecar pattern in a tech-note before shipping it. ADR-0042 covers per-PR docs; a research digest of the pattern is the right place to claim novelty.
2. Catastrophic forgetting is a real risk if the sidecar bias term is ever fed back into the shipped predictor. Industry prescription: replay buffer of "anchor" examples (NF Public Drop VI-DVC) interleaved with new data. We should design this before the sidecar pattern moves from scaffold to closed loop.
3. Federated VQ training (the user's K>10-contributor vision) is still genuinely open ground — no published precedent. Do not build the infrastructure yet (cost > benefit at K=1), but write the ADR specifying the trigger conditions for when we would.

7. Training corpora — UGC vs cinematic balance¶

What the field is doing:

Where we stand:

For FR we are well-served (NF Public Drop + BVI-DVC). For NR we are critically under-served — KoNViD-1k alone, with KonViD-150k pending. The field's flagship NR corpus is LSVQ, not KonViD-150k.

Gap / opportunity:

1. Ingest LSVQ in parallel with KonViD-150k, not after. LSVQ has ~3.7× more ratings than the proposed KonViD-150k workflow and is the de facto SOTA-comparison corpus. Hugging Face hosts it (teowu/LSVQ-videos).
2. Maxwell (4 543 videos × 13 quality dimensions) is the right corpus if we ever ship an explainable NR-VQA head — i.e. one that says "this video is quality 65; the dominant degradation is motion blur". Worth a feasibility ADR.
3. CHUG (HDR-UGC, 856 videos) is the right corpus for an HDR-NR head; pairs with our existing HDR work in vmaf-tune hdr.

8. Tiny-model deployment — ONNX runtime patterns¶

What the field is doing:

• ONNX Runtime quantization — INT8 (S8S8 with QDQ default), FP16, dynamic range optimisation, supports U8U8/U8S8/S8S8.
• Selective Quantization Tuning (arXiv 2507.12196, July 2025) — per-op, per-device quantization tuning is an active research area.
• Quantization Robustness for Object Detection (May 2025, arXiv 2508.19600) — empirical study of FP32 / FP16 / Dynamic UINT8 / Static INT8 across YOLO models, real-world degradation robustness.
• Industry default for video models: FP16 dominates GPU paths (Tensor-core support); INT8 for CPU/edge. Mixed-precision is the emerging frontier (NVIDIA NVFP4 2026 report).
• ONNX op allowlist as a deployment standard: not industry- standard. Most projects ship ONNX with whatever ops the exporter produces; the allowlist discipline is rare outside safety-critical ML.
• Sigstore signing for ML model artefacts: the sigstore/model-transparency project (v1.1.1, Oct 2025) is the canonical infra; Red Hat published "Model authenticity and transparency with Sigstore" (April 2025). The pattern is emerging, not yet standard. No specific public adopters listed in the model-transparency README.

Where we stand:

Our stack ships all five disciplines simultaneously:

• ONNX op allowlist (per-model; CI gate).
• Per-model PTQ accuracy budget (quant_accuracy_budget_plcc, e.g. 0.01 for learned_filter_v1 and nr_metric_v1).
• Dynamic INT8 sidecar for size-sensitive deployments.
• Sigstore keyless signing with bundled *.sigstore.json per model.
• ONNX opset pinning at 17 (per-model record).

This is ahead of every public open-source VQA project I could find. DOVER ships PyTorch checkpoints, no ONNX, no signing. FAST-VQA same. Q-Align same. The commercial counterparts (pVMAF, VMAF-E) presumably have signing/quant pipelines internally but do not publish them.

Gap / opportunity:

1. We have a publishable practice paper here. Not a research paper — a "how a small VQA project ships ONNX with the full supply-chain stack" tech-note. There is no public counterpart.
2. FP16 vs INT8 study on our models — we currently ship dynamic INT8 for some (learned_filter_v1, nr_metric_v1); an FP16 path would benefit GPU-deployed users (whoever embeds libvmaf in an ONNX-Runtime CUDA / DirectML pipeline). Cost: small ablation.
3. The op allowlist is strict — confirm via CI gate that we have not silently regressed against the published allowlist when we add new model types (transformer / Mamba / attention-heavy NR head).

