The project

About

This is an encyclopedia of the ways an AI system can be given binding norms — and of how much of each way has actually been built.

The field’s best-known answer is a written constitution: a document settled before deployment, applied to every case, and resolved into a single answer whenever its clauses pull against each other. That is one arrangement among many, and it is not obviously the right one. Political theory, jurisprudence and moral philosophy have spent centuries on the same problem and arrived at several quite different answers — precedent accumulating case by case, custom nobody ever wrote down, cultivated character, a procedure convened fresh each time. Each of those is a buildable design for a machine, and each has a literature standing behind it that the alignment field has barely read.

So the map crosses two questions — where a system’s binding norms live, and what it does when those norms conflict — into 15 regimes. Every regime carries the tradition it draws on, the approaches placed in it, and every work behind those approaches. Where a regime is nearly empty, that is the finding.

regimes
15
approaches
67
works indexed
239
working methods
93

How it is built

Everything on the site is derived from one workbook, and every number is a count of rows in it. Nothing is estimated and nothing is modelled.

The grid is a map of methods. Each work carries exactly one kind, and only a method counts toward a regime’s build-out: a mechanism given in enough operational detail to reimplement and shown working by its own authors. Both halves are required — a governance scheme is a proposal, however carefully specified, until somebody built and ran it. Of the 239 works indexed, 93 clear that bar. Benchmarks, analyses, applications, resources, proposals and precedents from other disciplines all stay indexed, searchable and attached to the approach they bear on, and count toward nothing.

Two measures, kept apart. A regime is scored on how much of it is built, and separately on how much of its borrowed tradition the built work actually carries. Collapsing the two was the site’s original mistake: once every regime had at least one working system, a single badge stopped saying anything. Kept apart, they state the map’s central finding — the two correlate negatively. Constitutionalism has a working method in every one of its approaches and carries about a fifth of Hart, Fuller and Raz.

A claim is separated from a gap. Each approach is read concept by concept against its tradition, and each reading also records whether the approach ever claimed the concept. An absent concept the authors never set out to deliver is a gap in the field; an absent concept they did claim is a shortfall they can be held to. The site never colours the first as a fault.

References are checked, and the check is structural. Whether a reference resolves is recorded per row as a flag, not inferred from a free-text note — a note saying a row was once wrongly called fabricated is not itself a fabrication flag. Every one of the 239 rows currently carries that flag.

Behind the 15 regimes sit 56 source works from law, political theory and philosophy, 73 named concepts drawn out of them, and 198 readings of an approach against one of those concepts.

Authors

To be filled in by the research team.

Version history

Changes to the corpus and to how it is read. Version numbers were assigned retrospectively — they are a reading order, not a release process.

  1. v0.55 September 2026Current

    A front page you can search, and three new ways to read the map

    • The landing page became a search box over the whole corpus; the grid moved to a page of its own.
    • Added the historical timeline: every indexed work on a year axis, one lane per regime, locus or approach.
    • Added the mind map: the taxonomy branched rather than crossed, down to the individual approaches.
    • Added the configuration finder: three questions that resolve to a regime and report what has actually been built there.
    • Redrew the built-out and adherence bars in the grid’s own green ramp, and dropped the traffic-light colours.
  2. v0.431 August 2026

    One number per cell

    • Reclassified the corpus and rebuilt the subcategories underneath it.
    • Replaced the grid’s status tint, row colours and legend with a single green ramp carrying each cell’s standing.
    • Added the three-mark corner gauge, so the tint has a countable reading beside it.
    • Added an admin area for reviewing and editing corpus entries without a rebuild.
  3. v0.326 August 2026

    Seven kinds, and a map of methods only

    • Every work now carries exactly one Kind. Only a METHOD — a mechanism given in enough detail to reimplement and shown working by its own authors — counts anywhere on the grid.
    • Split the single status badge into two independent measures: how much is built, and how much of the borrowed tradition the built work carries.
    • Two indexing sweeps added 181 works and 56 approaches, filling the last three empty regimes.
    • Made the fabrication flag structural rather than inferred from prose, and checked every reference against an index.
  4. v0.225 August 2026

    Reading approaches against their traditions

    • Coded whether each approach claims the concept it is being read against, across all 115 readings — which separates a shortfall against something claimed from a gap the authors never set out to fill.
    • Gave every concept of every tradition a plain-language definition, assuming no background.
  5. v0.124 July 2026

    First build

    • The two-axis taxonomy, fifteen regime entries, and the approach and paper records behind them.

Roadmap

What the encyclopedia is missing, in roughly the order it is likely to get it. Nothing here is a commitment, and an item marked considering may never happen.

  • Finish the fourteen regimes beside Constitutionalism

    Next

    Constitutionalism carries the research team’s own reading end to end. The other fourteen have a first-pass definition, limitation and commentary that still need the same treatment before their prose can be leaned on as hard as their counts.

  • A page per work

    Next

    Papers currently resolve to the approach they instantiate. Each needs a page of its own: the reading, the divergence from the tradition, the verification trail, and what else cites it inside the corpus.

  • Citation counts on a schedule

    Planned

    Counts are a single Semantic Scholar snapshot. They should refresh on a schedule, record the index and date beside each figure, and keep the known undercounting of law reviews visible rather than silently wrong.

  • A second reader on every concept reading

    Planned

    Presence and claim are judgment calls made once. Recording a second independent reading, and the disagreement rate between them, would put an error bar on the adherence figure the map turns on.

  • The corpus as a citable dataset

    Planned

    CSV and JSON exports of every table with a stable schema, a version number and a DOI, so the map can be cited and disagreed with rather than only browsed.

  • Proposal → method lag on the timeline

    Planned

    Mark, per regime, the year the first proposal landed against the year the first working method did. The gap between them is the quantity the map is really about.

  • Read an approach against a tradition it did not borrow

    Considering

    Every approach is currently read only against its own cell’s tradition. Reading it against a neighbouring one would show which regimes are near-misses of each other and which are genuinely different animals.

  • Coverage beyond arXiv and beyond English

    Considering

    The corpus leans on indexes that under-cover law reviews, policy literature and work published outside English. Naming that gap is easy; closing it needs a different search strategy.

  • Shareable reports from the configuration finder

    Considering

    A configuration is already a link. It should also be a short standing report — what exists, what does not, what it would cost you — that somebody can hand to a colleague without the reader having to work the form themselves.

Contributing

Corrections and additions are welcome, and the most useful contribution is the least glamorous one: a work that belongs on the map and is not on it, or a row that is on it and is wrong.

Proposing a work. Send the reference and, more importantly, the argument for where it goes: which regime it places in and why, which existing approach it instantiates or what new approach it establishes, and which kind it is. That last one decides whether it counts on the grid at all, so it is worth being strict about — if the authors did not build the mechanism and show it running, it is a proposal, however good the paper.

Disputing a placement. A placement is a claim about what a paper does, and several in the corpus are already recorded as contradicting the grid’s own account. Say which cell you think it belongs in and which part of the paper puts it there. Disagreements that survive get written into the entry rather than resolved away.

Disputing a reading. Every concept reading is one person’s judgment about whether an approach carries a concept and whether it ever claimed to. Both halves are arguable, and a second reader on every reading is on the roadmap for exactly that reason.

A submission address for the research team goes here.

In the meantime, the quickest way to see whether something is already covered is to search the corpus — it indexes every regime, approach, work, source and concept.