the claim you can try to break
Once a recomputable proof of where an AI agent stayed exists, deploying an
agent without one stops being prudent and becomes negligent — the same
turn the 1932 T.J. Hooper ruling made for a tugboat sailing without a weather radio
once radios worked. That proof exists now, on your own laptop, for free.
recompute it yourself — no login, no call, no trust asked
$ npx thetacog-mcp attest-demo
→ placement [B,A1] · σ 1.5 · LANE: IN · ed25519 signed ✓
The mechanism — three signed steps, no vote taken
- step 1 · node a, the buyerSigns the spec. A human-legible intent compiles to coordinates on a fixed, published 144-region map of meaning and is sealed — a contract boundary, not a vibe.
- step 2 · node b, the vendor / agentSigns the payload. The deliverable is hashed and bound to the exact spec it answers and to who produced it. No arguing later about what was asked vs. shipped.
- step 3 · the silicon oracleThe chip returns the verdict. Re-derives, from the two signed artifacts alone, where the work landed — a deterministic MATCH / no-MATCH plus σ. Any third party re-runs it offline, byte for byte, without calling us.
Real receipt, open it yourself: thetadriven.com/commit/60d4a3c6b · verify the signature with no server in the loop at /trust.
The one real number
sealed · dogfooded · a loss ratio, not a vibe
~13%
of agent actions, across our own dogfooded work, land outside the lane they were authorized for. That is the first loss ratio ever computed for AI competence. Your operation’s version of this number is not high or low — it is blank.
what the receipt proves — and does not
Proves: whether the agent’s action landed inside the domain you authorized — decidable, signed, recomputable by a stranger.
Never claims: that the work was good, correct, or “legally defensible.” That question is Rice-undecidable — no receipt, model, or vendor can honestly settle it. We price the lane, not the quality.
the founding partner package · per-agent license
$20 / agent / year
flat, no tiers — a fleet is the unit × your agent count.
Below ~15,000 licenses (≈$100K): you are buying licenses at the flat rate. Nothing else attached — no audit, no consulting.
Cross ~15,000 licenses (≈$100K) and the purchase auto-attaches a Discovery & Grounding audit: a forensic mapping of where your agents’ actions already land against the domains you actually authorized — today, under your practices as they stand, not a hypothetical future risk.
This is not consideration for assurance that your AI is correct, good, or safe — that claim is undecidable by construction and no one can honestly sell it. It is consideration for the audit itself: the mapping, handed to your own counsel as something concrete to act on.
Nomination-based, not promised — no seat is guaranteed by reading this. To nominate who belongs, use the form at /dinner.