Pre-release documentation. This guide describes the current product design and evaluation build. Packaging, availability and commercial terms may change before release.
T0: Decided
T0 performs symbolic resolution over the sealed graph with no model call. The same version, facts, as-of time and operating point produce the same decision path.
- Best for regulated and high-consequence checks
- Bit-reproducible
- Works with no model configured
T1: Decided with judgment
T1 adds bounded scalar oracles. An oracle may evaluate a prose condition, normalize an ask or narrow candidates. It returns a boolean, scalar or identifier list, never prose presented as authority.
- Each judgment records the seam, model, input and prompt hashes, output and latency
- Replay reads the recorded judgment instead of invoking the model again
- An oracle may narrow a candidate set, never widen it
T2: Reasoned
At T2, a model emits a closed typed plan. Rulebook validates the entire plan against the pinned graph, then a constrained interpreter executes it without model, network, filesystem or clock access.
- Allowed steps are apply rule, derive fact, calculate, lookup, combine and cite
- Every premise must rest on confirmed authority
- Invalid plans are rejected as a whole
- Execution is replayable from the stored plan
T3: Explained
T3 composes a labelled narrative over a T0 to T2 result and checks that the narrative is entailed by that result. The explanation is not reproducible and never changes the outcome.
Operating points and human review
The caller chooses a ceiling: strict stops at T0, fast permits T0 to T1, balanced permits T0 to T2, and best permits T0 to T3. A caller can also reject an answer above its own threshold by inspecting the returned tier and receipt.
- Needs review is a first-class decision outcome
- The response names missing facts, ambiguous rules or blocked conditions
- Human approval can be an enforcement obligation at any reasoning tier