9. Ensemble-with-gating¶

What the field is doing:

• Deep ensembles — well-established (Lakshminarayanan 2017); median pooling and weighted averaging both reported as performance improvements in NTIRE 2024 short-form UGC-VQA challenge.
• Stochastic Weight Averaging (SWA) and EMA — used by NTIRE 2024 top-team SJTU MMLab for training stabilisation; PLCC + SROCC > 0.9 reported.
• Conformal prediction — distribution-free uncertainty quantification; ACM Computing Surveys 2025 calls it a "data-perspective" framework. No published video-quality assessment paper using conformal prediction in our search. There is general work on CP for image segmentation (MICCAI 2025 paper 3902), CP for NLP (TACL 2024), CP under adversarial attack (VRCP 2025). The video quality field has not adopted it.
• Production gates: NTIRE / AIS challenges report "best-of-validation" weights; no two-criterion (mean + spread) gate published.

Where we stand:

Two genuinely-novel patterns:

1. ADR-0303 production-flip gate — flips the registry's "production model" pointer only when both mean LOSO PLCC ≥ 0.95 and max-min LOSO PLCC spread ≤ 0.005. The two-criterion shape ("ensemble is good and tight") is not in the published literature I could find; competitors gate on a single metric.
2. ADR-0291 probabilistic head with conformal calibration — conformal prediction for VQA is genuinely first-of-its-kind in the public literature. The field uses Bayesian / MC-dropout / deep ensembles for uncertainty, not conformal coverage.

Gap / opportunity:

1. Publish the production-flip gate as a short tech-note. The shape is publishable, the math is trivial, and naming first wins citations.
2. Publish the conformal-prediction-for-VQA result as an arXiv preprint. We have a working implementation; the field has a literal gap. Estimated effort: 2–4 weeks. ROI: high (first arXiv preprint that cites lusoris's ADR system; compounding).
3. Watch-out: the production-flip gate's thresholds (0.95, 0.005) are set by hand. A Bayesian or empirical-Bayes calibration of these thresholds — derived from a held-out gate-failure rate target — would be a stronger positional claim.

10. The community-data-loop¶

What the field is doing:

• CHUG (2025) — open-call UGC-HDR uploads with consent + AMT ratings. The only published open-call UGC dataset I could find in the 2024–2026 search. Not a closed feedback loop — the contributors uploaded, the dataset was built once, AMT rated it.
• YouTube UGC dataset — sampled from CC-licensed uploads, not a feedback-loop community.
• Crowdsourced subjective tests (arXiv 2509.20118, Sept 2025) — comparative methodology study. Standard AMT-style worker recruitment.
• Federated VQ training — no public production system found.
• Personalised-quality models — academic only.
• Anonymisation patterns for video metadata — generic privacy / GDPR literature; no VQ-specific best practice published.

Where we stand:

The user's vision — community-uploaded encode results contributing to training — is genuinely novel ground. The closest published analogue is CHUG, which is a one-shot open-call dataset, not a continuous feedback loop. Federated VQ training does not have production prior art. The user's "the real cheese" instinct is correct: this direction is open, in the strong sense of "no one has published it".

Gap / opportunity:

1. Do not build the infrastructure yet. At K = 1 contributor (the user themselves), the cost of a federation infrastructure is enormously larger than the benefit of the data. Defer until K ≥ 10.
2. Do build the data-format and consent-and-licence ADRs now. The hardest part of a community data loop is not the federation software — it's the rights-transfer agreements, anonymisation, redistribution licence. CHUG handled this with individual "rights transfer agreements"; we should specify ours before accepting first contribution. Trigger condition: we should not accept community data for training without the legal scaffolding in place, even from K = 1.
3. Position-paper the pattern — write a short ADR (or tech-note) describing the pattern at the abstraction level "shipped predictor + community-uploaded encode-result deltas → updated sidecar bias term". This pre-claims the novelty without committing to the federated infrastructure.

