Knolo Cortex

Knolo Cortex is reasoned evidence retrieval over a mounted V4 pack or a V5 Knowledge Image. Import it from @knolo/core/cortex. The root @knolo/core entry does not load it.

createCortex on the root export is a different API. That helper is the memory overlay. Knolo Cortex is createCortexSession.

The host owns any model. Core does not call a network, does not embed a model, and does not decide that an answer is true. verifyCortexAnswer recomputes traceRoot, answerRoot, and evidenceRoot, checks evidence identity, replays the retrieval receipt, and recomputes KAR references. Evidence must be selected by a replayed round. A dossier must match the trace body. A passing verification means the trace matches the declared evidence. It does not mean the claim is proven.

Rust, Python, and ICP do not implement Cortex. They stay on 5.5.1.

Loop

Exampletext

normalize question
  → retrieval round (V4 lexical receipt, or V5 EQL SEARCH)
  → optional host model selection
  → optional KAR evaluation
  → gap or refined query, inside the policy budget
  → answer from selected evidence only
  → Cortex trace

The model sees the question, hit ids, and excerpts of at most 500 characters. It never receives the corpus. A selected id, citation, or claim that names an id outside that round fails closed with CORTEX_MODEL_UNBOUND. A planned subquestion still runs when an earlier round adds no evidence. Refinement stops on that stalled round.

Semantic rerank stays off unless allowSemanticRerank is true and the caller passes a query embedding for vectors that are already on the pack. There is no vector database and no embedding provider inside this entry.

Modes

Mode	Result
`find_relevant`	Selected evidence and no prose. `ANSWERED` when a hit survives.
`evidence_answer`	Prose only from the host `synthesizeAnswer` adapter, with citations. Without a model, and when abstention is allowed, the status is `INSUFFICIENT_EVIDENCE`.
`compare_sources`	Needs two distinct sources. Fewer than two is `PARTIAL` or `INSUFFICIENT_EVIDENCE`.
`prove_or_abstain`	Requires a bound Committed Evidence Graph and a KAR plan. `SATISFIED` means the evidence requirements were met.
`coverage_review`	Reports subquestions that retrieved nothing. A bound plan adds the KAR status as a gap. It does not load a Domain Pack.

Answer statuses are ANSWERED, PARTIAL, and INSUFFICIENT_EVIDENCE.

Policy

Defaults are 4 rounds, 80 candidates per round, and 12 evidence items in the answer. Hard caps are 8, 200, and 32. The caller supplies createdAt when it wants one. Core does not read the clock, and createdAt is not part of the three roots.

API

Examplets

import { askCortex, createCortexSession, verifyCortexAnswer } from '@knolo/core/cortex';

const session = await createCortexSession({
  image,
  policy: { maxRounds: 2, allowAbstain: true },
});
const trace = await askCortex(session, {
  question: 'Where is the valve inspection?',
  mode: 'find_relevant',
});
const verified = await verifyCortexAnswer({
  image,
  question: trace.question,
  result: trace,
});

verified.ok means the roots match. The dossier field verified: true means the same thing. Neither one is a truth verdict.

CLI

knolo cortex find, knolo cortex ask, knolo cortex explain, and knolo cortex verify read local files. They do not take an API key. Pass a model only through createCortexSession in the host.

Public vectors: conformance/cortex/v1 in Knolo Core.