Method

The method is the set of rules by which the Broken-Born program labels, checks and reviews its own claims, together with the working arrangement between Paul and Claude that the rules serve. Every claim carries a tag saying where it came from, every asserted number is to be computed by a script kept in the repository, the scripts run in one gate that must pass before work counts as finished, dead claims get obituaries of their own and stay dead, and manuscripts go to blind referees only after the expected verdict has been written down. The rules exist because physics drafted by an AI has a known failure mode, fluent confabulation, and because Paul cannot independently check most of the technical physics; checkpoint #11 says that “every claim here carries a tag saying where it came from” for that reason (§2). Most of the rules were adopted after a specific failure, and their pages record it.

Stands: Lakatos’s methodology (established), the collaboration, provenance tags, claim tiers, negative results and dead claims, verification, blind review — the rules in force · Died: nothing in this section (the graveyard lists the program’s dead claims) · Open: the † literature battery, open problem 10 of #11 §14 (open problems and the queue)

Map#

The section runs from the frame the program judges itself by, through the arrangement that produces its claims and the labels that qualify them, to the machinery that checks them. Only the first page is established material from outside the program; the rest record how this program works and the episodes that shaped each rule.

  • Lakatos’s methodology of research programmes comes first because it supplies the vocabulary of the program’s self-description. It explains, as established philosophy of science [T], the hard core and protective belt, the two heuristics, and the grading of problem shifts as progressive or degenerating by whether they predict novel facts. It then shows how #11 §15 sorts the record under those grades, how two claims lost their novelty, where the calc notes flagged degeneration, and why round-3 Decision 2 declined to rescue the soliton candidate with a free parameter.
  • The collaboration: roles, attribution and conventions says who does what. Paul (P.R.), an independent developer, supplies conjecture and direction and holds a veto; Claude supplies derivations, counterexamples, verification and drafting. The page gives the attribution line and its earlier wordings, the calibration Paul relies on and its four-round record, the decisions he took, delegated or has still to rule on, the chunked working pattern, the prose register, and the move from claude.ai to Claude Code for code execution.
  • Provenance tags and citation marks defines the labels every claim carries ([D] [T] [P] [O] [V], and † or [S] on each citation), with an example of each from #11, the combined and qualified forms, and the three ways a tag has changed in the record: a derivation found in the literature, a memory citation retrieved, and a flag turned into a mark. It then covers the variants in other documents, including the handoff’s ✓ (flag 16) and three of #13’s tags flagged during the wiki build (flags 29 and 30 in review/wiki-build-flags.md), the system’s growth from the lab notebook to #11, and how the wiki carries the tags.
  • Claim tiers and selection bias ranks the claims by how far they can be trusted: Tier 1, the seven machine-checked tests of the battery; Tier 2, literature pending source checks; Tier 3, novel structural physics at a stated prior of ~50–80% per claim (#11 §2). It explains why negative results are the most reliable Tier-3 items, and where the selection-versus-derivation bias operates: the frame-tier coupling class, “selected, not derived”, and the qualifier dropped from “unique known consistent structure” (flag 7 in review/wiki-build-flags.md).
  • Open: the † literature battery (OP10) follows the tags because it is the open task they create: every reference cited from model memory in #11 and #13 is to be checked against its source before any external use. The page lists the entries by subject, says what a check involves, records the citation errors found so far (one suspected in #11 itself, flagged during the wiki build as flag 20 in review/wiki-build-flags.md) and the partial progress, and places the battery in the first item of the current queue.
  • Negative results and dead claims states the rule that a dead claim gets an obituary of its own and stays dead. It lists the deaths in the record, by calculation and by review, with the numbers that killed them; shows how a death is reported, from #11’s abstract to the obituary format; records the rescue the program declined and what each death left standing; and records the misses in the method itself.
  • Verification holds the rules for numbers (the iron rule, independent anchors, the reduced-model and error-bar rules), the gate scripts/check.sh, and one page for each family of scripts the gate runs.
  • Blind review closes the section: fresh AI referees, the expected verdict written down in advance (pre-registration), the referees’ measured failure mode, and the four review rounds.

Boundaries#

Backward Superseded statements Lakatos's methodology of research programmes Forward