The Cognitive Substrate
Surfaces 8 and 9 together form a substrate through which an organization remembers what its agents decided and why — the layer that compounds.
Surfaces 8 and 9 are not "two more fields on the contract." Together they constitute a new category of organizational infrastructure: a substrate through which an organization remembers what its agents decided and why.
What the Cognitive Substrate Is
The agent contract specifies nine surfaces (see Agent Contracts). Surfaces 1-7 govern operational behavior. Surfaces 8 (Confidence) and 9 (Decision Trace) together form a cognitive substrate — the infrastructure through which agent decisions become observable, replayable, and pattern-detectable across the entire organization, indefinitely.
Conceptually, the cognitive substrate is the durable record of three things:
- What the agent decided — the structured output
- How confident the agent was — computed, not generated; defensible by replay
- How the agent arrived at its decision — the trace, including inputs, tools invoked, reasoning steps, resource consumption, version information
The substrate is not the agent. The substrate is the record the agent leaves behind.
Organizational Cognition Lineage
Organizational cognition lineage: the durable record of decisions an organization's agents made, why they made them, and how that reasoning evolved over time.
Traditional organizations do not have this. Human decisions are recorded in fragments — emails, meeting notes, design docs — most of which lose specificity within weeks. The cognitive substrate inverts this: every agent decision produces a structured record automatically. No human has to remember to document. The trace is mandatory, structured, replayable, and durable.
What This Substrate Enables
Five operations become possible that were not possible before:
- Replay — given an input hash and version information, an invocation can be replayed; deterministic agents reproduce the same output, stochastic agents a measurably similar one.
- Audit — every tool invocation is recorded with input, output, duration, and cost; contested decisions are resolved with evidence, not memory.
- Pattern Detection — the Inspector Pipeline reads trace metadata across cells to detect better prompts, reliable tool compositions, and calibration drift.
- Calibration — declared confidence is compared against actual outcomes to measure and refine agent self-assessment.
- Lineage Reconstruction — the current state of any capability or pattern can be traced back through the decisions that produced it.
Why This Is Different From "Logging"
| Logs | Cognitive Substrate |
|---|---|
| Optional, frequently sampled | Mandatory at L2+, no sampling permitted |
| Free-form text | Structured schema enforced by kernel |
| Service-level | Per-invocation, per-decision |
| Read mostly for debugging | Read for replay, audit, pattern detection |
| Retention measured in days to weeks | Retention measured in years |
| Privacy implicit | Privacy structural (metadata accessible, content gated) |
| No replay guarantee | Replay-by-input-hash is a contract |
Logs are an artifact of execution. The cognitive substrate is infrastructure for organizational memory. The first is operational; the second is architectural.
Why "Computed, Not Generated" Matters
Surface 8 (Confidence) requires that confidence values be computed from observable signals, not generated by asking the model. This rule is what makes the cognitive substrate honest. If confidence were generated, replay would be meaningless and pattern detection would be detecting hallucination patterns, not real confidence patterns. The Butler's gating, the Inspector Pipeline's calibration analysis — all of it depends on confidence values being defensible by replay.
The Decision Trace as Engineering Artifact
A decision trace is the closest agentic-systems equivalent to an architectural decision record (ADR) — but produced per invocation, automatically, at scale. Where ADRs are written by humans for major decisions and frequently neglected for minor ones, decision traces are written for every decision, large or small, with no human intervention required.
Without the substrate, agentic systems produce output that ages into mystery within weeks. With the substrate, agentic systems produce a record of their own reasoning that becomes more valuable with time.
Closing
The kernel doesn't run the cells. It defines the contract they all honor. And the most important thing the contract defines is the substrate through which the organization will remember what its agents did and why. This is the layer that compounds.