My methods page lays out a four-tier standard for claims: demonstrated (shown, with an artifact you can open), plausible (built or designed for it, not yet proven), aspirational (an intention, stated as one), and unbuilt (named so it is never mistaken for something real). A standard is easy to publish and cheap to praise. The test is whether I will apply it to my own work, in public, where it costs something. So here it is, turned on the three projects in my lab — three projects, four tiers each.
Possible Futures
Possible Futures is the decision-simulation program in my lab. It runs a real decision forward many times, every stakeholder modeled as a separate AI agent arguing its own case, and every move, ruling, and outcome written in plain language you can read back afterward — auditable by construction, not by audit.
Demonstrated. I rebuilt one documented crisis — the US Army War College’s 2015 ISIS Crisis game, played in the open by expert teams — and ran it 126 times across three language models, at about four cents a game. Most runs came back as messy as the real conflict: no clean winner. Among the runs configured to match the experts’ exact setup, the decisive ones split almost symmetrically — roughly one in five drove ISIL to collapse, one in five let it win outright — and the experts’ full documented outcome turned up in about 12% of them: present, reachable, and rare. The whole run of every game is on the record.
Plausible. It is built to map the space of how a multi-actor decision could go — the dynamics that hold across most runs, which are what a strategy has to survive, and the rare ones, where leverage might live. I have shown that on one scenario; the design is for many.
Aspirational. I am preparing the first client-facing runs now, written up the same way this one was — including the parts that do not work.
Unbuilt. Whether the method holds on a second crisis is untested — one validated case is a rigorous result, not a validated system. Whether a real decision has ever changed because of a run: zero, so far. A finished causal account of what actually drives each outcome: not done. And Possible Futures is not a product — I keep an internal bar for product-readiness, and the program is deliberately still behind it, which is why nothing in the lab is for sale.
Doxa
Doxa is a file format built to make one expert’s reading judgment travel with their knowledge — a single small database that carries not just the documents but the instructions for how to read them.
Demonstrated. The format and its tooling exist and are tested: an open specification, an agent skill, a server, and a test suite that runs green — 207 checks collected, 192 passing, the rest waiting on an API key. At its center is a guidance table: the methodology, the order to consult things in, and the known limits of one specific body of knowledge, written by the person who built it.
Plausible. It is designed so that a generic agent, handed a doxa it has never seen, reads the knowledge the way the author would — same priorities, same cautions, same sense of where the edges are. That is the design intent.
Aspirational. The near-term plan is to test precisely that, and to build a second doxa in a different domain, by a different author. I will publish what happens, including if the answer is no.
Unbuilt. The test that matters has not run: hand an agent a doxa cold — no coaching, no custom prompt — and check whether the guidance layer makes it perform like a domain specialist instead of a search box. Until it does, “teaches agents how to read” is a design claim, not a demonstrated one. And only one substantial doxa exists so far, built by the same person who wrote the spec; a format proves itself on the second author, not the first.
Concordance
Concordance is the quietest of the three: an attempt to rebuild the scholar’s concordance as an AI-native instrument — linguistic fingerprints over a public-domain corpus.
Demonstrated. A working analysis pipeline over a small corpus, with a tested backend.
Plausible. Built toward an instrument for reading a corpus structurally — across many books at once — rather than one volume at a time.
Aspirational. Notes to come.
Unbuilt. There is no deployed surface yet, the corpus is small, and not one finding has been written from it. Until those three exist, Concordance is a notebook entry, not a program — and no claim on this site rides on it.
What this doesn’t show
This note tiers my claims; it does not prove the work is useful to you. The tiers are honest as of the date at the top of the page — and they are meant to move. Demonstrated should grow; unbuilt should shrink; the day one of these crosses a line, the claim crosses with it, in public.
And the one item sitting in unbuilt for all three projects is the same item: none of this has yet changed a real decision for a real person. That is the test I care about most, and the one I have not passed. When I pass it, that will be a note here. If I try and fail, that will be a note too.
If you are holding a decision with several actors and a date on it, the Brief is where that conversation starts.