Method

How it's built

These studies are produced by a multi-agent AI system, engineered so that it is safe to trust. The architecture below is what makes it so.

For a publication about epistemic integrity, that disclosure isn't a liability — it's the demonstration. The interesting question was never whether to use these systems. It was whether one could be built whose primary design constraint is that it must not be able to quietly deceive you — including about itself.

This page is the honest answer, including the parts that are unglamorous: what the machine enforces, what only discipline enforces, and how you can check either one.

The architecture

A table of agents, and a human who decides

The seats at this roundtable are not a metaphor for disciplinary perspectives. They are agent roles — each instantiated from its own written role specification, each with a defined scope, its own methodological toolkit, and explicit limits on what it may claim. Nine agent roles are specified — an Author and eight expert seats — alongside a reusable template from which new seats are instantiated. The Editor is human. A study convenes only the seats it needs; a seat that did not sit is not credited.

01 / EDITOR

The human

Convenes the table, chooses the study, sets the priors, and decides what publishes. Every contested judgment terminates here, not in an agent.

Human — not an agent
02 / AUTHOR

Builds and integrates

Constructs the model, the analysis, and the reader-facing surfaces. Reproduces every external result before it enters a study. Explicitly may not referee its own work.

Agent role · own specification
03 / EXPERT SEATS

An adversarial bench

Domain authorities convened per study, instantiated from a shared template. The first is an Economics & Statistics Referee whose job is to try to fail the study.

Agent roles · rotating

The separation is the point. An agent that builds a result does not get to bless it, and no agent decides what publishes.

The bench, in full

A seat owns a lens, not a topic. Most studies convene several; the reference study convened exactly one, and says so.

RoleOwns
Author / CataloguerBuilds the model, the analysis, and the reader-facing surfaces; reproduces every external result before integrating it; keeps the library consistent. May not referee its own work.
Economics & Statistics RefereeIncidence, distribution, uncertainty, causal design. The adversarial seat. Sat on the reference study
Energy & Biophysical EconomicsNet energy and EROI; the physical substrate beneath the money.
Earth-Systems & Ecological ScienceClimate, biodiversity, biogeochemical cycles, overshoot.
Complex-Systems & PolycrisisCoupling, cascades, tipping points, resilience.
Institutions & CommonsCollective action and governance, in the Ostrom lineage.
Cognitive Science & Media EcologyCollective-cognition failure; the attention economy; information integrity.
Game Theory & Mechanism DesignThe mathematics of multipolar traps, and the design of escapes.
Technology & AI GovernanceExponential capability against the wisdom and governance to hold it.

Eight of these nine were written after the reference study was complete. They have not yet sat at a table — and until they do, they are not claimed as contributors to anything.

The pipeline

Question Source & model Simulate (seed + raw draws) Reproduce independently Build surfaces Referee Propagate to every artifact Publish

The rule that does the most work is reproduce before you integrate. No external result enters a study until it has been recomputed from the raw draws and independently reimplemented — different code, different random seed. If a fresh implementation on a fresh seed diverges, the result was a coding accident or a seed artifact, and it does not ship.

On the first study, the Referee seat rebuilt the Monte Carlo engine from the model parameters in its own code on seed 99, and the coupling engine again on seed 7, against an original seeded 20260605. The headline tail reproduced. That check is why the numbers are published at all.

What is published, always

Every study ships its floor: the working paper, the live formula-driven model, the simulation engines, the recorded statistics, and the raw draw vectors themselves — all 200,000 of them, so any reader can recompute any percentile without trusting ours. The published commitment is blunt: if a number in the report and a number in these files disagree, the files win — and we want to know.

Uncertainty is reported as two bands, never one false-precision interval: a historical-volatility floor, and a coupled band that lets channels fail together. The gap between them is the deliverable. And a published study is immutable — never rewritten, never quietly corrected. Everything learned afterward is appended, dated, and layered around it.

The distinction that matters

What the machine enforces, and what only discipline enforces

Any system can claim rigor. The question an engineer should ask is which guarantees are structural and which are merely intended. Blurring that line would be exactly the error this project exists to catch — so here is the split, plainly.

Enforced by mechanism
Holds whether or not anyone is paying attention.
  • The random seed is fixed in the engine source; the engine writes its raw draw vectors on every run. Reproduction is not a promise — it is a file.
  • Every reported statistic is recorded to a results file at generation time, not transcribed by hand afterward.
  • The git repository is the system of record. Every published change is a commit; nothing reaches a reader that did not pass through it.
  • The corrections endpoint verifies its bot-challenge token server-side, sanitizes and length-caps every field, whitelists permitted values, restricts source URLs, escapes all output, and fails closed if misconfigured.
  • Structural checks run against the artifacts themselves before publication: scripts must parse, structured data must validate, the workbook must recalculate without errors.
Enforced by discipline
Real, and load-bearing — but not machine-guaranteed.
  • Provenance tagging. Every input is labelled measured, modelled, assumed, or design-not-yet-run. This is a standing rule, checked by review — not by a compiler.
  • Evidence tiers. Claims are tagged established / contested / exploratory. An editorial gate, not an automated one.
  • Reproduce-before-integrate. The reproduction is mechanical; the obligation to perform it is protocol.
  • Caption–data consistency and the ban on presenting a modelled number as a measured one. Enforced by an adversarial reviewer, and by a human editor who can refuse to publish.

There is no continuous-integration gate enforcing the right-hand column. It is enforced by an agent whose role is to fail the work, by a human who decides what ships, and by publishing enough raw material that a stranger can catch what both of them missed. That is a weaker guarantee than code, and saying so is the only honest way to describe it.

It is also, so far, a guarantee that has held — because it has been tested.

Evidence, not assertion

What the architecture has actually caught

A design is a hypothesis until it fails something. This one has failed three things that matter — twice against its own work.

01 — The founding catch

The first study exists because a figure in circulation was wrong. An annual inflation rate had been applied as a monthly one and compounded, overstating the household impact of an energy shock by roughly fifteen to twenty times — describing hyperinflation the United States has not seen in the post-war era. The verification pass caught it before it was built on. A number that exaggerates a real crisis does not help the people it describes; it hands ammunition to anyone who wants to dismiss them.

02 — The referee forced real changes

The Economics & Statistics seat returned Accept with revisions, then held the work through three rounds before Accept — final. It forced a single interval to be replaced by two honest bands; forced the income base of an inequality index to be disclosed rather than assumed; and caught a caption describing a pattern the data did not show — where the correct fix was to rewrite the caption, not to "fix" the data that had been right all along.

03 — It caught itself, after publishing

After the study went live, the Referee seat swept the published materials and found that one reproducibility file had been generated from a superseded parameter run — its 99th-percentile figure disagreed with every other artifact. It reported its own error. The file was regenerated, independently reproduced to byte-identical output by a second implementation, and the full raw draws were published so the discrepancy could never hide again. The correction is on the public log, against ourselves, dated. It is the first entry.

Colophon

Who built this

Designed, built, and operated by

Brandan Borgos

Attorney and software architect, Minneapolis.

The Roundtable — thethirdattractor.org — is his: the multi-agent architecture, the integrity protocols, the studies, and the site itself. The editorial framing draws on the work of Daniel Schmachtenberger and Nate Hagens; the engineering does not.

Corrections, disagreements, and caught errors are welcome and wanted — the same lens, turned inward.