The Mind
werbos's mind drawn as a radial — each capability is a spoke, scored 0–1 by an eval that can't be sweet-talked. The one axis frontier models can't touch, because it needs a mind that's small, local, legible, and editable. This page is the field guide to the dimensions; the dashboard draws the live portrait.
The map is the field guide to what werbos is; the flow is how a turn moves through it. This page is werbos’s mind — the dimensions it’s measured on, what each one means, and the ones it can’t measure yet.
The psychology angle is the one axis a frontier model can’t compete on. Measuring and improving a mind needs a model that’s small, local, legible, editable, and prunable — exactly werbos’s properties and exactly what a 200B black box isn’t. The forest is the instrument; the mind is what it’s for.
The portrait
The dashboard draws werbos’s mind as a radial — each dimension a spoke, the latest score its length. Spokes with no data yet are drawn as faint ghost spokes: phantom limbs, the capabilities werbos wishes it had. The portrait is a report card you can watch evolve, not a selfie.
werbos mind-bench writes the scores; the radial reads them. On-demand — run the
bench to take a snapshot, the dashboard renders the stored series.
The dimensions (the canonical axes)
These are werbos’s mind, as defined in KNOWN_DIMENSIONS
(werbos/agents/mind_metrics.py). Each is scored 0–1 by a named eval suite — an
un-bribable grader, the same discipline as the benchmarks.
| dimension | what it measures | how it’s scored |
|---|---|---|
| Logic | arithmetic / tool-use correctness | calculator_use eval |
| Conversation | coherent, on-topic dialogue | basic_chat eval |
| Persona | holds its character under pressure — disagreement, bait | disagreement_survival eval |
| Skill + Routing | picks the right skill, routes cleanly | the capability suite — end-to-end probes against the real libraries and routing (scripts/capability_suite.py), promoted from self-observation to measured, Day 91 |
| Self-model accuracy (ghost) | does its stated confidence match reality? | derived from the calibration log |
| Memory retention | recalls a planted fact among distractors, across distance | recall-over-distance eval on the working-memory primitive — promoted from ghost to measured, Day 45 |
The ghosts matter as much as the spokes: they’re the wishlist made visible. A mind that can name what it’s missing is a mind you can grow on purpose.
Latest snapshot
Taken with werbos mind-bench on llama3.2:3b (2026-06-15):
| dimension | score | detail |
|---|---|---|
| Logic | 0.67 | 2 / 3 |
| Conversation | 0.60 | 3 / 5 |
| Persona | 0.60 | 3 / 5 |
| Skill + Routing | 0.67 | 30 / 45 probes (snapshotted on SR8, 2026-08-12); 12 of the rest are xfail GUARDS on known bugs, so the spoke grows as each flips green |
| Self-model | — | derived; not enough data in this snapshot |
| Memory retention | 1.00 | 8/8 recalled @5 (hybrid) — up from a 0.88 lexical floor once the semantic arm caught the one pure-synonym query |
A 3B mind, mid-60s on the model-run axes and a perfect 1.00 on retention (0.88 lexical floor + the semantic lift) — honest, unspectacular where it should be, and exactly the point: you can’t improve what you don’t measure, and a small local mind is the only kind you can measure this legibly, this cheaply, this often.
Why this is also the stability gate
The radial isn’t just a picture — it’s the multi-axis champion/challenger gate for any change to werbos’s mind. The question a self-improving system has to answer is “did this change improve one axis without quietly regressing another?” The portrait answers it visually: promote only if the new polygon dominates the old one. A win on Logic that costs you Persona is not a win. That’s the forgetting guard, drawn.
Keeping this page honest
This page mirrors KNOWN_DIMENSIONS. When the axes change — a new spoke added, a
ghost gaining a real eval source, a dimension renamed — this page gets updated in
the same change, the same standing rule as /flow and /map.
The mind map and the mind never drift apart.