Web-evidence policy closure and v8 launch plan

Resolved all previously pending direction questions using external evidence on model trustworthiness, validation methodology, structure tooling, source API constraints, and security controls; converted decisions into an immediate research execution plan.

Decision Outcome

Pending Questions Closed 16 / 16

Primary Direction v8 + Exports

Risk Control Guarded APIs

Frontend Mode Executive Read-only

Answered Questions (Condensed)

ID	Resolved Direction	Operational Effect
Q1	Zone-A rescue first, then disjoint scaling.	Tightens boundary behavior before growth.
Q2	Corrected cohorts only for training.	Reduces label-noise leakage into model fit.
Q3	Keep interpretable baselines; add GBDT challengers.	Preserves continuity while testing uplift.
Q4	Keep threshold anchor at 0.38.	Comparability preserved; adaptive thresholds remain secondary.
Q5	Start lightweight explainability every pass.	Improves traceability for research and executive review.
Q6	Keep frontend read-only.	Limits operational and security surface area.
Q7	Integrate Mol* next pass.	Upgrades structural interpretation quality.
Q8	Nightly canary for pass35 probes.	Adds freshness without destabilizing ETL.
Q9	Run v8 corrected-only training.	Launches next-model research track.
Q10	Prefer Python `usd-core` path first.	Lower runtime/storage overhead.
Q11	Bind viewer to model outputs plus fallback UI.	Turns structure view into decision support.
Q12	Add computed channels with confidence/provenance gating.	Broader coverage while controlling confidence risk.
Q13	Add export bundles next pass.	Improves collaboration and executive handoff.
Q14	Broaden retrieval before volumetric overlays.	Coverage-first sequencing.
Q15	Unfreeze for targeted structure-feature v8.	Begins measurable model-improvement phase.
Q16	Exports before live orchestration panels.	Safer release path with lower API abuse risk.

Evidence Highlights

Label quality: noisy labels are documented as degrading generalization performance, supporting corrected-only training cohorts.
External validation: model transportability requires external validation and calibration attention before policy drift (e.g., adaptive thresholds).
Transparency: modern health AI guidance emphasizes explainability and clear logic communication.
Structure UX: Mol* is production-proven for PDBe/RCSB usage and third-party integration, supporting next-pass adoption.
Computed model caution: AlphaFold confidence is local and can be misread at domain/global orientation level without additional metrics.
Source operations: OpenAlex now enforces key-based quota economics; Europe PMC permits OA protocols but restricts non-OA bulk automation.
Security: OWASP API Top 10 continues to prioritize authorization/SSRF classes, supporting deferred live orchestration from public UI surfaces.
Toolchain pragmatism: usd-core is available as a lightweight pip path; Blender supports UI-less background rendering for targeted runtime probes.

Research Continuation Plan (Immediate)

Track A: Model

Run v8 corrected-only cohort training with structure-derived feature additions as optional flags.
Benchmark against v7 using fixed threshold anchor plus calibration reports.

Track B: Frontend

Add export bundles (JSON/CSV/manifest) and preserve read-only executive mode.
Integrate Mol* for model-linked structure drilldowns with confidence labeling.

Track C: Data/Runtime

Promote OpenAlex/EuropePMC pulls into a canary nightly schedule with budget/rate guardrails.
Use usd-core as default OpenUSD path; run heavyweight containers only where probe goals require them.

References

Inference note: the selected answers are applied assumptions based on the above evidence and current repository state; they can be revised as new data arrives.

Generated 2026-05-20 08:35 UTC