the ghostie stack

AI you own, that remembers, stays local, and proves its own work.

Today's AI rents you a memory you can't read, sends your code to someone else's servers, and asks you to trust an answer it can't justify. The ghostie stack is built the other way: small, local-first tools that keep your context in a store you own, run their deterministic core on your own machine, and verify their own output.

local-first architectureairgap-capable componentsdeterministic verification provider-agnostic memoryuser-owned context

what it is

One persistent stack, not a dozen disconnected sessions.

Small tools that each do one thing well and snap together into a loop: understand the code, render what's known, remember it, act on it, prove the result. Small parts that compose, never one monolith.

context compounds with every pass TrailTrackerunderstand Ontoloomrender ghostieremember Nightshiftact Lockstepprove

A codebase enters, and TrailTracker builds its structural model.

ghostie
memory & continuity

Provider-agnostic memory you own: your context persists across sessions, agents, and devices, in a store you can read and carry.

shipped · public
TrailTracker
the understand node

Builds a codebase's structural model on your own machine, the offline first step the rest of the loop reads from.

built · private
Ontoloom
knowledge rendering

An airgapped GUI that renders those models as navigable graphs and lets a human author ontologies on top.

shipped · public
Nightshift
verified migration

The act node: a behavior-preserving migration orchestrator. It freezes a behavioral contract, drives the candidate through Lockstep, and refuses rather than guess when it cannot verify.

shipped · public
Lockstep
verification

Runs a program and its rewrite over an explicit corpus and issues replayable, tamper-evident evidence that they behave equivalently at the observable boundary, naming exactly what it could not prove.

shipped · public
Auftragstaktik
the standard

The shared contract and health checks every tool follows, so the pieces stay small, composable, and honest as the stack grows.

shipped

how the pieces compose

The loop, stated plainly.

The whole loop runs end to end today on one example. Scaling it past that example is the ongoing work.

  1. TrailTracker reads a codebase locally and builds the structural model the loop works from.
  2. Ontoloom lets an engineer inspect that model and correct where the machine got it wrong.
  3. ghostie preserves the assumptions and decisions, so they outlive the project.
  4. Nightshift orchestrates the migration and hands both implementations to verification.
  5. Lockstep runs both against the same corpus and issues a replayable certificate.

The loop runs end to end today on a first example, proven in public CI: Nightshift freezes the contract, hands the candidate to Lockstep, and gets back a replayable EQUIVALENT certificate. At this milestone the candidate is supplied, so what is proven is the wiring, not regeneration quality. General, agent-driven regeneration is the next step.


what it brings

Three things hosted AI doesn't let you fully own.

Hosted AI will remember you, run your code, and hand you answers. What it won't do is let you own any of it: the memory is rented and opaque, locked to one vendor. This stack makes your context owned, inspectable, and portable, and holds the rest of the loop to the same standard.

01

You own the memory.provider-agnostic

Hosted assistants do remember you now. But that memory is rented: opaque, unexportable, locked to one vendor's product. The ghostie stack keeps your context in a store you own and can read, portable across sessions, across agents, and across devices. What one agent learns, the next already knows, whoever's model it is.

02

It's sovereign where it counts.airgap-capable

The deterministic core, TrailTracker, Ontoloom, and Lockstep, runs entirely local: zero external dependencies, no telemetry, working with the network unplugged. The memory layer talks to whatever model you choose, so the claim stays scoped and honest; the parts that have to be trustworthy are the parts that never phone home.

03

It proves its work.anti-black-box

It doesn't ask you to trust its confidence. It runs deterministic analysis, issues self-verifying certificates, and makes a test fail before it trusts one. The output comes with evidence attached, and it stays honest about what it cannot prove instead of bluffing past it.


shipped

Working pieces you can read right now.

The repositories linked below are real. Clone them, read the tests, read the gate. The deterministic tools carry zero external dependencies and are gated in CI.

Lockstep

public

Behavioral-equivalence verification for code migration: runs the old and new programs over the same inputs, compares them at the observable boundary, and issues a deterministic, hash-chained certificate that includes what it could not prove.

Rust · std-only, zero deps · #![forbid(unsafe)] · 75 tests · self-audited

Ontoloom

public

An airgapped, dependency-free ontology & knowledge-graph builder with a visual browser GUI served on localhost, exporting Neo4j-compatible formats. No cloud, no telemetry, no packages.

Rust · std-only · browser SPA on loopback · Neo4j-compatible exports

ghostie

public

The memory layer: provider-agnostic context you own, persisting across sessions, agents, and devices in a store you can read and carry, whichever vendor's model sits on top.

Rust · std-only, zero deps · own-your-memory store + deterministic recall with why-explanations · 287 tests

gatekit

public

The stack's verification discipline, packaged for any project: one script is the gate, CI is a thin wrapper, and the gate dogfoods the product on a real workload with byte-stability and tamper checks. It is the working implementation of the discipline Auftragstaktik, the stack's standard, requires.

self-dogfooding gate · byte-stability + tamper checks · green on a fresh clone

Nightshift

public

The act node: a behavior-preserving migration orchestrator. It freezes a behavioral contract from the program, drives a candidate through Lockstep, and refuses-and-flags rather than guess when it cannot verify. The one impure step, regeneration, is boxed by deterministic trust on both sides.

Rust · std-only scaffolding · the loop closes end to end on a first example, proven in public CI · candidate supplied at this milestone

TrailTracker

private

The stack's understand node: builds a codebase's structural model, fully offline. The mature private engine the rest of the loop reads from.

Rust · std-only · fully offline
walkthrough available on request

Honest status: the stack is modular and already useful before every component is complete. Lockstep, Ontoloom, gatekit, ghostie, and Nightshift are public and working; TrailTracker is the mature private engine underneath. The loop closes end to end on a first example today; what remains is scaling regeneration past it. The vision on this page is the direction, and these are the parts already standing.


who's building it

Jonathan Popham

Systems and AI engineer. I was the founding engineer at a code-intelligence startup, hired pre-product. I helped build the engineering team and designed the parsing, classification, LLM-routing, and evaluation stack that carried it to production.

My lane is deterministic understanding and verification: code that has to be right, and provably so. I work in Rust when correctness matters (zero-dependency, forbid(unsafe), byte-stable), and across the stack when it doesn't. I'm most useful where AI throughput meets senior-engineer judgment, when someone has to make the machine's output trustworthy before it ships.

The ghostie stack is where that conviction lives. It's a bet that the durable form of AI tooling will be owned, persistent, and self-verifying, rather than rented, vendor-locked, and opaque.

focusUnderstanding · evals · agentic infra
languagesRust, Python, TypeScript, Go
educationB.S. Mathematics

work with me

I build deterministic infrastructure around probabilistic systems. If that's the problem your team is facing, let's talk.

jonathan.popham@gmail.com