Research line: research tooling and rules

This line is not about the model. It records the rules, checkers and command-line tools the project built to keep its own research honest: who may run what, how claims are worded, and which banned operations must never creep back in. It holds 185 records and not one is a model result. Most of the effort came in two days in mid-October, when the whole tool was ported from Python to Rust.

Source: archive/catalog.jsonl, counted with jq. Terms follow the glossary; ACTIVE, READY and PARKED are defined on the mechanism board.

flowchart LR
  A["Jul to Sep<br/>Claim-language rules,<br/>handoffs, reviews"] --> B["12 Oct<br/>Owner bans softmax,<br/>CPU compute, matmul"]
  B --> C["13 to 14 Oct<br/>Harness ported<br/>Python to Rust"]
  C --> D["15 to 17 Oct<br/>Write-path fixes,<br/>ratchet hardening"]
  D --> E["18 to 19 Oct<br/>Docs reconciled,<br/>charter adopted"]

The line in four phases. The second box explains much of the third: the Rust-first rule (D27) required the ports.

What it set out to do

Make a many-lab research project reproducible and hard to fool. Three jobs recur: keep claim wording honest (formal vocabulary), keep banned mechanisms out (scripts that count softmax-like code and dense matrix products and fail if the count grows), and let several AI labs share one repository without overwriting each other (claims on branches, check-ins, receipts).

What happened, in order

1. Jul to Sep. Normative claim-language rules, a traceability matrix of the formal analysis, handoff documents and audits of outside repositories. No code was adopted from the audits (external research audit).
2. 12 Oct. The owner banned softmax everywhere (D23), CPU model compute (D24) and dense matrix products (D25). A scan found all five active code paths still used softmax (softmax census). The first harness and its multi-lab version arrived the same day. An agent also removed 11,171 research files from git to save disk; they were restored byte for byte (integrity record).
3. 13 to 14 Oct. About 105 records port the harness commands and checks to Rust, each compared against the Python output. The comparisons kept finding faults in the checkers themselves: a parity check that could never fail (harness CI), a green report when the binary was not built and nothing was checked (skipped is not a pass), and a check command that ran none of its 41 rules (dispatcher).
4. 15 to 17 Oct. Fixes to how the tool writes state, branch-cleanup proof for squash merges, and hardening of the ban scans. Five rules that were registered but never executed were wired in (record).
5. 18 to 19 Oct. Documentation reconciliation against the source, the offline-float decision D37, the owner-phrase rule D40 and many one-paragraph documentation syncs (course correction).

Key results

What it taught

Every checker needs a test that can fail (lesson 1 in what the project knows). Several tools here reported success while checking nothing. A process that forbids running code produces nothing (lesson 21). Section 3 of the synthesis separates owner-mandated rebuilding from discretionary gating as causes of the drop in model results. Branches are temporary; main is the authority, and a merged claim must be released.

Current state

Active but slimmer. The charter adopted on 19 October requires a written sheet only before GPU runs or milestone-claim runs, and keeps one notebook per line (how research is done now). No record in this line is a model result.

Go deeper: the generated catalog page lists all 185 records.