An open specification for treating reasoning as durable infrastructure.
Most AI forgets how it decided. LIA is the contract for systems that don't.
Most current AI integrations conflate distinct concerns into single opaque components. Reasoning, retrieval, validation, orchestration, and human review collapse into one prompt and one model call. The result is systems that are brittle at the edges and expensive at scale. When an answer is wrong, there's no way to tell whether the failure came from what was retrieved, what was inferred, or what was invented.
LIA defines deliberate seams between concerns, so brittleness is detected early, cost is controlled at the routing layer, and the system's reasoning can be inspected, evaluated, and improved over time. It is a contract, not a library.
Each layer accepts a structured object from the layer above and produces one for the layer below. The contracts are the protocol: two implementations that emit the same contracts can swap internals freely.
CanonicalInputObject
RoutingDecision
ExecutionResult
FeedbackSignal
The formal JSON Schema definitions, with validated examples, live in /schemas. They are the canonical, machine-validatable definition of what a conformant implementation must emit at each boundary.
"routing_decision": { "trace_id": "9f2c…", "route": "known_complex", "tier": "frontier", "evaluators": { "precedent": 0.31, "complexity": 0.78, "risk": 0.64 }, "thresholds_version": "2026.06", "latency_budget_ms": 20000, "checkpoint_required": false }
A component that mixes types is non-conformant, because a component that both retrieves and reasons can't be tested, priced, or replaced independently.
Take the most common AI system in production today: a retrieval-augmented support bot with one prompt and one model call.
Reasoning, logic, routing, and a human checkpoint, fused into a single call. When it breaks, you can't say which part failed. The primer reads this exact system through the framework and shows the five bounded changes that fix it.
All four layers exist with correct boundary contracts. Node types are correctly classified. A profile is declared. Trace is emitted.
Level 1 plus a functioning closed loop with declared feedback contracts, externalised routing thresholds, and complete observability.
Level 2 plus calibrated confidence aggregation, versioned thresholds, documented loop effectiveness, and a published profile test suite.
Implementations declare the level they claim and the profile they conform to. Conformance test documentation is planned; the levels are normative in the base specification today.
Routes, evaluators beyond the required three, tier taxonomies, output shapes, and feedback semantics are domain decisions. A profile makes them for one domain.
The Resolution Atlas
Customer support AI, with institutional resolution patterns treated as load-bearing infrastructure, not informal practice. Read the profile
The Evidence Graph
Investigative reasoning held as a revisable body of claims and evidence, where confidence moves in both directions. Read the profile
The Creative Genome
Creative AI products, with individual creative reasoning specified, refined through feedback, and, where appropriate, licensed. Contributed separately.
"Creative Genome" is a trademark in formation held by DREAMVISION AI. When the Creative Profile is published, its conceptual structure will be open under Apache 2.0 and freely implementable; the term itself is reserved as a product designation. New profiles for other domains are welcome; discuss yours in an issue before drafting.
Find them. Open an issue where the specification seems wrong or unclear. Accepted corrections are credited by name. Profile authorship is open; the base specification is editorially controlled.