TypeSafe System One FORMAT (not a clone)

What we stole

From signed-in console.typesafe.ai/hook + docs:

Primitive	Returns	Trading use
Noul	P(yes)	Is this claim supported by ledger facts?
Choice	option + probs + confidence	allow / abstain / deny
Score	weighted level + confidence	harm if asserted falsely
Confidence routing	second axis	low confidence → abstain

Operator path

# Offline honesty (CI / no key)
.venv/bin/python scripts/typesafe_claim_gate.py --offline \
  --claim "Put credit has proven positive expectancy"

# Live Jev (needs TYPESAFE_API_KEY)
.venv/bin/python scripts/typesafe_claim_gate.py \
  --claim "Put credit has proven positive expectancy"

Exit codes: 0 allow · 2 abstain · 3 deny.

Credentials

• Keychain: service TYPESAFE_API_KEY, account hermes-fleet
• Fallback file: ~/.resume_secrets/TYPESAFE_API_KEY (mode 600)
• Console keys: TypeSafe keys
• Global retrieve skill: /typesafe-ai-api
• Official API skill (Claude): ~/.claude/skills/typesafe-ai (upstream)

Never print the API key. Never commit it.

Parallel questions (cookbook)

Source: Parallel questions

• Batch independent Noul/Choice/Score questions in one system_one call against the same state.
• Cookbook: 13 questions / one ~54k doc → ~12× cheaper and ~10× faster vs N serial calls; answers unchanged.
• Claim gate: 3 questions / 1 call. Skill suggest: batched Choice+gates, then batched fits (2 calls total — progressive disclosure).
• NEVER one system_one per question when they share state.
• Next ROI: one-document ledger/RAG briefings with many parallel Nouls in a single request.

Hard don'ts

NEVER	ALWAYS
Clone TypeSafe / sell a Jev wrapper SKU	Steal typed questions + confidence routing
Claim edge from paper theater	Gate profitability claims through this CLI
Require paid API for CI	--offline deterministic heuristic
Dual-edit sibling Typesafe WIP without claim	AGENT-651 worktree + vault claim

Related

• src/adapters/typesafe_client.py
• scripts/typesafe_claim_gate.py
• scripts/superpowers_verify_complete.py (completion evidence)
• LL-651