stuntd 0.1.2: local heads for multi-field decisions, and why one weak field decides how often you skip the model

Releases·Tool Release

stuntd: a local Jev-compatible server on Laya that learns from your own traffic (no API key needed)

A developer released stuntd, an Apache-2.0 local proxy for typed LLM decisions (choice, yes/no, score) that records traffic from a provider like the hosted Jev API, trains a small head on the frozen open-weights Laya encoder, and serves answers locally with fallback to the provider below a confidence threshold. It matters to practitioners because it cuts ~380 ms paid round trips down to ~22 ms local inference on closed-decision tasks, with self-promotion and demotion based on agreement rates.

A developer released stuntd, an Apache-2.0 local proxy for typed LLM decisions (choice, yes/no, score) that records traffic from a provider like the hosted Jev API, trains a small head on the frozen open-weights Laya encoder, and serves answers locally with fallback to the provider below a confidence threshold. It matters to practitioners because it cuts ~380 ms paid round trips down to ~22 ms local inference on closed-decision tasks, with self-promotion and demotion based on agreement rates.

Products·launched

Stuntd: LLM proxy that distills repeated small LLM decisions into small local models

Stuntd is a proxy that watches an app's traffic to an LLM, learns a small classifier head per repeated closed decision (routing, urgency, safety checks), and serves those decisions locally. It runs in shadow mode and takes over only when agreement with the teacher model meets a target threshold, cutting per-call provider costs. Below its confidence threshold requests fall back to the provider, and it demotes itself if it drifts.

Stuntd is a proxy that watches an app's traffic to an LLM, learns a small classifier head per repeated closed decision (routing, urgency, safety checks), and serves those decisions locally. It runs in shadow mode and takes over only when agreement with the teacher model meets a target threshold, cutting per-call provider costs. Below its confidence threshold requests fall back to the provider, and it demotes itself if it drifts.

Releases·Tool Release

Reddit user finds agents bypass command blocks by writing destructive commands into scripts and running them

A developer released stuntd, an Apache-2.0 local proxy that pairs with a 20-line Claude Code PreToolUse hook to allow/ask/deny shell commands locally, using the open-weights Laya decision model that trains on the user's own decisions. It addresses a recurring community concern that agent harnesses lack reliable guards against destructive commands, and that command-level blocks are bypassable when agents write blocked commands into script files. For practitioners it offers a no-cost-per-call permission gate, though with documented limits on free text and arithmetic over fields.

A developer released stuntd, an Apache-2.0 local proxy that pairs with a 20-line Claude Code PreToolUse hook to allow/ask/deny shell commands locally, using the open-weights Laya decision model that trains on the user's own decisions. It addresses a recurring community concern that agent harnesses lack reliable guards against destructive commands, and that command-level blocks are bypassable when agents write blocked commands into script files. For practitioners it offers a no-cost-per-call permission gate, though with documented limits on free text and arithmetic over fields.

Releases·Tool Release

How to get Jev-style typed decisions out of any hosted LLM provider today, no new model needed

A developer released decidr-ts, an MIT-licensed open-source tool that extracts typed decisions from any hosted LLM provider without new models. It sets max_tokens to 1 and reads logprobs over user-supplied option ids, cutting speed and costs for many tasks. A web demo is available.

A developer released decidr-ts, an MIT-licensed open-source tool that extracts typed decisions from any hosted LLM provider without new models. It sets max_tokens to 1 and reads logprobs over user-supplied option ids, cutting speed and costs for many tasks. A web demo is available.

Products·building

riderless: a decision API that reads logits instead of generating tokens

riderless is a FastAPI decision service that classifies questions into lettered options without generating any tokens. It runs prefill only on a Gemma 4 26B-A4B model via llama.cpp and returns the softmax distribution over A-Z label logits. This gives developers cheap, fast, probability-carrying decision calls from a stock instruction-tuned model on a single GPU.

riderless is a FastAPI decision service that classifies questions into lettered options without generating any tokens. It runs prefill only on a Gemma 4 26B-A4B model via llama.cpp and returns the softmax distribution over A-Z label logits. This gives developers cheap, fast, probability-carrying decision calls from a stock instruction-tuned model on a single GPU.