Adapt the template to a document-only shape

[lvelho]

Instantiated from visgraf/math-ai-method at 81a638e. This repository holds one technical report, its markdown draft, and the pass that verifies its claims. No library, no experiments, no Python.

Note on how this landed. The specification asked for the first commit direct to main, with PRs after that. main has a ruleset requiring pull requests, so the push was rejected and this is that first commit as a PR instead. Nothing about the change differs.

Deleted

experiments/exp000_example/, src/research/, tests/test_scaffold.py (fc-003); spikes/ (fc-004); pyproject.toml (fc-002); .github/workflows/ci.yml (fc-001). report.yml is now the only workflow and the entire automated surface.

One correction to the premise. test_scaffold.py was described as having one job (the spikes/ import ban). It had four, and the .reference/ guard protects a convention this repository actively depends on — the pinned checkouts are how every claim gets verified. It was deleted anyway because its precondition is that executable code exists to scan, and there is none: it would have passed by scanning an empty set. The convention now rests on a comment in .gitignore. Recorded in fc-003 rather than absorbed.

Kept, against the obvious call

• docs/state.yaml (fc-005) — the ledger's argument never depended on there being experiments. It exists so a later session can tell we decided against this from nobody thought of it, and a deleted directory leaves no trace at all. The evidence is empirical: this adaptation generated six foreclosures, five open decisions and a stopping rule before any report work began. What did not transfer is marked, not tidied — falsifier_verdicts and constraints are empty and may stay so.
• docs/inherited-measurements.yaml (fc-006) — a report in which every figure is inherited is this file's ideal case, not a category error: a count in a repository at a pinned SHA is exactly recheckable. claims-verified.md cites ledger ids as its Source field, so deleting the ledger breaks the retained instrument.

Found while deciding fc-006

The instance's verification tally appears twice and disagrees with itself:

source	figure
report/claims-verified.md	22 claims — 10 confirmed, 3 corrected, 8 under-qualified, 0 unsupported
report/draft.md	eighteen load-bearing claims, seven stood as written

They agree exactly on 3 corrected and 8 under-qualified, and differ on the total and the confirmed count — reconcilable if "load-bearing" names a subset, which neither passage says. That is an UNDER-QUALIFIED defect in the two documents that define the category. Neither sentence is edited. They are mn-001/mn-002; resolving them at source is od-002, blocked on gap-001.

Also

CLAUDE.md's two marked blanks are filled. In Domain invariants the template's "declare it at the function that returns the value" rule is inert — nothing here produces a value — so that section becomes the only place conventions can live.

Verified: make -C report builds clean from distclean (exit 0, no LaTeX errors), both ledgers parse as YAML, and no reference to a deleted path dangles.

🤖 Generated with Claude Code