The agent loop: search results are gated before reading — real Jev calls, 60% of context never ingested. See examples/.
Drop-in agent skill: gate search/page ingestion through a cheap typed-decision model (Jev) BEFORE reading — keep agent context healthy.
Don't read into context what a 0.0001¢ classifier can judge first. Every
candidate page, search result, or snippet is one typed "noul" (yes/no
probability) question to Jev, batched into a single API call. Only pass: true
items get read into the agent's context.
Source: gate report from a real multi-source research session (2026-09-20, 277 gate judgments / 97 passed, ~85K tokens kept out of context, ~$0.007 — see REPORT.md §verification for the reproduced numbers).
| Metric | Without gate | With gate |
|---|---|---|
| Judgments | read everything | 277 gate judgments, 10 batched API requests |
| Passes | — | 97 passed (35%) → only those fetched/read |
| Spend | ~20 pages of raw HTML in context | ~$0.007 total |
| Tokens kept OUT of context | — | ~85K (42 gated-out page bodies ≈ 336K chars, plus 61 gated-out HN comments/snippets) |
| Page-level read cut | — | ~70% |
| Engine | p50 latency | Accuracy (24 cases × 3 q) |
|---|---|---|
| Jev (API) | 0.468 s | 0.83–1.0 on 2 of 3 q types |
| Laya (local MLX clone) | 0.028 s | 0.67–0.96 |
# Hermes
git clone https://github.com/nedzen/decision-gate ~/.hermes/skills/decision-gate
# Claude Code
git clone https://github.com/nedzen/decision-gate ~/.claude/skills/decision-gate
# Raw shell
git clone https://github.com/nedzen/decision-gate && ./decision-gate/scripts/gate.py --help# single text (stdin or --file); exit 0 = pass, 1 = fail
cat snippet.txt | python3 scripts/gate.py --question "is this about X?"
# batch: TSV lines id<TAB>text — all candidates, ONE API request
printf 'a1\tTitle: ...\n Snippet about transformers\na2\tTitle: ...\n Recipe blog\n' > cands.tsv
python3 scripts/gate.py --question "this text discusses transformer architecture or benchmarks" --batch cands.tsv --threshold 0.6Output: JSON lines {"id":"a1","score":0.93,"pass":true}. Batch mode prints one
line per item; single mode prints one line and sets the exit code.
Also included: scripts/flatten_md.py — flattens word-wrapped transcript
markdown into paragraphs (handy for prepping text before gating).
| Variable | Default | Meaning |
|---|---|---|
DECISION_GATE_URL |
https://openrouter.ai/api/alpha/decisions |
decisions endpoint |
DECISION_GATE_MODEL |
typesafe/jev-1.13 |
model id |
DECISION_GATE_API_KEY |
falls back to OPENROUTER_API_KEY env, then an OPENROUTER_API_KEY= line in ~/.hermes/.env |
bearer key |
Jev pricing: $0.042/MTok input, output free, 70–500 ms latency, 64K request / 32K state context (docs.typesafe.ai). At ~2000 chars per candidate, a 300-item batch costs ~$0.0007.
A local Jev clone on Apple silicon is the free/private option: Laya
(laya-typed-decisions-mlx, MLX) measured p50 0.028 s per decision vs
0.468 s for the hosted API, with usable accuracy (0.67–0.96 across question
types) in our head-to-head bench.
Status: roadmap, not shipped. We verified that laya is currently an
in-process MLX Python library (laya_mlx.Agent.predict(text, schema)) — it does
not expose the HTTP {model, state, questions} → {answers} decisions contract
that gate.py speaks. Wiring it up needs a small localhost server adapter;
until that exists, DECISION_GATE_URL pointing at localhost is NOT supported.
PRs welcome.
- The unsure band (0.3–0.6) is real. Scores in that band are neither yes nor no; usually it means your question is badly phrased, not that the threshold is wrong. Rewrite the question instead of lowering the threshold.
- Language mismatch dilutes scores. Non-English text scored against an English question came back 0.45 and was the most relevant source of our run. Phrase the question in the source's language or describe content generically.
- Extraction quality gates upstream of the gate. Snippet passed 0.97 but the
body failed 0.23 → that was a subscribe-wall extraction problem, not
irrelevance. Check
len(content)and header noise before trusting a low score. - Gate the atomic unit, not the container. Whole HN thread pages score 0.04–0.49; the same comments gated individually gave 4/65 sharp precision.
- Jev is a classifier, not a reasoner: decision-seam questions only ("can you write every possible answer before seeing the input?").
This project stands on the shoulders of:
- uehaj/jev-semgrep by Junji Uehara
— grep-by-meaning with Jev, one file, zero dependencies, and the
Zenn article
that went viral on X. Our local environment-key overrides
(
SEMGREP_URL/SEMGREP_MODEL), a cost line, and--hlhighlighting are being contributed upstream as PRs to that repo (fork after 14 days of no response — follow-up, not blocking this release). - TypeSafe — Jev itself, and the docs (typed questions, fan-out patterns, skill-suggestion cookbook) that define the decisions contract.
- classifier.dev — free zero-shot classification API built on Jev; their smart-tier benchmark ("re-ask only the unsure items: AG News 87.5% → 90.0%") is the best published calibration of the unsure band.
MIT — see LICENSE. Copyright 2026 nedzen. uehaj/jev-semgrep is MIT by Junji Uehara; semgrep.mjs itself is NOT included in this repo.