Concepts

The architecture of
verifiable intelligence.

Infrastructure for governed memory, replayable execution, and traceable reasoning.

Five Layers. One Governed System.

Knowledge, Memory, Learning, Governance, and Execution — operating in continuous coordination. SimCapsules specify how repeatable tasks are executed — tools, scope, permitted outputs, escalation conditions — governed at admission.

LAYER 05 · EXECUTION

Agent Runtime

Episode ingress · Admitted actions only

Inbound

LAYER 02 · GOVERNANCE LAYER

Admission gates · Constraint enforcement · Pre-execution verification

STEP 01

Admit gate

Auth check

›

STEP 02

Constrain

Schema · ACL

›

STEP 03

Verify

Pre-exec

Retrieval

Governed write

LAYER 01
EXTERNAL

Knowledge Banks

Curated corpora. Governed retrieval.

Curated Cited

LAYER 04
LEARNING

SimCapsules

Procedural Knowledge.

Governed Versioned

LAYER 03
MEMORY

Memory Runtime

Governed state. Causal history. Replay.

Provenance Replay

Cited results

Audit record

Governed response

Cited · Auditable · Provenance-stamped

Audit feedback: Memory Runtime continuously informs the Governance Layer.

Cognition loop: episode records feed into SimCapsules, which condition governance at next admission.

system
active
0
transitions
governed
provenance
preserved
retrieval
governed

Eight core properties.

Not design goals. System guarantees.

Governed Memory

Analogous to Disciplined observation and recordkeeping

Memory as a controlled system of record. Every state transition is policy-gated, every write is logged with its causal chain intact.

✗ Normal stacks mutate memory without policy — losing reasoning lineage entirely

✓ Every memory write is governed, auditable, and replayable

Provenance

Analogous to Citation, lineage, and traceable evidence

Every result carries a traceable chain back to its source. Citation-root provenance means you can always answer: where did this come from?

✗ Outputs appear without traceable evidence of what was relied on

✓ Full decision lineage preserved from source to output

Reproducibility

Analogous to Repeatable experiments and deterministic replay

Any reasoning state can be reconstructed exactly. Deterministic replay means you can re-derive any prior answer from the same governed inputs.

✗ State transitions are ephemeral — past reasoning cannot be reconstructed

✓ Content-addressed snapshots enable perfect deterministic replay

Constraint

Analogous to Methodology, protocol, and experimental controls

What rules bounded a system during a given reasoning episode. Policies are not aspirational; they are enforced at admission time, before state can change.

✗ Guardrails run post-hoc or in parallel, so actions may execute before policy checks complete

✓ Pre-execution policy gates; no state mutation without constraint satisfaction

Verification

Analogous to Audit, peer review, and evidence inspection

What can be independently checked. Exportable evidence packs let any auditor inspect the full reasoning chain without access to the live system.

✗ Audit means reconstructing context from scattered logs after the fact

✓ Evidence packs are exportable, inspectable, and independently verifiable

Knowledge Banks

Analogous to Governed literature base and curated corpora

Not generic retrieval. Curated corpora with citation-root provenance. The reference base every reasoning episode can cite and trace.

✗ Generic retrieval mixes governed and ungoverned sources with no lineage

✓ Curated corpora with full citation lineage and access-controlled retrieval

Agents

Analogous to Autonomous workers operating within governed boundaries

Runtime entities that submit work to the governed infrastructure. An agent does not authorize its own state mutations — every action is policy-gated, episodically bounded, and fully attributable. Agents execute within the procedures that SimCapsules specify.

✗ Agents call tools and mutate state directly — nothing tracks what was done, why, or whether it was authorized

✓ Every agent action runs through the Policy Layer. Episodes are bounded, attributed, and replayable

SimCapsules

Analogous to Versioned procedure specification with explicit constraint schema

A SimCapsule is a governed procedure specification. It names which tools may be used, which knowledge sources must be retrieved, which memory namespaces are in scope, what outputs are permitted, and when escalation is required. Versioned and governed — it only activates when the execution profile matches and governance clears. The procedure is the artifact. Evidence follows execution.

✗ Behavioral patterns live in prompts — unversioned, ungoverned, and indistinguishable from authorization

✓ Explicit constraint schema: permitted tools, retrieval sources, memory scope, output bounds, escalation conditions — governed at admission

Next step

The layer your system is missing.

Permanent governed infrastructure that sits outside the model — admission boundaries, decision lineage, and auditable state transitions built in from the start.

See it in practice →

Admission boundary

Every input policy-checked before it enters governed state

Decision lineage

Every decision attributed, linked, and replayable

Auditable ledger

Every state transition committed, exportable as evidence