Methods

The AI can write the code. It doesn't get to write the audit trail.

PortaStack is the governance kernel I built to enforce that line. Authority, evidence, verification, and approval stay explicit and auditable, no matter how much of the work an AI actually did. It's the control architecture underneath every project on this site.

Why this exists

AI-assisted development has a trust problem. An agent can write the diff, run the tests, and report that it's done, and there's no structural reason to believe the summary over the raw evidence. PortaStack closes that gap with nine deterministic invariants.

The nine invariants
01Orchestration
Stops an agent from freelancing outside an approved workflow.
02Scoped execution
Stops an executor from touching anything outside its declared domain.
03Evidence recording
Stops an action from happening with no append-only, hash-chained record.
04Independent review
Stops a reviewer from trusting the executor's own summary instead of raw evidence.
05Tiered context
Stops a model from quietly promoting an assumption into settled fact.
06Policy gates
Stops a material change from landing without an explicit check, approval, or evidence.
07Capability qualification
Stops a model from doing work it has never been evaluated for.
08Reproducibility
Stops a result that nobody else can reproduce from the recorded inputs.
09Rollback
Stops write authority from being irreversible. It can always be revoked back to observation-only.

No industry assumption, no certification regime, no customer baked in. It's a control architecture that gets mapped onto whatever governance model is already in place.

Macro photo of a BIOS chip and eSPI traces
The shape of it

An orchestrator routes work to a scoped executor. The executor writes to an append-only evidence journal. A reviewer reads that journal directly and never sees the executor's summary. A policy gate checks for both evidence and approval before anything releases.

routes to writes read independently checks evidence + approval missing either disposition Tiered Context Working → Long-lived → Canonical Orchestrator Scoped Executor Evidence Journal append-only · hash-chained Reviewer Policy Gate Release / Merge Blocked

The reviewer never receives a summary from the executor. It reads the same append-only journal independently. That single design choice is what turns "the agent says it passed" into something you can actually check.

Maps onto any stack

PortaStack doesn't assume an orchestrator, a language, or a toolchain. Whatever a system already has, it maps onto the same control-plane roles:

Your system has...
...maps to this control-plane role
An orchestrator profile
Workflow controller
Specialist / domain agents
Scoped executors
Something recording what happened
Append-only audit recorder
Something checking that record independently
Evidence reviewer
Session vs. project vs. approved-forever knowledge
Working / Long-lived / Canonical context
Merge / release gates
Policy-as-code workflow gates
Which model does which job
Qualified capability profiles
Where it runs
Developer workstation
IDE/CLI calls the orchestrator, read-only by default. Production credentials never touch the model.
Orchestration service
Authenticates through the real identity provider, resolves policy and capability qualification, stamps every operation with a correlation ID.
Execution runners
Ephemeral, least-privilege, network-restricted, split by class: source, verification, infra, release.
Evidence plane
Append-only, immutable retention for anything material. Reviewer identity stays separate from executor identity, always.
Context plane
Working (session) → Long-lived (repo knowledge) → Canonical (human-approved, never AI-promoted).
Rolling it out without betting the farm
01
Observe only
Read-only, just recording decisions and evidence.
02
PR authoring
Bounded code changes through normal review.
03
Controlled execution
Approved build/test actions in isolated runners.
04
Release preparation
Evidence and manifests assembled. Existing release authority still signs off.
05
Deeper integration
More permission, only after the mapping earns it.

A control plane that can tell you what actually happened.