High-impact next moves (ranked, for the project's planning round)¶

1. Run an external benchmark vs x264-pVMAF (Quortex OSS) and DOVER-Mobile on the BVI-DVC + Netflix Public Drop test splits. Publish the digest as docs/research/00NN-fr-and-nr-external- benchmarks.md. Effort: 3–5 dev-days. Leverage: this is the single most credibility-defining external comparison the project has not yet done. Touches: tools/vmaf-tune/, ai/scripts/ eval_loso_*.py, possibly a new ai/scripts/external_benchmark_pvmaf.py.
2. Write the ADR-0303 production-flip-gate + ADR-0291 conformal-VQA short tech-note for arXiv, citing the ADRs by number. The math is settled; the field-side novelty-claim is real. Effort: 2–4 weeks. Leverage: first arXiv preprint that names lusoris's ADR system; compounding citation value. Output: docs/papers/ tech-note-production-flip-gate.md (tracked source) + arXiv submission.
3. Ingest LSVQ alongside KonViD-150k (CC-BY licence, Hugging Face hosted, ~3.7× ratings of KonViD-150k). Effort: 2–3 dev-days (download tooling + corpus ingestion). Leverage: makes our NR numbers comparable to DOVER / FAST-VQA / Q-Align. Touches: a new ai/scripts/fetch_lsvq.py + corpus-ingestion ADR.
4. Ship saliency-ROI for AV1 (SVT-AV1 segment map, since SVT-AV1-PSY is archived). Effort: 1–2 weeks. Leverage: be the first fork-public AV1 saliency-ROI integration on the internet. Touches: tools/vmaf-tune/src/vmaftune/codec_adapters/svtav1.py, new ADR.
5. Distil ViNet-S into a fork-trained tiny temporal saliency student, replacing DUTS-TR-trained saliency_student_v1. ViNet-S is 36 MB; a distilled student can be ~150 K params. Effort: 2–3 weeks. Leverage: temporal coherence on real video, the missing axis from our static saliency story. Touches: ai/scripts/train_saliency_student.py, new model-card.
6. Document the predict-then-verify GOSPEL/RECALIBRATE/FALL_BACK verdict semantic as a publishable abstraction. The pattern is already in ADR-0276; promoting it to a named convention has compounding value. Effort: 2–3 dev-days for the write-up.
7. Write the community-data-loop legal-scaffold ADR before accepting any external contribution, even from K = 1. Specify consent format, redistribution licence, anonymisation, rights-transfer agreement. Effort: 1 week. Leverage: cannot safely accept any external contribution without this. Touches: docs/adr/0NNN-community-data-loop-legal-scaffold.md.

Discarded directions (with reason)¶

• Skip federated learning infrastructure for now — at K = 1 contributor (the user) the cost of a federation stack is several orders of magnitude larger than the benefit of the data. Revisit at K ≥ 10. Source: ACM Queue federated-privacy survey 2024 + general cost-of-infra reasoning.
• Skip Q-Align / LMM-based VQA as a deployment target — Q-Align is multi-billion params and depends on a vision-language foundation model. Even Q-Align's evaluation requires a GPU. Outside our "tiny" envelope. Worth tracking as a benchmark teacher (knowledge distillation from Q-Align into our tiny student is a possible research direction at K ≥ 100), not as a deployment artefact.
• Skip Mamba state-space NR-VQA (MVQA, Q-Mamba) for now — the speed-up vs DOVER is real but Mamba ops are not in our ONNX allowlist and are not stable across ONNX Runtime versions. Revisit in 2027 once SSM ops stabilise.
• Skip DOVER-as-teacher distillation until we have an LSVQ- ingested corpus to train against. Distilling DOVER on KoNViD-1k alone would just inherit DOVER's KoNViD-1k overfit.
• Skip x264-pVMAF as a code dependency — GPL-2.0 is incompatible with our BSD-3-Clause-Plus-Patent stack. We can run it as an external benchmark but not link or vendor it.
• Skip SVT-AV1-PSY as an upstream-track target — archived 2025-04-20, no future. Track svt-av1-hdr if the perceptual story matters for AV1.
• Skip Huawei PyTorch-VMAF as a build-time dependency — code not released as of latest report. Re-check in 6 months.

References¶

Numbered citation list. WebSearch / WebFetch dates are for 2026-05-08 unless noted.

