Where the methods are tested before they become advice.
SaltationAI is two things at once: a consulting practice and a small lab. The lab is where I work out the methods, formats, and experiments that — when they hold up — become the practice. Some stay in the lab.
Nothing here is a product. Nothing is for sale. It's public so the thinking can be examined — including what hasn't worked yet.
The lab has one agenda with two programs: decision simulation you can audit (Possible Futures), and knowledge that travels between agents (Doxa). Everything else is notebook.
Project · 01
Possible Futures
A multi-actor simulation engine for mapping what could happen next.
Possible Futures runs a decision or a situation forward many times, with each actor modeled as a distinct AI agent with its own incentives, constraints, and tolerances. Every step — every argument, every adjudication, every outcome — is written in plain language and auditable after the fact. The engine doesn't predict the future. It maps the space of plausible ones, and surfaces the dynamics that hold across most of them.
A single run is a story. Across a hundred-plus runs, patterns separate: the dynamics that appear in nearly every run are load-bearing — they're what your strategy has to plan for. The rare ones are where to look for leverage — places where a small change in framing may produce a different outcome.
The kinds of situations the method is built for:
- A nonprofit board considering a transformational grant with restrictive conditions — how the executive director, board chair, program staff, and major donors actually respond once the strings are made explicit.
- A leadership team planning a public reorganization — which teams, peer organizations, and external stakeholders move first, and how the rollout sequence changes who feels blindsided.
- A coalition of organizations negotiating a shared policy ask — which combinations stay aligned under different framings, and which fracture.
- A research team preparing to publish a contested finding — anticipating how different professional communities will receive it, and which framings open or close the conversation.
A nonprofit board considers a transformational grant with restrictive conditions. Who moves first, and how do the strings change the room?
- Executive Director
- Board Chair
- Program Staff
- Major Donors
- Donors raise the conditions before the ED does. Every run.
- Program staff align with the chair, not the ED, when strings tighten.
- A quiet second offer reframes the room when it arrives early.
One finding from the validation work: I rebuilt a documented War College crisis scenario and ran it 126 times across three models. Most runs came back messy, like the real thing — no clean winner. But the 42 runs I matched to the experts' exact table broke decisively about as often each way: ISIL crushed in roughly one run in five, triumphant in roughly one in five, around a large undecided middle. The experts' own documented outcome showed up in five of those 42 — reachable, but one path among many the same crisis can reach. Mapping that range — the common and the rare, side by side — is what running it many times is for.
Status: validated on one documented scenario against a human-played reference; cross-scenario generalization is the open question we're testing next. We're preparing the first client-facing runs now.
Project · 02
Doxa
A portable, AI-native knowledge format built for large corpus work.
What does it mean to share knowledge in an AI-native way? Most documents are written for humans to read top-to-bottom. AI agents need something different: structured claims, relationships, provenance, and instructions for how to read the rest.
A .doxa file is a single SQLite database containing
structured knowledge — concepts, relationships, provenance, confidence scores, semantic
search — alongside a _guidance table that teaches any AI
agent how to navigate that specific doxa at runtime. The broader system is an open
spec, a universal agent skill, and an MCP server. Doxa began as a way to make a
long-running personal research corpus portable — knowledge an agent could be taught to
read. Concordance and Possible Futures need the same layer. The larger question is
social: how does research compound across collaborators when citations and provenance
travel with the data?
- Claude · MCP
- ChatGPT · skill
- Local agent
- Cursor
- concepts id · term · gloss · domain
- relations subject · predicate · object · confidence
- provenance source · author · retrieved_at · license
- embeddings concept_id · vector · model
- _guidance when_to_query · how_to_navigate
sqlite · FTS · vec · open schema
_guidance table teaches each agent how to traverse the rest.
Status: the spec and tooling are tested; the open question is whether the guidance layer makes a generic agent behave like a domain expert. That test hasn't run yet.
From the notebook
Concordance — rebuilding the scholar's concordance as an AI-native instrument: linguistic fingerprints over the public-domain corpus. A working analysis pipeline over a small corpus; an exploration surface in progress. Early; notes to come.
The Lab
Why the lab exists
I've kept a studio-and-lab practice for more than twenty years, across architecture, public art, community organizing, and affordable housing. The lab side has always had the same shape: a place where ideas are tested against real material before they become advice to clients. AI is the current medium.
The lab is not separate from the consulting. It's where the consulting is kept honest. When a framework I teach holds up against a new situation in the lab, it earns its place in the practice. When it doesn't, I learn that here, not in front of a client.
Things in the lab are not products. Some become methodology used in client work. Some stay in the lab. A few may eventually become open-source releases.
Get in touch
Work with the lab
If your situation fits the shape of one of these experiments — a decision with many actors, a corpus that wants a new way to be read, a body of knowledge that should travel between agents — I'd be glad to talk. The lab is not a consulting product. It's a conversation, sometimes a collaboration.