🧠 Decision Kernel

The judgment API — a decision that was too expensive per row now fits in your pipeline.

Decision Kernel exposes TypeSafe Jev (a System One decision model) as a fast, calibrated, low-cost judgment API. Send text; get department (triage), urgency, spam, tone, and whether it needs human review — all in one request, ~0.5s, with explicit confidence.

It also exposes /guard — an agent-safety gate: before an agent spends money, deletes data, or touches production, ask Decision Kernel to judge approve / review / block with Jev's verdict.

Deployed on Cloudflare Workers. Judgment proposals only — no money movement, no financial/medical/legal accountability (classification + routing suggestions only). The guard never executes anything itself.

Why judgment APIs

Most agent loops ask an LLM to pick between a few known answers, token by token, slowing every call and paying for prose nobody reads. A decision model returns a typed answer and a calibrated probability in milliseconds. Where code already knows the possible answers, a judgment is the right interface.

LLMs think. Jev decides. Tools act.

Try it

• Live web demo: https://decision-kernel.pickaxe.workers.dev/
• API: POST /classify (email/ticket triage)
• API: POST /guard (agent action safety gate)

/classify — route a customer message

curl -X POST https://decision-kernel.pickaxe.workers.dev/classify \
  -H "Content-Type: application/json" \
  -d '{"text":"my app keeps crashing when I try to export my data, please help"}'

{
  "judgments": {
    "triage":  { "choice": "technical", "confidence": 1 },
    "urgency": { "score": 0.4 },
    "spam":    { "score": 0.03 },
    "tone":    { "choice": "angry", "confidence": 0.45 },
    "risk":    { "score": 0.57 }
  }
}

/guard — stop an agent from doing something dangerous

curl -X POST https://decision-kernel.pickaxe.workers.dev/guard \
  -H "Content-Type: application/json" \
  -d '{"action":"delete the leads database file","cost":0,"goal":"send a newsletter","lastActions":[]}'

{
  "guard": {
    "verdict": "block",
    "spend": "deny",
    "confidence": 0.99,
    "off_goal": 0.95,
    "duplicate": 0.03,
    "undoable": 0.2,
    "last_ok": 0.33,
    "lead_real": null
  }
}

Verdict logic: spend:deny → always block; review or low-confidence → review; otherwise approve. A deny stops the action no matter what the numbers say. A judgment proposal only — code in the calling agent decides whether to proceed.

The judgments (one request, all parallel)

Judgment	Type	Meaning
triage	choice	Which department (billing / technical / account / sales)
urgency	noul	Needs immediate attention (0–1)
spam	noul	Is this an automated pitch (0–1)
tone	choice	positive / neutral / angry
risk	noul	Needs human review before any automatic action (0–1)

Guard field	Type	Meaning
spend	choice	approve / review / deny
off_goal	noul	Off today's stated objective?
duplicate	noul	Already done in recent history?
undoable	noul	Can it be reversed if it goes wrong?
last_ok	noul	Did the last step succeed?
lead_real	noul	(when a lead is passed) is this a real enquiry?

Every judgment returns a calibrated probability, so routes can split on confidence (act automatically / confirm / escalate to a human).

Pricing (planned)

Tier	Price	Includes
Free	$0	10 judgments / day (demo)
Starter	$29/mo	2,000 judgments, all 5 types
Pro	$99/mo	10,000 judgments, parallel, priority
Enterprise	$499/mo	Custom judgment types, SLA

Guard Skill (paid): the skill/decision-kernel-guard/SKILL.md — a drop-in agent skill file that wires /guard into Hermes, Grok bot, or any agent, with policies (cheap+reversible runs at ≥0.7, money/deletion requires ≥0.9, deny always stops, hold on 3s no-response). Request via the pricing above.

Cost is a Cloudflare-free-tier Worker + Jev (~$0.04 / M input tokens) — >90% gross margin. Target: small SaaS / support teams / agent developers in English / India / Indonesia.

Verified results

40-agent-action day, simulated against the live /guard (no funds moved):

Category	Result
10 malicious (pay transfer, DB delete, credential leak)	blocked 10/10 ✅
21 normal (fetch, render, validate, preview)	approved ✅
9 borderline (bulk send, CSV export, schedule)	routed to review ✅
false-block rate on safe actions	0 (after verdict fix) ✅

Demo: demo/guard_demo.py

Agent discovery

Decision Kernel is built for AI agents:

• llms.txt — index & briefing
• agents.txt — capability card
• agents.md — human-readable guide
• openapi.json — machine-readable schema
• robots.txt — explicit AI-crawler allow
• sitemap.xml
• /.well-known/agent.json — AgentCard
• /.well-known/ai-catalog.json — ARD registry

Run it yourself

npx wrangler deploy
# then set the secret (never committed):
echo "$YOUR_KEY" | npx wrangler secret put TYPESAFE_API_KEY

Stack

• Cloudflare Workers (free tier, serverless)
• TypeSafe Jev (System One decision model)
• MIT License

---

Judgment proposals only. No money movement.