# runtime cognition

Reasoning Runtime — how VAOS thinks

Between Scene Contract and Validation sits cognition. A plugin orchestrator that runs rules, models, and policies against the structured scene, against memory, against site context — and emits candidate actions with scored confidence.

Reasoning is a layer, not a model. SceneLM is one engine that runs inside it. So are rules, local LLMs, hybrid policies, and cloud adapters. The runtime orchestrates them — and hands the result to the validation envelope before any action leaves the system.
# inputs to reasoning

Six inputs. One coherent context.

🗺️

Scene Contract

Structured perception output — objects, regions, events, sensor reliability.

💾

Memory

Persistent local history — prior scenes, prior decisions, prior outcomes.

📐

Site policies

Configured rules — restricted zones, allowed actions, escalation paths.

⏱️

Temporal context

Trends, persistence, cooldowns, expectation models, time-of-day priors.

🧩

Plugin outputs

Whatever upstream extensions added — fused scores, custom detectors, derived intents.

🎛️

Sensor reliability

Per-stream confidence, surfaced into reasoning as evidence weights.

# runtime flow

From scene to candidate action

Plugin orchestration runs each cognition engine, scores its candidates, then hands the top-N to validation.

🗺️
Scene
Structured perception in
01
🧩
Plugin orchestrator
Runs each engine in priority order
02
🧠
Cognition engines
Rules · SceneLM · LLM · hybrid
03
⚖️
Scoring & fallback
Top-N candidates with confidence
04
🛡️
Validation handoff
Action only if envelope passes
05
# cognition engines

Supported engines

Mix and match. The orchestrator routes by capability, latency budget, and policy.

EngineRoleTypical latency
rulesDeterministic safety, escalation policies< 1 ms
scenelmEdge-distilled contextual reasoning20–80 ms
ollamaLocal LLM with structured output200 ms – 2 s
vllmHigh-throughput batched inference50–300 ms
cloudOptional cloud adapter (OpenAI, Claude, custom)500 ms – 3 s
memoryTemporal influence — priors, expectations< 5 ms
hybridComposed orchestration (rules + model + memory)varies
# reasoning trace

A real example

[scene]
event:       restricted_zone_entry
region:      loading_bay_03
objects:     person × 1
reliability: 0.92

[reasoning]
plugin:     scenelm
confidence: 0.87
memory:     prior_entries=2 in last_60s
intent:     "notify security; log incident"

[validation]
C_v:        0.84
threshold:  0.70
status:     PASSED

[action]
emit:       notify(target="safety-team", level="medium")
log:        scene + reasoning + envelope archived
# temporal reasoning

Reasoning over time, not just frames

📈

Trend analysis

Detect direction of change — crowd density rising, queue stalling, anomaly accumulating.

Persistence

Don't fire on a single frame. Require the same evidence over a window before escalating.

📊

Expectation modeling

Compare observed reality to historical baseline — when the world deviates, mark it.

🚨

Escalation

Stepped responses — log → notify → alert → page. Each step gated by a separate threshold.

❄️

Cooldown

Suppress repeated alerts. The runtime remembers the last action and waits a tunable window.

🎯

Predictive priors

Future SceneLM variants will inject expected next-state, so the runtime catches surprises faster.

# the architectural distinction

Reason ≠ Execute

Most AI vision systems collapse reasoning and execution. VAOS keeps them separate so the validation envelope can sit between them.

Reason → Validate → Execute. Every candidate action passes through the envelope. Below threshold, the runtime yields a no-op + diagnostic instead of an action — auditable, reversible, safe to deploy in physical environments.

Cognition that's portable.

Swap engines without rewriting the runtime. Compose hybrids. Trace every decision.

SceneLM Runtime Validation Engine Memory Engine