How research is done now

UOR-R4 is a research project, and for months it ran research the heavy way: one folder per pull request, a written plan before almost every step, and a review before each run. In the week of 12 to 18 October only 11% of the recorded lab work was a model result (see history). Since 19 October 2026 the owner's charter keeps the rigour but drops the bulk. Each line of research has one notebook page with one row per result. A written plan is required only before a GPU run or a run that a milestone claim rests on. Old records are kept, untouched, in archive/.

flowchart LR
  A["Question or idea"] --> B["Notebook page<br/>for its research line"]
  B --> C{"GPU training or<br/>evaluation, or a<br/>milestone claim?"}
  C -- no --> D["Build, test, run<br/>host checks"]
  C -- yes --> E["One-page sheet"]
  E --> F["Review by someone<br/>who is not the author"]
  F --> G["Owner releases<br/>the sheet once"]
  G --> H["Run on the GPU"]
  D --> I["New notebook row"]
  H --> I
  I --> J["Sealed artifacts and<br/>old records go to archive/"]

The path of one piece of work. Only the right-hand branch needs a sheet, a review and an owner release.

Terms follow the glossary. Claims are labelled proof, measurement or hypothesis as defined in formal_vocabulary.md. The rules below are the research clauses of the charter; section numbers point there.

The working rules

One notebook per research line (charter §5.6). A line is one sustained question, such as "how the discrete model reads context". Its notebook lives at docs/research/notebook/<line>.md. Each result is one row: date, what ran, the number, the bar it was tested against, the verdict and a link to the artifact. A dated evidence folder is kept only when a run seals artifacts. The one-page plan for the whole project is NOW.md; with the notebooks, it replaces the old running log (current-state.md).

A sheet only where it decides spend (§4.5). A sheet is a pre-registration of one page: the bar, the noise, the controls, the seeds, the cost and what each outcome would decide. It needs a review by a non-author, called a FIT review, before:

• any GPU training or evaluation, and
• any run that a milestone claim rests on.

A fresh-context subagent counts as a non-author. Builds, finite kernel tests and host checks need no sheet. Old float sheets come back only as new sheets with a new FIT. A FIT review releases nothing by itself.

One owner release per sheet (§5.1). The owner types one release phrase personally for each sheet. It covers all the sheet's arms, seeds and registered stop rules. Labs stage the commands and then check the records; no lab types a release for the owner.

Mechanism statuses (§4.1). Every mechanism is ACTIVE (on the model path), READY (built and tested but not yet trained in) or PARKED (the tested forms failed or broke a rule). A PARKED page lists what would reopen it. There is no terminal status. The mechanism board shows the current state. This is the rule from decision D22 (a negative closes a configuration, never a mechanism).

What a verdict closes. A negative result closes the exact tested configuration: the recipe, dose, window, key and objective. It never closes the mechanism, and stopping a novel mechanism takes an owner decision. A run that meets its own pre-registered mechanism target without a significant headline loss is a KEEP, and the work iterates. A tooling or numerical failure is a blocker to fix, not a negative. Source: D22 sections 1 and 2 in archive/integration/DECISIONS-2026-10-19.md.

Instrument first (§4.2). Before calling a result, decide whether the check itself is faulty. An instrument defect is a fix, never a negative. No target is set from a vacuous instrument, and no single number rules out a line. Findings shows why: a metric that read zero, a control that checked zero rows and a typed-in number each misled the project once.

Power and transfer (§4.3). A decisive run uses at least two seeds or at least 200 fresh rows. Its bar is never narrower than the measured noise. The mechanism must be trained in, not bolted on. Every gain on a development panel is reported together with a fresh transfer number.

Controls and pivots (§4.4). Compare against lawful non-neural baselines, which today means count tables. After three results on one line that do not move its headline number, write a pivot note: stop, or run one decisive test.

Plain verdict words. NOT_RUN means nothing ran; it is not a pass. UNAVAILABLE means a needed input was missing; it is not a negative. A passing test is not trained capability, and a passing control is not a language result (§4.6).

Preserve everything (§4.7). Nothing is deleted. Negatives, unique source and other labs' work go to the archive, byte-identical.

The lines of research

The catalog groups its 1,173 records into 20 lines. The ten largest are below. Counts are catalog records, which mix lab records and lane results, so they show where effort went and not how many experiments ran.

Source: archive/catalog.jsonl (fields line, date), counted with jq. The remaining ten lines hold 14 records or fewer each. The descriptions come from the line names and milestone titles, not from a field in the catalog. Every line becomes a notebook page as it is touched.

Where old records live

Every dated record before 18 October 2026 now sits under archive/, at the same relative path it had under docs/, except the reference documents that stayed in docs/ (such as docs/scaling_law.md and docs/inference_contract.md). Records dated 18 and 19 October stay in docs/labs/. The archive never changes: many experiment sheets are pinned by hash, and SHA256SUMS files cover the records. So read an archived record as a faithful copy of the day it was written, including any claim that was corrected later.

To find a record, use the generated index archive/CATALOG.md, with the same data in archive/catalog.jsonl. Its columns are date, lab, line, milestone, mechanism, verdict and path. It also holds the old-path to new-path map, so an old link still resolves. The catalog is an index, not a verdict. Follow its path and read the record before you rely on a number.

Go deeper

• What the project knows: the lessons and the measured results behind them.
• What is still open: the tests nobody has run and the cheapest decisive ones.
• How the project got here: five phases and 32 eras, 10 June to 15 October.
• Mechanism board and the charter.
• NOW.md for the current plan, and docs/harness/README.md for the research harness.