Skip to content

concepts

Intended Documentation

Semantic Intelligence

How the Large Intent Model interprets an action into ranked hypotheses and risk signals — and how those signals shape, but never weaken, the authority decision.

intermediate4 min readimplemented

Semantic Intelligence#

Semantic intelligence is the interpretation layer of the authority runtime. Before any decision is made, the Large Intent Model (LIM) reads the proposed action and projects structured meaning onto it — what the action is, how confident the interpretation is, and which risk signals matter. The policy engine then decides; LIM does not. Its job is to make the decision better-informed, not to make it.

LIM interpretation
A proposed action enters the Large Intent Model, which emits ranked hypotheses (a primary plus alternates) and derived signals — confidence, action classification, extracted risk, and process conformance. Signals feed the Authority Engine and can only raise severity.

LIM is an interpreter, not a classifier: it emits ranked hypotheses and signals. The primary drives the decision; alternates are recorded for explainability.

An interpreter, not a classifier#

A classifier picks one label and discards the rest. LIM instead emits a ranked set of hypotheses — a primary interpretation plus up to four alternates — each carrying its own confidence, rationale, assumptions, and any missing fields. Only the primary hypothesis drives the authority decision; the alternates are written into the audit chain so a reviewer can see what else the action might have meant. LIM runs on a configurable backend per tenant: a deterministic rules backend, an llm-api backend (the one that produces multiple hypotheses), or a local-model backend.

Derived signals#

From the interpretation, LIM derives the signals the policy engine and your policies can reason over:

SignalWhat it captures
confidenceHow sure LIM is of the primary interpretation (0–1)
action classificationThe kind of action being interpreted
touchesProductionThe action affects a production system
requiresPrivilegedAccessThe action needs elevated privilege
containsSensitiveDataThe action handles sensitive/regulated data
assumptions / missingFieldsWhat LIM had to assume, and what was absent
process-step match & sequence conformanceWhether the action fits the expected business process and order

These extracted risk signals are the same factors that feed the risk score. The conformance signal comes from a deterministic sequence-conformance scorer (coverage and order against the tenant's process maps).

Signals can only raise severity#

This is the central guarantee of the semantic layer: signals are upgrade-only. They can turn an ALLOW into an ESCALATE or a DENY, but nothing LIM produces can turn a DENY into an ALLOW.

Warning

The semantic layer never grants authority. It can only add caution. Concretely: low interpretation confidence, a conflict between the risk LIM extracts and the risk the caller declared, or a low process-conformance score will push a decision toward escalation or denial — never toward approval.

Representative thresholds applied after policy matching:

  • Confidence — below 0.3 denies; below 0.5 (and below 0.85) escalate for review.
  • Signal conflict — if LIM detects production/privilege/sensitive-data exposure that the declared risk context understates, the decision escalates.
  • Conformance — a sequence-conformance score below the tenant's deny threshold denies; below the escalate threshold escalates.

Scope today#

Info

Accuracy note: the digital (text) interpreter classifies actions into a small, deterministic action vocabulary today — it does not yet classify free-text actions into all 173 categories of the Open Intent taxonomy. The full OI-NNNN classification is produced by the physical structured-goal classifier (see the Physical AI track). Treat broad-taxonomy classification of arbitrary text as roadmap, not current behavior.

Semantic Intelligence | Intended