v0.12.0 — fifteen papers read

The first release since fifteen papers were read in full.

	v0.11.0	v0.12.0
Records	319	363
Claims	112	349
Papers read down to claims	7	27
Claims checked by a person	43	43
Open questions	4	13
Organizations · people	—	210 · 2,390

Reading moved a long way; reviewing did not move at all. That row is the
honest one.

What changed

Fifteen papers read. The collusion work, the multi-agent security papers, the workplace and market benchmarks, the economics of agent firms — the anchors the corpus already held and had never read. Every quote verified verbatim against the source.

One question now reaches six papers that share no author, each arriving at the same missing instrument from a different direction: how do you evaluate a system of agents rather than the agents in it? That is the case the open-problems layer exists for.

MIRA's vocabulary throughout. A statement is a claim, a finding is evidence, a method is a protocol. The old names raise rather than alias, so there is one vocabulary and not two.

The borrowed foundations are held, at section 15: Coase, Williamson, Aghion and Tirole, Jensen and Meckling, and eleven others. Pointed at rather than read, so a claim can cite where a term came from rather than whoever last used it.

What is in the release

nodes.jsonl            1.86 MB    every node, one JSON object per line
relationships.jsonl    1.61 MB    every edge, with where it came from
graph.mira.jsonld      1.06 MB    the whole corpus in the MIRA schema
mira-report.json                  what MIRA cannot yet carry, counted and named
taxonomy.json                     103 topics, 16 sections
metadata.json                     counts, and the taxonomy version
checksums.txt                     so this release means the same thing next year

The MIRA projection conforms — 12,556 triples against the schema's own generated shapes, pinned at commit 365d4040. You can re-run that check yourself:

python3 -m pip install '.[mira]'
python3 scripts/validate_mira.py data/releases/v0.12.0

What this is not

Nothing here is finished. 27 of 363 records have been read down to their claims, and 43 of 349 claims have been checked by a person against the paper they came from. Both figures are published rather than rounded up, and the second one is the binding constraint.

Corrections are the most useful thing you can send: the form is here.