On reading the engineered systems of power.

States, source code, and the gap between the people who report and the people who build.

Forward Indicators Team · May 2026

Most analysis of state power treats the state as a black box. A small group of officials meets behind a curtain, deliberates, and emits a press release; the rest of us are asked to infer intent from the release and the body language of the people delivering it. The implicit advice this frame produces is always the same: collect rumors, count tea leaves, hire a former diplomat.

That advice is convenient for the people selling it. It is also a methodological dead end. It treats every surprise as an act of will rather than as the predictable output of a system whose specifications were published years earlier. It rewards access over evidence. It produces narratives that are unfalsifiable by design, because no one remembers what last quarter's briefing predicted.

The cost of this frame is not academic. It compounds every time a regulator is caught flat-footed, every time a board reads a sanction list as a shock instead of as a slow-cooked policy, every time a venture portfolio is restructured at 11pm because something “came out of nowhere.” Almost nothing comes out of nowhere. Almost everything was on a timetable.

There is a better frame, and it is available to anyone willing to read code instead of cables.¹

Modern power leaves a source trail. Doctrine is a specification. Five-year plans are architecture diagrams. Regulations are interface contracts. Procurement schedules are build pipelines. Subsidy cascades are reference implementations. Talent restrictions are access-control lists. Each of these is published, dated, and indexed; each one constrains what the system can do next.

The reason this trail is rarely read as a system is that the people best equipped to read code are not the people writing about policy, and the people writing about policy were trained to read motive rather than mechanism. The professional incentives reward narrative continuity. The mechanism does not care about narrative continuity. It cares about the inputs it was specified to compile.

Treating a state as an engineered system does not mean denying agency, ideology, or surprise. Engineered systems have bugs, exceptions, and emergent behavior. So does any production codebase. The point is that the surprises are bounded by the architecture. If you know the architecture, you can describe the surprise space before the surprise arrives — and you can put a probability on each region of that space.

This is not new. It is how analysts read corporate earnings, how epidemiologists read genomes, how engineers read other engineers' commits. The novelty is applying the same reading to the artifacts of state power, in public, with the model's reasoning shown.

Forward Indicators runs the same pipeline on every question. Five steps, one architecture, different corpora.

Ingest. Every doctrinal document, regulation, ministerial readout, sanction list, and procurement notice goes into a versioned corpus. Provenance is preserved. Translations are stored beside originals.²

Encode. The analytical prior — system prompts, few-shot examples, feature definitions — is versioned and viewable. When the prior changes, the diff is logged. You can read why the model says what it says, and you can read what it used to say last month.

Predict. Every output is structured: a dated claim, a confidence, a falsification criterion, a citation pack. No prediction is published without the evidence it was conditioned on.

Register. Predictions land in an append-only log. No retroactive edits. When a prediction is wrong, the wrongness is permanent and visible.

Evaluate. Brier score, calibration plot, lift over baseline. Three baselines run on every scoreboard: random, last-observed, and feature-only logistic regression. If the language model is not adding lift over the simple regressor, the scoreboard says so before we do.

On top of the five-step spine, an auto-research loop runs weekly: it proposes new features, tests them against held-out historical data, and updates the prior only when calibration improves. The loop is the laboratory. Everything else is plumbing.

China is where the method bites first, for reasons of corpus density rather than ideology. The doctrine is the most explicit. The cascade from plan to ministry to subsidy to firm is the longest and best-documented. The policy machine is the most legible to anyone willing to read the original-language sources at scale.³ Decades of five-year plans, work reports, and central committee plenums constitute the richest publicly available training set for a model of how a modern state translates doctrine into deployment.

It is also where the gap between the engineering frame and the prevailing analytical frame is widest. The state is run by engineers; it is read by everyone else. The May 2026 reporting on travel curbs for Chinese AI researchers was a representative case: the policy direction had been signaled in formal documents for over a year, but most coverage treated the news as a surprise.⁴ The signal was in the specification. The surprise was a reading failure.

The method, however, is not specific to one country. The same spine runs on US semiconductor export controls and their cascading licensing exceptions. It runs on the EU AI Act's enforcement schedule and the gap between statutory text and member-state implementation. It runs on Gulf sovereign technology spend and the talent-flow patterns that follow it. Wherever there is a published specification and a downstream cascade, there is a pipeline to be built. The lab will broaden as the spine hardens.

Four operating principles. None are aspirational; each one is enforced by the pipeline itself.

Predictions are dated. Every claim is logged with a date, a confidence, and a falsification criterion. The log is append-only.

The prior is public. System prompts, few-shot examples, and feature definitions are versioned and viewable. When the prior changes, the diff is in the log.

Baselines stay shipped. Random, last-observed, and feature-only regression run alongside the language model on every scoreboard. Lift over baseline is shown, not assumed.

Calibration over accuracy. A seventy-percent prediction must be right seventy percent of the time. Not ninety. Not fifty. The reliability diagram, not the point accuracy, is the headline chart.

The work is open by default and slow on purpose. We would rather publish three calibrated predictions a quarter than thirty uncalibrated ones a week. We would rather be wrong in writing than right in private.

If you work on geopolitics, AI policy, or US–China tech relations and want to follow along, the research log is where the loop is visible. Write in.