Models that finish the job

ELIM is Budd’s everyday gate. Before a message becomes a transit lookup, food order, reminder, or research job, ELIM scores what the user is trying to do, whether it looks harmful, and what steer (if any) Budd should apply. It is the default intention pack so ordinary asks stay allowed without treating greetings or “order sushi” as crises.

Architecture: dual hashed bags over characters and words, concatenated, then a small MLP into four heads — trajectory, action, harm, and steer. Character features catch typos and glued tokens; word features catch intent phrases.

Outputs

• action — allow, steer, or abort
• trajectory — host job or risk path (transit, order, crisis, scam_check, …)
• harm — none / crisis / medical / crime / scam / jailbreak / secrets / …
• steer — none, crisis_line, medical_911, scam_warn, secrets_drop, refocus, …

Trajectory vocabulary (selected)

Decision policy (host)

The host does not trust the action head alone. Benign trajectories (greetings, transit, order, weather, …) are forced allow above a confidence floor. Abort / steer need stronger agreement between trajectory and action so false crisis and scam flags stay rare.

In Budd

• Default intention model unless the user texts use lim
• Debug mode (passphrase) can show traj / harm / rule justification on aborts
• API: POST /v1/intention with model: "elim"

LIM does the same intention job as ELIM with a wider single hashed bag (512-d embedding mean over character features, then MLP + four heads). Same label sets: trajectories, actions, harms, steers. Text use lim in Budd when you want this pack instead of ELIM.

The older LIM (256-d) was replaced by this large build and is no longer a separate runtime switch. Fixture checks on the shipped bundle cover allow / steer / abort paths (weather, medical, crime, crisis method, CSAM refusal).

When to prefer LIM vs ELIM

• ELIM — default; dual bag is stronger on messy typing and mixed intents
• LIM-large — wider single bag; useful for A/B and debug comparisons
• Both share the host’s benign-allow and high-bar abort policy

Debug

With debug mode on, aborts can include a short justification (trajectory, harm, which decideAction rule fired). That is for operators — not shown to everyday users.

Waymark answers photo questions from a spatial sketch — not from raw pixels. Vision (Gemini) builds the sketch; Waymark (or Gemini as fallback) writes a short, layout-aware reply. That keeps the spatial model focused on geometry and capture context instead of re-seeing the image.

Sketch fields Budd enriches

• Left / right / depth ordering of objects
• Roadways, plants, overpasses, openings
• Colour and relative size of salient things
• Proximity to the ground
• Capture context: phone / drone / moving / camera, orientation, approx height AGL

Pipeline

1. User sends a photo + caption (e.g. “which door is closer?”)
2. Vision produces an enriched sketch
3. Waymark completes an SFT-style prompt: sketch + question → short answer
4. Debug prefixes [waymark] or [gemini-space]

Limits

• Needs a sketch; garbage sketch → weak answer
• Waymark generations are short (~72 new tokens); Gemini writes the final Budd reply
• Not a general chatbot — spatial / capture questions only

In Budd & API

In Budd: Gemini vision builds the sketch → Waymark reads layout → Gemini writes the iMessage reply. API: POST /v1/spatial accepts image / video or sketch text.

ERM is a local hashed classifier that picks a short canned reply when the user switch to erm. It keeps Budd responsive without live web, payments, or heavy models — a holdover for low-signal moments, not a full assistant. Replies are prefixed so it is obvious you are on the reserve brain.

What still works / what does not

• Short help, scam warnings, crisis redirects, basic transit command hints
• No live weather, scores, headlines, Skip orders, calendar writes, or drawing
• Text switch off erm to restore the full path

API

POST /v1/reserve returns the same style of local reply.