1. Synamedia. Real-Time Video Quality Assessment with pVMAF. Blog post. https://www.synamedia.com/blog/real-time-video-quality-assessment-with-pvmaf/ — pVMAF architecture (1–2-hidden-layer MLP), inputs (QPs + pre-analysis + PSNR), accuracy (PLCC 0.985, SROCC 0.988 sequence- level), CPU overhead (~0.06 %). WebFetch 2026-05-08.
2. Synamedia / Quortex. Unlocking Real-Time Video Quality Measurement with x264-pVMAF. Blog post (2024-11-03). https://www.synamedia.com/blog/unlocking-real-time-video-quality-measurement-with-x264-pvmaf/ — open-source x264-pVMAF (35× faster than VMAF, frame-level SROCC 0.991, FHD progressive 4:2:0 medium-preset). WebFetch 2026-05-08.
3. Quortex. x264-pVMAF. GitHub repository. https://github.com/quortex/x264-pVMAF — GPL-2.0; inference-only; no training code; SIMD-optimised C; 3 214 commits at WebFetch. WebFetch 2026-05-08.
4. MainConcept. VMAF-E. Product page. https://www.mainconcept.com/vmaf-e — neural-net VMAF approximator; ±2 VMAF; 10× faster than VMAF; in-coding-order integration. WebSearch 2026-05-08.
5. MainConcept. vScore and VMAF-E (IBC 2025). Press release (Sept 2025). https://www.mainconcept.com/ibc2025-vscore-vmafe.
6. Anastasia Antsiferova et al. Hacking VMAF and VMAF NEG. Semantic Scholar paper id 71c676b4ec1465ed6a52684c1cf5ffea7a717c45. https://www.semanticscholar.org/paper/71c676b4ec1465ed6a52684c1cf5ffea7a717c45. (Cited only for context; the search hit, not a primary distillation result.)
7. Cloud BU, Huawei Technologies. VMAF Re-implementation on PyTorch: Some Experimental Results (Sept 2023, latest revision Dec 2023). arXiv 2310.15578. https://arxiv.org/html/2310.15578v3 — full pipeline reproduction; ≲ 10⁻² VMAF unit discrepancy; gradient-based; code release deferred for security review. WebFetch 2026-05-08.
8. Wu, Haoning et al. Exploring Video Quality Assessment on User Generated Contents from Aesthetic and Technical Perspectives (DOVER, ICCV 2023). arXiv 2211.04894. https://arxiv.org/abs/2211.04894. Repo https://github.com/VQAssessment/DOVER. DOVER-Mobile architecture convnext_v2_femto-inflated, 9.86 M params, GFLOPs 52.3, PLCC KoNViD-1k 0.853 / LSVQ_test 0.867. WebSearch 2026-05-08.
9. Wu, Haoning et al. FAST-VQA: Efficient End-to-End Video Quality Assessment with Fragment Sampling (ECCV 2022, TPAMI 2023). arXiv 2207.02595. https://arxiv.org/abs/2207.02595. Repo https://github.com/VQAssessment/FAST-VQA-and-FasterVQA. WebSearch 2026-05-08.
10. Wu, Haoning et al. Q-Align: Teaching LMMs for Visual Scoring via Discrete Text-Defined Levels (ICML 2024). Q-Future GitHub organisation https://github.com/Q-Future. WebSearch 2026-05-08.
11. Wu, Haoning et al. Towards Explainable In-the-Wild Video Quality Assessment: A Database and a Language-Prompted Approach (Maxwell database, ACMMM 2023). arXiv 2305.12726. https://arxiv.org/abs/2305.12726. WebSearch 2026-05-08.
12. Ying et al. Patch-VQA: 'Patching Up' the Video Quality Problem (LSVQ database). Repo https://github.com/baidut/PatchVQ. Hugging Face dataset card teowu/LSVQ-videos. WebSearch 2026-05-08.
13. Saini, Shreshth et al. CHUG: Crowdsourced User-Generated HDR Video Quality Dataset (Oct 2025). arXiv 2510.09879. https://arxiv.org/html/2510.09879. 856 sources, 211 848 AMT ratings, 700+ raters, open-call uploads with rights-transfer agreements. WebFetch 2026-05-08.
14. Wu, Hao et al. MVQA: Mamba with Unified Sampling for Efficient Video Quality Assessment (April 2025). arXiv 2504.16003. https://arxiv.org/abs/2504.16003. WebSearch 2026-05-08.
15. Y. Sang et al. Constructing Per-Shot Bitrate Ladders using Visual Information Fidelity (Aug 2024). arXiv 2408.01932. https://arxiv.org/html/2408.01932. Extra-Trees regressor, BVT-100 4K corpus, 6 res × 16 CRFs, ~20 % bitrate savings. WebFetch 2026-05-08.
16. Wu, Tianyu et al. Quality-Constant Per-Shot Encoding by Two-Pass Learning-Based Rate Factor Prediction (2022). arXiv 2208.10739. https://arxiv.org/pdf/2208.10739. WebSearch 2026-05-08.
17. Ren et al. Per-title and per-segment CRF estimation using DNNs for quality-based video coding (Expert Systems with Applications 2023). doi 10.1016/j.eswa.2023.120469. https://www.sciencedirect.com/science/article/pii/S0957417423007911. WebFetch 2026-05-08.
