The reality-oracle thesis
The founding point of view, expanded — convert events into affirmable facts and real probabilities.
The thesis is one sentence: turn the stream of events the world produces into facts you can affirm and probabilities you can defend. Everything else is engineering.
A reality oracle is not a system that knows what is true. That system does not exist, and any product claiming to be one fails the first question a regulator asks. A reality oracle is a system that turns the world into two things an examiner can use: affirmable facts (this happened, at this version, from this source) and defensible probabilities (this is how the future forks, here is the shape of the fork, here is how much I disagree with myself). The first is provenance; the second is a calibrated distribution. The thesis is that those two, together, are the substrate a regulated decision actually needs.
Affirmable facts, not ground truth
An affirmable fact is a record with provenance, not a claim about reality. The GNL does not certify that a node is true; it certifies which version of which layer produced an output and that the computation over those versions is reproducible. A NodeVersion is immutable and hash-linked (BR-842), so the state of the world at any past moment is reconstructable bit-for-bit. That is what makes a fact affirmable: not that it is correct, but that the chain of how it got here is sealed and replayable. We are pre-revenue, building the POC; the layers are built but inert until calibration fills them.
Real probabilities, not confident numbers
A real probability is a number that earned its unit through a calibration loop, not a number an LLM said with conviction. The sisters produce a distribution; the Oracle normalizes it in exactly one place (deino-oracle), with entropy in nats so a CRO can read the spread. Calibration is the test: shadow runs against a realized timeline, scored with Brier and log-loss via deino-eval, gated by Gate 1b before any forecast side is positioned as accurate. A probability without that loop is decoration.
Affirmable facts and defensible probabilities are two different outputs from one stream. The oracle does not conflate them, and neither does a regulator.
Why the two belong together
Most AI systems produce one or the other. A data warehouse produces affirmable facts (if it is versioned) but no probabilities. A model produces probabilities (if it is calibrated) but rarely versioned provenance. The thesis is that the regulated decision needs both in the same artifact: the manifest attests how the forecast was produced, the distribution is the forecast, and the calibration loop is what makes the probabilities real. Separating them is what makes each one weaker.
- Affirmable fact: a versioned, hash-linked record of what happened — provenance, not truth.
- Defensible probability: a calibrated distribution over what may happen — earned by a loop, not asserted.
- The manifest binds them: the forecast is reproducible because the facts it read are sealed.
The line the thesis refuses to cross
Certified is not true. The manifest certifies provenance and reproducibility; it does not certify that the forecast was correct. The refusal to certify truth is what makes certifying provenance credible — an oracle that claimed to know what was true would fail procurement the moment an examiner asked the question. The thesis is honest about the boundary: we affirm how we got here, and we forecast how it forks, and we do not pretend either is the same as knowing what is real.
A reality oracle does not know what is true. It converts events into affirmable facts and defensible probabilities, and the difference between those and truth is exactly what makes it usable in front of a regulator.
— reality-oracle thesis note