Layer 1 — PASR Local Runtime
Gate proposed interventions and create evidence.
- gate
- evidence / readout
- event store
- geometry snapshot
- readout memory
- Docker memory-loop evaluation
PASR Local · architecture
PASR Architecture
PASR is a white-box governance runtime for proposed interventions.
PASR sits between an AI agent, operator, simulator, workflow, or actuator-related proposal and the downstream system that could be affected. It checks whether a proposed intervention should proceed, records a bounded decision, reconstructs the shape of repeated decisions, and exports safe evidence.
PASR does not control actuators. PASR governs whether a proposed actuator-related intervention is allowed to proceed toward the manufacturer's control plane.
The core chain
proposed intervention → gate / evidence → bounded decision event → append-only local memory → geometry snapshot → resonance recall → intervention_memory readout → safe export
- gate / evidence
- checks whether a proposed intervention should proceed.
- bounded decision event
- records what PASR decided without raw payloads.
- append-only local memory
- keeps an auditable local trail.
- geometry snapshot
- reconstructs folds, creases, seams, and stitches.
- resonance recall
- recalls prior creases / seams relevant to a similar new proposal.
- intervention_memory readout
- shows operator-facing memory context.
- safe export
- human / JSON / JSONL artifacts — not the private memory engine.
Three-layer architecture
Layer 2 — FoldEngine Memory
Private geometric memory / projection / resonance system.
- decision event = fold
- repeated blocker = crease
- readiness gap = seam
- operator action = stitch / repair
- resonance recall over similar intervention geometry
- private reconstruction / compression logic remains internal
- not exposed as a raw memory engine
Layer 3 — Export Adapters
Safe projections for humans and systems.
- human-readable Markdown
- bounded JSON
- event-only JSONL
- future ROS/ROS2 governance event adapter
- future SIEM / OpenTelemetry-style export
- future ledger-compatible summary
- future customer storage export
The export is not the memory engine; it is a safe projection of what the memory engine knows.
Why this is not just an allowlist
An allowlist says whether an action is permitted by policy. PASR records and reconstructs what happens across repeated intervention decisions:
- repeated BLOCK events form a crease
- NOT_READY forms a readiness seam
- WAIT_FOR_PLANE_READY becomes the suggested stitch
- similar future proposals can recall prior creases before a path proceeds
Recall adds context, review, or block hints. It does not silently bypass PASR gates.
Robotics / actuator boundary
Manufacturer controller
- physical control
- timing
- firmware
- sensors
- interlocks
- safety certification
PASR
- proposed-intervention governance
- readiness / evidence checks
- bounded decision memory
- operator explanation
- safe export artifacts
PASR does not replace robot controllers, firmware logs, ROS/ROS2 logs, diagnostics, or safety systems. It can add a governance event layer before proposed interventions reach the control plane.
What remains private
- private FoldEngine Memory internals
- resonance scoring internals
- Core3 memory state
- kernel internals
- engine_codex internals
- proprietary reconstruction / compression logic
- raw reconstruction graph
- raw path weights
- raw compression state
What can be exported safely
- bounded PASR decision events
- geometry snapshot summaries
- intervention_memory readout sections
- human-readable Markdown
- bounded JSON
- event-only JSONL
- reason codes
- blocker codes
- mapped operator actions
- readiness states
- evidence refs
- receipt refs
- continuity refs
What this architecture is not
- It is not hardware control; PASR does not control hardware.
- It is not actuator control; PASR does not control actuators.
- It is not safety certification; PASR is not safety-certified.
- It is not an autonomous actuator controller.
- It is not a ROS/ROS2 runtime today, and does not replace ROS/ROS2.
- It is not ledger writes today.
- It is not a raw memory dump, and does not store raw payloads.
- It is not full Core3 included.
- It does not expose private FoldEngine Memory internals.
- No cloud is required for local evaluation; there is no cloud dependency.
See it in practice
Sample: a BLOCKED tool call
What a single gate / readout artifact looks like.
Sample: memory loop evaluation
What the memory-loop readout and exports look like.
The opt-in Docker memory-loop evaluation is available for technical review. PASR Local Evaluation Pack is delivered manually by request while the product is in early evaluation.