Volunteer Speed Profile
0.9794Latency: 158.61 ms
Best for high-volume field reports where responders need fast safe routing.
Run: EDG-307 r0 / 20260427T200056Z
Metrics source
Current validated frontier.
A judge-friendly summary of what the numbers mean, which runs support the public claims, and why the public story uses a trusted full-50 frontier instead of the raw maximum row in a research ledger.
The canonical repository file is docs/CURRENT_FRONTIER.md.
What these numbers mean
Two profiles, two response needs.
Volunteer Speed Profile is for high-volume field sorting; Critical Accuracy Profile is for slower, more careful review. Both are measured on the same 50-sample MEDIC/QCRI gold set and stay far below the 4-second response budget, so the tradeoff is about operational fit, not waiting minutes for a model.
Decision
We present a Pareto frontier, not a single magic number.
Speed matters for volunteer sorting. Accuracy matters for critical incident review. Judges can inspect both profiles in Optimization Mode while the Web UI keeps only two top-level experiences.
Critical Accuracy Profile
0.9818Latency: 237.97 ms
Best for high-stakes review where a small latency cost is acceptable.
Run: EDG-480 r2 / 20260515T093558Z
4,000 ms ceiling
< 0.3sThe current profiles preserve most of the time budget for human judgment and communications.
Metric discipline
Why not just show the raw maximum?
results.tsv is an experiment ledger, not a leaderboard. It includes diagnostics, guarded CPU runs, low-VRAM checks, and historical artifact rows. Public claims therefore use trustworthy comparable full-50 rows from docs/CURRENT_FRONTIER.md, with exact run IDs preserved for auditability.
Readable source details
Open the evidence without leaving the demo.
For safety and judge usability, this page exposes curated, versioned excerpts instead of serving every raw research log or repository file from the public site.
Canonical file docs/CURRENT_FRONTIER.md View excerpt
The repository source of truth says to present two validated profiles rather than one headline number:
- Volunteer Speed Profile: 0.9794 weighted F1 at 158.61 ms.
- Critical Accuracy Profile: 0.9818 weighted F1 at 237.97 ms.
- Decision: EDG-480 is the current Pareto/frontier candidate; EDG-479 ablations should not replace it alone.
Experiment ledger results.tsv View rows
Judge-facing rows from the run ledger:
| Run | Weighted F1 | Latency | Decision |
|---|---|---|---|
| EDG-307 r0 | 0.9794 | 158.61 ms | Keep for speed |
| EDG-478 r1 | 0.9798 | 286.26 ms | Superseded by EDG-480 |
| EDG-480 r2 | 0.9818 | 237.97 ms | Accuracy candidate |
| EDG-479 r1 | 0.9594 | 260.33 ms | Do not use alone |
| EDG-479 r2 | 0.9595 | 214.79 ms | Do not use alone |
Demo data site/data.json View data
The interactive cards are powered by a small static JSON file. It is safe to expose because it contains only public metrics and curated demo examples.
{
"frontier": [
{ "profile": "Volunteer Speed Profile", "f1": 0.9794, "latencyMs": 158.61 },
{ "profile": "Critical Accuracy Profile", "f1": 0.9818, "latencyMs": 237.97 }
]
}