18. KTH Royal Institute of Technology. Visual Attention Guided Adaptive Quantization for x265 (thesis, 2023). https://kth.diva-portal.org/smash/get/diva2:1788172/FULLTEXT01.pdf. WebSearch 2026-05-08.
19. MSU Video Group. x264_saliency_mod (x264 with custom-saliency- map input). https://github.com/msu-video-group/x264_saliency_mod. WebSearch 2026-05-08.
20. ViNet-Saliency. vinet_v2 (ICASSP 2025). GitHub https://github.com/ViNet-Saliency/vinet_v2. ViNet-S 36 MB, 1000+ FPS, U-Net decoder; ViNet-A 148 MB. WebSearch 2026-05-08.
21. Wang, Wenguan et al. Revisiting Video Saliency: A Large-scale Benchmark and a New Model (DHF1K, CVPR 2018, PAMI 2019). https://github.com/wenguanwang/DHF1K. WebSearch 2026-05-08.
22. SalFoM. Dynamic Saliency Prediction with Video Foundation Models (Springer 2024). https://link.springer.com/chapter/10.1007/978-3-031-78312-8_3. WebSearch 2026-05-08.
23. SVT-AV1-PSY (archived 2025-04-20). README final state. https://github.com/psy-ex/svt-av1-psy/blob/master/README_old.md. Recommended successor: juliobbv-p/svt-av1-hdr. WebFetch 2026-05-08.
24. ONNX Runtime. Quantize ONNX Models (current docs). https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html. WebSearch 2026-05-08.
25. Selective Quantization Tuning for ONNX Models (July 2025). arXiv 2507.12196. https://arxiv.org/html/2507.12196v1. WebSearch 2026-05-08.
26. Sigstore. model-transparency. GitHub https://github.com/sigstore/model-transparency. v1.1.1 (Oct 2025), 232 stars. WebFetch 2026-05-08.
27. Red Hat Emerging Technologies. Model authenticity and transparency with Sigstore (April 2025). https://next.redhat.com/2025/04/10/model-authenticity-and-transparency-with-sigstore/. WebSearch 2026-05-08.
28. NTIRE 2024 Challenge on Short-form UGC VQA: Methods and Results. arXiv 2404.11313. https://arxiv.org/html/2404.11313v1. SWA + EMA strategies, top-team SJTU MMLab PLCC + SROCC > 0.9. WebSearch 2026-05-08.
29. ACM Computing Surveys. Conformal Prediction: A Data Perspective (2025). https://dl.acm.org/doi/10.1145/3736575. WebSearch 2026-05-08.
30. ResearchGate publication 381882542. The impact of model size on catastrophic forgetting in Online Continual Learning. WebSearch 2026-05-08.
31. CORE: Mitigating Catastrophic Forgetting in Continual Learning through Cognitive Replay (2024). arXiv 2402.01348. https://arxiv.org/html/2402.01348v1. WebSearch 2026-05-08.
32. Springer Nature. Federated Learning for Scalable Video Streaming (2025). https://link.springer.com/chapter/10.1007/978-3-031-84651-9_3. WebSearch 2026-05-08.
33. Xinyi Wang et al. BVI-UGC: A Video Quality Database for User-Generated Content Transcoding (Aug 2024). arXiv 2408.07171. https://arxiv.org/html/2408.07171v1. 60 ref × 18 dist, 3 500 raters. WebSearch 2026-05-08.
34. Wang Yutian et al. Comparative Study of Subjective Video Quality Assessment Test Methods in Crowdsourcing (Sept 2025). arXiv 2509.20118. https://arxiv.org/abs/2509.20118. WebSearch 2026-05-08.
35. NVIDIA Technical Blog. Calculating Video Quality Using NVIDIA GPUs and VMAF-CUDA. https://developer.nvidia.com/blog/calculating-video-quality-using-nvidia-gpus-and-vmaf-cuda/. WebSearch 2026-05-08.
36. ScienceDirect, Memory-VQA: Video quality assessment of UGC based on human memory system (2025). doi via journal. WebSearch 2026-05-08.
37. Lakshminarayanan, Pritzel & Blundell. Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles (2017). Cited as the deep-ensemble baseline reference; not separately WebFetch'd.

Process notes¶

• This digest is research-only; no code or tests are changed by the PR that lands it.
• Every quantitative claim about an external system (PLCC, SROCC, param count, GFLOPs, dataset size, model architecture) is sourced from a WebSearch result or a WebFetch'd primary source on 2026-05-08. Where a number could not be confirmed, the entry is tagged [UNVERIFIED].
• The "TL;DR" three "novel" claims (predict-then-verify verdict semantic, ADR-0303 production-flip gate, conformal-VQA) are negative-search claims — i.e. I searched and found no published prior art. A future reader should treat this as "no obvious public precedent at write-time" rather than "provably first".
• Pairs with the existing planning ADRs ADR-0303 / ADR-0291 / ADR-0309 / ADR-0310; does not itself create an ADR. Concrete next-move bullets in §"High-impact next moves" are candidates for separate ADRs.