The plain-language case
If AI must reread the whole world every second, that is not understanding.
Treating a model like a full-speed camera is what happens when a system has no visual causal projection. APR's design proposition is to keep bounded state, model ordinary continuity, and spend perception when change or uncertainty actually justifies it.
APR attitudeDesign proposition: model the ordinary, observe the difference, verify before acting.
Brute-force multimodality is not intelligence
Reading every frame, the whole frame, at maximum depth can imitate awareness by spending more compute. It does not prove that the system understands continuity, predicts consequences, or knows what no longer needs attention.
- Brute force Full frame → full read → expensive Provider → forget → repeat.
- APR Change event → perceptual need → targeted evidence → bounded state → verify.
The five-step shortcut
The APR proposition does not require a complete copy of the world every second. A capable implementation should preserve what has been verified, track what may have changed, and escalate observation only when the next decision needs more evidence.
- Change Detect an event or difference before requesting a full read.
- Proposition Use verified state to model ordinary continuity; re-observe when the model is insufficient.
- Need Name the fact that is unknown, stale, conflicted, or action-critical.
- Read Choose the cheapest sufficient region, depth, and modality.
- Verify Re-observe the outcome when an action can change the world.
What users and investors actually care about
Less avoidable Token use can mean lower latency, fewer API cost spikes, less context pressure, longer-running sessions, and a system that can scale without treating every quiet second as a brand-new universe.
- User Faster responses and longer useful multimodal sessions.
- Operator Fewer unnecessary Provider calls and clearer budget control.
- Investor A path from impressive demo to economically repeatable runtime.
- Agent Persistent evidence instead of perpetual visual amnesia.
Token savings are only the first visible result
APR is also about persistent world state, evidence provenance, visual causal projection, action gating, outcome verification, and recovery. Compression alone cannot tell an agent when it is safe to act or when reality has invalidated its prediction.
Current implementation boundary
The current v0.10 runtime does not yet predict arbitrary visual scenes. It implements event, evidence, world-state, perceptual-need, budget, action-gate, outcome-verification, and recovery primitives; learned visual causal projection and measured savings remain research work.
No magic savings claim
Actual savings depend on the workload, scene-change rate, task risk, available structured signals, and Provider pricing. APR removes avoidable perception by method; benchmarks must measure the result instead of inventing a universal number.
Source evidence
Immutable links pinned to the candidate source reference.
- runtime_architecture
docs/runtime/ARCHITECTURE.md - event_runtime
docs/runtime/EVENT_NATIVE_RUNTIME.md - need_graph
docs/runtime/PERCEPTUAL_NEED_GRAPH.md - budget
apr_runtime/budget.py - outcome
docs/runtime/ACTION_OUTCOME_VERIFICATION.md - apr_04
papers/APR_Paper04_Persistent_World_State_Differential_Reobservation_v0.1.md - whitepaper
docs/theory/WHITEPAPER_APR_Runtime_v1.0.md