Questions the mathematics makes askable.
The framework underneath our products — budgets that hold, guarantees that compose, receipts that prove — keeps suggesting things that don't exist yet. This page is where we think out loud. Nothing here is a product; no numbers, on purpose. When a question earns its numbers, it graduates to the product pages and to Proof.
Questions we are working on
Detection ceilings
What if a certificate could state what NO detector can achieve on a given signal? Procurement could price monitoring claims against a proven ceiling instead of a vendor’s slide.
Alarm budgets that cannot be gamed
Could a false-alarm guarantee be proven to hold even against an adaptive adversary? Not “we catch attackers” — no one can promise that — but a budget that cannot be manipulated, which is the promise a regulator could actually use.
A grammar for combining guarantees
Safeguards multiply; promises don’t. What would it take for a fleet of safeguards to carry one provable promise — and to refuse, by construction, the combinations that would silently break it?
Counterfactual replay
What would last quarter have looked like under a stricter cap? Your own receipts may already contain the answer.
Risk budgets as a market
Could automation capacity be allocated like any other budget: departments buying against a shared, certified error budget, at prices that reflect the risk?
Segment isolation
Could one team’s traffic be proven unable to erode another team’s guarantee? One platform, many tenants, separate promises.
Receipts for reasoning
Could the chain extend from the decision to the deliberation behind it: evidence of what a conclusion was based on, not just what it was?
Same mathematics, new territories
Wherever decisions are costly and must survive an audit, the same questions apply. These are the ones that keep us up at night, in the good way.
Execution under a hard cost budget (finance)
Desks execute models under real-world costs. Could the execution itself carry a certificate: a hard ceiling on cost that holds through calm and through stress, with the price of the guarantee measured? Not a better model — a provable cap on what running yours can cost.
Model-risk certificates (banking)
How wrong can a model be before the decision changes? Could that plateau be certified, in the language model-risk teams already speak?
Decision records for regulated science (pharma)
Batch release, signal assessment, quality control: could the same receipts and error budgets map onto the evidence duties of medicines regulators, across a product’s whole lifecycle?
Assurance for systems that act (defense & safety-critical)
What may a machine do alone, and what must always reach a human? Could authorization chains be built to survive hostile scrutiny — with alarm budgets that hold even against an adversary? Catching every attacker is impossible to promise; an alarm budget that cannot be gamed may not be.
Compute budgets for reasoning (AI economics)
Reasoning models spend tokens like money, often past the point of value. Could the stopping rule itself be certified — the moment the next thought stops being worth its cost?
Questions, not commitments.
If one of them is a problem you own, we would like to compare notes.