工赋Young · New materials optimization
Evidence & Experiment Console
Current verified snapshot
System at a glance
Data interpretation
Candidate outcome space
Position shows the two predicted objectives. Proximity between predictions and returns is not prospective model validation.
Evidence readiness
Gate status
Engineering diagnostics
Baseline model checks
R² and RMSE use separate scales and units. They are pipeline sanity checks—not generalization or publication evidence.
Closed-loop composition
Evidence mix
Only accepted rows enter the next-cycle dataset. Failed and censored rows remain preserved as immutable evidence but are not relabelled as successful targets.
Bounded optimization output
Candidate experiments
| Rank | Score | Resin ratio | Cure temp | Cure time | Predicted Tg | Predicted flexural |
|---|
Immutable returned evidence
Accepted laboratory-result rows
| Rank | Evidence class | Outcome | Tg | Flexural strength |
|---|
Deterministic control plane
Workflow, approvals & safety
Dashboard reads query-only. JSON and JSONL remain inspectable mirrors.
Domain scientist, laboratory safety and data owner identities must remain distinct.
Final approval, release, event and checkpoint commit in one transaction.
Simulator jobs can only return synthetic evidence and cannot be relabelled measured.
Experiment approval never constitutes publication authorization.
The web agent cannot approve, release, dispatch, ingest or publish.
Append-only sequence
Lifecycle timeline
Transparent provenance
Sources, tools & licenses
Optional interpretation support
Advisory evidence agent
Local operator preferences
Agent settings
There are no controls for approval, release, laboratory dispatch, result ingestion, evidence classification or publication. Those remain deterministic CLI/service responsibilities.