What we can show you, and what we can’t.

This page exists to sell you something. Everything on it is either checkable by you, or labeled as something we can’t check either.

We don’t publish an accuracy number

Every company in this category leads with one.

The ceiling for this kind of work is how closely a person matches their own answer when asked the same question twice. People don’t match themselves. A percentage without that denominator, its question set, and its holdout protocol is unbounded — it can mean almost anything.

We have no published calibration record. Until we do, we won’t imply one exists.

What you can check right now

The record format is published. Timepoint Telemetry — tt-ontology/1.0 v2.1.0: 149 nodes across two disjoint lenses, 151 lateral edges, 26 bridges, 3 kernel bridges. 10 envelope conformance vectors and 39 classification verdict vectors with committed inputs, canonical bytes and hashes, and 86 tests across two independent implementations run in CI on every push (as of 2026-08-18).

Rerun the vectors and check the hashes byte for byte.
→ github.com/timepointai/timepoint-telemetry · source-available under BSL 1.1

Governance runs on a schedule, in public. Three change classes, two release windows at NYSE close, retirement only by deprecation, published compatibility criteria. Each version converts to Apache-2.0 four years after its own publication, so you can read your permissive date off a calendar.

Every simulated output is labeled. Fictional examples say so in the first line, not a footnote. No simulated result is presented as a real outcome, real traction, or a case study.

How uncertainty is reported

Ranked branches with the uncertainty written out in words. Where repeated runs disagree, the disagreement is reported rather than averaged away.

Someone at Timepoint reads every client deliverable against your stated boundaries before it reaches you, for what it’s arguing rather than what it says. The engine will route around a stated constraint semantically if you let it.

The offer: replicate a decision you already made

Give us a decision your organisation made at least a year ago whose outcome never became public. Tell us what you knew at the time and nothing about what happened. We simulate it blind and hand you the ranked branches. You compare against what actually occurred.

The non-public part is the whole test. A decision whose outcome was reported anywhere is inside the training data of every model we run — we’d be measuring recall, not simulation. Most real decisions were never public, which is why this works.

If you’d rather not take our word for the blindness, have someone on your side hold the outcome and score the result.

You keep the result either way. If it misses, you find that out too.

What we can’t show you

• No calibration record. Not published, no date promised.
• No third-party evaluation. Nobody outside Timepoint has assessed our output.
• No security certification. No SOC 2, no uptime SLA. If procurement is your gate, we will fail it.
• No named customers anywhere on this site. We name none, and we don’t claim a roster you can’t see.
• Named actors are what a simulation is about, not a fidelity guarantee. In testing, the engine has relabeled which side a named person was on.
• We run on off-the-shelf models.
• We are in alpha.

Sources

Every figure above, with where it comes from and how it was checked.