Reasoning, in five layers.

Every dataset documented at its source: methodology, funding, exclusions, limitations, and the decisions that shaped what was collected and what was not.

Most data arrives stripped of its history. By the time a figure reaches a model or a policy memo, the decisions that produced it — who paid for the measurement, which population was sampled, which outcomes were judged worth recording — have been erased. What remains looks like fact. It is, more honestly, the residue of a hundred choices no one is accountable for.

Data Provenance refuses that erasure. It documents a dataset where it is made: the methodology that generated it, the funding that commissioned it, the exclusions baked into it, the limitations its own authors already knew about. Not a citation at the end of a report — a record of constitution, attached to the data itself.

The hardest field to fill is the one almost nobody fills: what was left out. Every dataset has an edge — a question not asked, a group not counted, a window cut short. Provenance makes that edge explicit, because the edge is where most distortion lives. A dataset can be flawless inside its boundary and still produce a confident, wrong answer the moment you forget the boundary was a choice.

Explicit modelling of what the data implies under multiple scenarios, with assumptions stated and uncertainties acknowledged — including the downstream effects of the exclusions documented in Layer 1.

A number implies nothing on its own. It implies something only once you decide what to assume around it — a baseline, a trend, a threshold, a counterfactual. Change the assumption and the conclusion moves, sometimes quietly, sometimes entirely. Most reasoning hides this by committing to a single set of assumptions and presenting the result as if it were the only one available.

Consequence Simulation makes the assumptions the object of attention. It models what the data implies across several scenarios, states each assumption in the open, and carries the uncertainty forward instead of rounding it away. A conclusion that survives only under one convenient set of starting points is labelled as exactly that.

Critically, it inherits the exclusions named in Layer 1 and asks what they cost. The group that was never counted does not disappear from the world — it simply bears the consequences of a decision it was invisible to. Simulating those downstream effects is how reasoning stops treating "not in the data" as "not affected."

Identification of every party affected by the reasoning, including those absent from the data, with documented analysis of how the conclusions affect them.

The people a decision harms are rarely the people in the room when it is made — and almost never the people in the dataset it rests on. Empathy Mapping is the discipline of naming them anyway.

It identifies every party affected by the reasoning, and then does the harder thing: it looks for the parties who are affected but absent — unmeasured, unmodelled, unheard — and documents how the conclusion lands on them. This is not sentiment, and it is not a consultation exercise. It is a structural check against the most common failure in data-driven decisions: optimising for the measured majority and averaging the minority away until it rounds to zero.

Done honestly, it changes what counts as a finished analysis. A conclusion is not complete because the maths closes; it is complete when the question "who does this fall on, and did anyone model that?" has an answer on the record.

A traceable record of every significant decision in the reasoning process: who made it, when, and why, with alternatives considered.

When a data-driven decision goes wrong, the investigation usually ends in a shrug. The model did what it was trained to do. The analysts followed the brief. The data said what it said. No single person decided the thing that caused the harm — because the decision was distributed across a dozen small choices nobody wrote down.

The Accountability Ledger writes them down. Every significant choice in the reasoning — a threshold set here, a population excluded there, an assumption chosen over its alternatives — is recorded with who made it, when, and why, and crucially what else was on the table. The alternatives not taken are part of the record, because a decision only means something against the options it beat.

The point is not blame. It is that reasoning becomes answerable. You can return to a conclusion months later and find the exact junctions where it could have gone differently, and the named reasons it did not.

A formal mechanism for proposing alternative assumptions and comparing what different starting points produce, so that the relationship between choices and conclusions is permanently visible.

Disagreement about data usually dies as a stalemate: two parties, two conclusions, and no shared way to see where they actually diverge. The Fork Registry turns that stalemate into something productive. Anyone can take a document, change an assumption, and re-run it — and the fork is kept, permanently linked back to the original it departs from.

The effect is that the relationship between choices and conclusions stops being a matter of authority and becomes a matter of record. You can point to the single assumption two forks disagree on, and watch exactly how much of the conclusion rides on it. Disagreement gets an address.

This is also how the project resolves disagreement about its own work. Not by vote, and not by whoever has standing — but by forking, re-running, and showing better reasoning. A conclusion holds only as long as no fork beats it in the open.