The flight recorder for AI.
When an AI system makes a consequential decision, there is no trustworthy record of how confident it was at the moment it acted. CrashTrace seals the model's own uncertainty estimate — cryptographically, before the outcome is knowable — and makes it independently verifiable afterwards. Evidence that insurers can underwrite against, auditors can certify, and regulators can rely on.
After an AI failure, everyone asks the same question: what did the system know?
Aviation solved this decades ago: when something goes wrong, the flight data recorder provides an objective account that no party could have altered after the fact. AI has nothing equivalent. Logs are held by the operator, produced after the event, and editable by the very party under scrutiny. A model's stated confidence, reconstructed after an incident, proves nothing.
Operators can't prove diligence
Deployers of AI in healthcare, finance and safety-critical settings cannot demonstrate how much care their system applied to any individual decision — even when it acted responsibly.
Insurers can't price the risk
AI-liability underwriters publicly concede they lack trustworthy telemetry. Without tamper-evident records of model confidence, calibration is unverifiable and premiums are guesswork.
Regulators can't audit
Record-keeping and accuracy obligations for high-risk AI now carry legal force in the EU. Self-reported logs, editable by the operator, do not meet the standard of evidence.
Commit before the outcome. Verify after it.
CrashTrace is middleware that sits between an AI model and the systems that consume its outputs. It requires no access to model weights or prompts to verify — and it works with any model that can produce an uncertainty estimate.
1Seal
Before an output is released — while the outcome is still unknowable — the model's per-inference uncertainty estimate is sealed inside a cryptographic commitment. The estimate itself stays private.
2Anchor
Each commitment is appended to a tamper-evident, hash-chained log that is anchored externally, so not even the operator can rewrite history after the fact.
3Verify
Once the outcome is known, a reveal entry lets any third party — insurer, auditor, court — confirm what the system's confidence was, and prove it was recorded before the outcome existed.
Three forces converged in 2026.
EU AI Act high-risk obligations became enforceable on 2 August 2026 — including accuracy, robustness and record-keeping requirements that self-reported logs cannot satisfy.
AI-liability insurance is now a real market, with major underwriters and a growing Lloyd's presence — and an open need for verifiable calibration evidence to price cover against.
AI agents are doing consequential work at scale — clinical documentation, financial decisions, autonomous operations — while governance and evidence standards lag far behind.
Patent-backed technology.
The core CrashTrace mechanism — cryptographically verifiable pre-outcome commitment of a model's uncertainty estimate — is the subject of a UK patent application filed in July 2026, with search and examination requested. The technology is available for licensing, partnership and investment discussions.
Investment, licensing and partnership enquiries.
CrashTrace is led by Rich Cunningham, author of A Status-Based Accuracy Framework for AI Agents — a safety-engineering framework drawing on practice from aviation, anesthesiology and naval nuclear propulsion. Serious enquiries from investors, insurers, AI vendors and certification bodies are welcome.
Get in touchcontact@crashtrace.ai
Detailed technical material is available under NDA.