Fenesphere Console

fentrue-ai-dev

Console access

Paste the console token. This is CONSOLE_TOKEN — a different secret from the field app's, so a leaked field token grants no console access and either can be rotated without disturbing the other. Stored on this device only.

Kris — Morning Brief

Loading…

Machine state

Loading…

API health

No health data yet — Kris records this when she runs.

Agent roster

Grouped by status. IN TESTING agents carry the full character sheet — those are the cards a promotion decision is made from. Gate progress fills in as evals run; promotion stays disabled until every gate is green.

Testing Lab

A photo against a live model, through the same POST /classify the bridge API will use. Served from inference.fentrue.ai.
Drop a photo here
or click to choose a file

Holdout images — focused training

The frozen evaluation set. Click through, mark right or wrong, and the verdicts become Layer 2 data.
Loading…

Knowledge packages

One scenario, taught with several kinds of resource at once. Every resource carries what it is for — a photograph in a package is not automatically a training example.
Loading…

Pre-fill preview

What the technician would see in the Fentrue form. Mark each field correct or wrong — the verdicts and the photo become Layer 2 data.
No run yet.

Consensus assembly

What the user would get, which is not the same as what any one agent said. Avery combines the reads; conflicts on weight-4+ attributes leave the field blank rather than guessing. Rules in EVALUATION_SPEC “Consensus verification”.
2+ agents agree c = 1 − (1−c1)(1−c2), fires at ≥ 0.90
Single agent fires at ≥ 0.85, with a confidence indicator
Conflict, weight 4+ blank — “Unclear from photo, please confirm”
Conflict, weight ≤ 3 higher confidence wins, logged, no user flag
Nova + Mercer + Ada fires if each clears the weight floor (5 → 0.60, 4 → 0.50, ≤3 → 0.40); below floor, drop that agent and re-apply the 2- or 1-agent rule
Verified label above floor and confirmed here → Tier 1; unreviewed → candidate
Per-attribute agreement map, combined confidence and pre-fill decision render here once a photo can be submitted. Pre-fill accuracy will always be shown paired with coverage — blanked fields leave the accuracy denominator, so accuracy alone can be driven to 100% by answering nothing.

Gap signals

Every conflict logged this session, user-facing or not, with its training recommendation. Empty until a session runs.

Not every conflict is a knowledge gap: Mercer reporting casing: none against Nova’s profile read is two agents correctly doing different jobs, and a plan that cannot tell that apart from a real gap sends effort where there is nothing to fix.
Customer data: …
Loading…
Select a project.

Agent instructions

Select an agent.

Diff vs last commit

No local edits — the editor is read-only in this build, so the displayed content is the committed content.

Deploy

Commit & Deploy is not built yet, deliberately. It is gated on two things: roles/editor coming off the runtime service account, and the console token being separate from the field app token. Until both hold, a leaked token would mean write access to sixafterseven/fentrue-ai and a Cloud Run deploy, not just photo upload.

Analytics

Coming in Phase 3. Pipeline analytics (corpus growth, run history, verdict distribution) land first; agent usage and performance need the bridge API to be live before there is anything to chart.