PegasusSource
Layer 04 The membrane, in depth

What you optimize for is
what you'll answer for.

The membrane is the layer where an agentic system's decisions are authenticated, permissioned, and recorded — the boundary between what a system is free to do and what it must account for. What a system optimizes for is a decision like any other: it has to be specified, logged, and defensible in advance, not reconstructed after something goes wrong.

Reggie Britt·Pegasus Source·≈ 5 minute read
What optimization hides

An agent optimizes for something. Most systems can't say what.

Every autonomous system is driven by an objective — a number it is built to move. But the objective a system was given and the behavior it actually learns are rarely the same thing. It finds the shortest path to the metric, and that path is discovered in operation, not written down anywhere you can inspect. Ask most teams to produce the record of what their system optimized for, run by run, and there is no record — only the dashboard, which shows the number went up, not what the system did to move it.

That gap is not a moral failure; it is a structural one. The target is easy to measure and the behavior behind it is not, so teams instrument what they can see and infer the rest. The distance between the number moved and the system did what we intended stays invisible — until a drift, an incident, or a review forces someone to reconstruct it after the fact, from logs that were never built to answer the question.

What a system optimizes for is a decision. Decisions get recorded.

The membrane, defined

A system you can't account for is a system you don't control.

The membrane is the layer that closes that gap. It sits over every consequential action an agent takes and does three things without exception: it authenticates what is acting, permissions what it may do, and records what it did and why. Governance is not a list of prohibitions bolted on after the build — you cannot enumerate everything a capable system might do. What you can do is make every crossing accountable: a positive definition of what is allowed, an immutable record of what happened, and a gate that holds before anything ships.

Applied to optimization, the membrane makes the objective itself accountable. The target a system pursues stops being an implicit property discovered in production and becomes an inspectable, recorded decision — specified before the run, logged during it, and defensible after. Not defensible to a court. Defensible to your board, your next incident review, and yourself.

The answer · at the membrane

Three disciplines. Three things the membrane must hold.

Making an objective accountable is not a matter of intent — it is a matter of architecture. Three mechanisms turn "know what your system optimizes for" from a principle into something the build enforces. Each discipline is a thing you build.

The discipline · 01

Record what governed each run

Don't infer the objective from the dashboard after the fact. Capture what actually drove each consequential decision, as it happens.

What the membrane must hold → the why-layer

A decision trace on every run

Accountability needs a record that optimization targets are inspectable, not implicit. The DECISION_TRACE_SCHEMA makes each consequential run emit an immutable, signed record of what governed it. What drove the outcome stops being a quarterly guess and becomes a queryable projection over the log — the same immutable, correlation-ID log the membrane already keeps.

Rule Applied Exception + Justification Inputs Approver Precedent
The discipline · 02

Measure what you actually intended

Track whether the system is achieving what you meant it to — not just whether the proxy metric it was given is moving in the right direction.

What the membrane must hold → the gauge

The instrument, on the Four Pillars

This is the gauge the dashboard structurally lacks, wired into the independent GOVERN / ASSURE layer: Trusted Evals that benchmark for drift and foreseeable failure, not just accuracy; Searchable Logs as the substrate; a Human Review Queue with a named owner; Granular Rollback when a signal fires. A drift or harm signal routes direct to GOVERN by an independent path — the way a fraud signal does.

Trusted Evals Searchable Logs Human Review Queue Granular Rollback
The discipline · 03

Name what each behavior is for

Identify what a mechanic is meant to do and what it's allowed to do — and decide, before it ships, which behaviors you can actually stand behind.

What the membrane must hold → the contract

A spec per behavior, and a gate that can halt the build

Naming what a behavior is for is the AGENT_SPEC_SCHEMA discipline, applied to any mechanic: no spec, no ship. Purpose, Autonomy Tier, and Permission Envelope bound what it may do. The test is the Readiness Gate: a behavior you cannot account for is a Red, and a Red halts the build. Can you stand behind this becomes a gate the architecture enforces, not a question asked too late.

Purpose Autonomy Tier Permission Envelope Escalation Audit Trail
Links to ExO 3.0 · the Four Pillars of GOVERN/ASSURE

These three disciplines are the membrane doing its core job. The record, the gauge, and the gate are the production form of the Four Pillars of GOVERN/ASSURE in ExO 3.0 — the layer that proves what an agent did, stops what it shouldn't, and holds why it optimized the way it did. A rule states what should happen in general; the decision trace records what happened in this case, and why. They are not a private framework: each pillar operationalizes controls in NIST AI RMF, the OWASP LLM Top 10, and the CSA AI Controls Matrix. The full method →

Optimization targets are accountable decisions. Specify them, log them, and gate them — before the system runs, not after it drifts.

Our stance: a system you can account for is a system you control. The record that lets you answer for what a system did is the same record that keeps it honest while it runs.

The membrane, restated

What a system optimizes for is a decision — so it has to be held like one.

The principle and the architecture are one claim in two languages. Know what your system optimizes for is the plain-language version of what decision traces, the gauge, and specs operationalize: what a system is built to maximize carries accountability, so it has to be written down, recorded, and answerable in advance — whether the behavior lives in a workflow, a checkout flow, a pricing loop, or an autonomous agent.

That is why this layer is a membrane and not a wall. A wall blocks; a membrane lets action through while holding a record of it. The instrument that lets you answer for what a system did is the same instrument that keeps it aligned to intent while it runs. Build it before you need it, and the audit, the eval, and the answer are already written.

The fix is whatever gets you there — your own team, another partner, or us. What we sell is the reading and the depth, not a dependency.

Before the conversation — the depth

Every engagement starts the same way: a reading of where you actually stand — here, whether your systems can say what they optimize for, and whether the record to prove it exists. A direct conversation about the decision in front of you — no obligation, and both sides will know quickly if it's a fit.

Book a call ↗ Or send a message →
← Back to the layer in context
Layer 04, inside the whole surface
Where the membrane sits among the five, and what to work first.
The next layer down →
05 · The sovereign core
What happens when an agent is already inside — reached, never rewritten.

Own your source. Own your surface.

The membrane, in depth · continued