Current Plan

June 3, 2026 ยท View on GitHub

Last updated: 2026-05-27.

This file is the active roadmap only. Historical phase detail lives in dated experiment notes and research/experiments/registry.md. Git history preserves removed historical source imports and superseded strategy notes.

Current Position

  • C-CEILING is earned as the accounting frame: end-to-end gains are bounded by dense component share times component reduction.
  • C-PERSIST is earned for same-video follow-up latency inside the tested envelope. Warm follow-up multipliers remain the headline denominator, while setup-inclusive tables now expose the serving economics by session length.
  • C-VISION has bounded measured sparse-vision evidence. Do not describe it as a broad sparse backend or sparse LM prefill result.
  • RLT/Gemma cost accounting now reinforces C-CEILING: the prefill+vision stage model fits 19 Gemma cells at R^2=0.97097 and 1.72% mean absolute relative error. Treat admission-only rows as cost-accounting controls, not as a query-aware implementation result.
  • Candidate C-STREAM has a checked mixed/boundary bundle. It is paper-facing as a candidate scale-out regime, not as a fourth headline: default cache reuse is unsafe, topology-aware correctness can be restored without speed, prefix snapshots are promising but small-(N), and low-FPS dense remains a serious baseline.
  • 26B follow-up reuse remains diagnostic until a topology-safe path preserves correctness without surrendering the speed path.
  • OneVision x VLMaxxing refresh planning is currently a parity/boundary result, not a codec-over-pixel win. The pooled N=57 Qwen/VideoMME-short rows are dirty-tree/advisory, in-sample, and matched to the pixel proxy's reuse budget; simple codec sources are only +2/57 over pixel there, and the disjoint per-item holdout ties pixel exactly. Positive refresh language now requires a clean frozen-threshold transfer run against the pixel proxy.
  • OneVision x VLMaxxing sparse vision pruning is bounded Qwen evidence, not a broad speedup claim. At VideoMME short / Qwen2.5-VL-7B-4bit / 8 frames, codec_novel_coded at kr=0.7/layer=2 is the best tested sparse arm by point estimate over magnitude_norm, but paired tests remain inconclusive and that cell lacks the matched random-keep control. The follow-up sweep adds two boundaries: Qwen random beats magnitude on 4/4 seeds at kr=0.5/layer=2 and codec does not beat random where that control exists, while TOMATO motion stays near chance and is not rescued by codec scoring.
  • OneVision x VLMaxxing sidecars are the clean systems result: live PyAV codec-score extraction is replaced by precomputed score sidecars with zero sparse-choice drift in the landed Qwen/Gemma gates. This is an extraction-path speedup, not a 10410^{4}x model-pipeline speedup.

Active Gates Before Paper/OSS Freeze

  1. Finish incoming experiment bundles.

    • local phase-2 chain: finish remaining follow-on gates and update the registry/status docs from the artifacts
    • scale-out bundle: import only checked artifact bundles, not sibling markdown, personal handoff prompts, or screenshots
    • close the RLT/VLMaxxing documentation around the 2026-05-20 cost-accounting result, then keep M5 as scale confirmation only
    • natural-dialogue C-PERSIST and one adjacent-method comparison are the highest-value main-track science gaps after the current integration pass
    • defer query-aware implementation until a separate branch can inherit the stage-cost ledger and run fresh held-out fixed/random/admission controls
    • OneVision follow-up: sidecar gates have landed; Gemma N=10 smoke clears the cross-family wiring gate; TOMATO is a low-headroom boundary result; and pooled H.264 refresh evidence is advisory until a clean frozen-threshold transfer run. Use M5 only for preregistered confirmation cells. The Qwen kr=0.7 random-control protocol is preregistered and must run through scripts/run_ov6_m5_qwen_kr070_random_control.sh; the Qwen kr=0.7 parity/timing launcher must either point at that committed four-seed random-control preregistration or explicitly close that clean-control window with a committed closure note. Do not claim net codec model-pipeline speedup; only claim the measured extraction-path sidecar speedup.

  2. Freeze artifact provenance.

    • every paper table/figure cell needs a source artifact path or a visible pending label
    • generated build metadata must be regenerated from the clean release tag
    • no paper automation should read a sibling checkout by default

  3. Keep claim language bounded.

    • C-PERSIST: distinguish median warm follow-up speedups from mean setup-inclusive session economics
    • C-VISION: bounded measured sparse-vision envelope
    • C-STREAM: candidate streaming state reuse until a topology-safe fast path and matched native-streaming baselines land
    • imported targets: reproduction targets only, not local evidence

  4. Release hygiene.

    • default repo should foreground code, tests, checked artifacts, manuscript tooling, validators, schemas, and concise reproduction docs
    • historical source imports, stale review packets, and one-off legacy harnesses should stay out of the release tree because git history is sufficient

Source Of Truth