The second programme runs beside the first, in the same repository, on its own Register

The first programme's lab closed on 2026-09-12 when its falsifier fired (ADR-0033): three mechanism verdicts, all null, on a floor and a count declared before the first bar was read. ADR-0033 says what closing is — terminal, a publication, not repaired by new data — and the Register enforces it: nothing registers after LabClosed (#33).

The author then asked for what the first programme did not do: a full set of strategies tested across multiple histories, depth and breadth, in batch, with the hypotheses and results written and logged legibly. That is a different premise from the one that closed, and ADR-0033 already names its form: a new question is a new programme.

Decision

Programme 2 runs in this repository beside programme 1, and never in its Register.

• Its own Register and queue: register/programme-2.jsonl and register/programme-2-queue.jsonl. The first programme's Register is never written again; the console renders both, the first marked closed.
• Its own configuration: configs/programme-2.toml, gitignored like every configuration. It carries the author's floor convention, falsifier count and alpha split for this programme, declared after python -m occams whatif --archive archive --register … has shown two tables the first programme lacked: the falsifier arithmetic — P(the first N mechanism verdicts are all null) at a range of base rates of true edges, with the declared count marked — and universe affordability — the trades each universe affords on the measurement partition after the design effect at its own same-day ρ, and the largest floor that count can detect at each sweep size. The numbers stay the author's; the tables say what they buy. A count that closes the lab before the base rate can show is a test of edges are common, not of an edge exists — that is what programme 1's count of three was, and it is written here so it is not chosen twice by accident.
• Its own frozen calendar at its first act (ADR-0038), per universe.
• The same apparatus, unchanged: the spec, the engines, the guards, the accountant, the loop, the console. Every command takes --register and --config, so the two programmes share code and nothing else.
• Breadth by survey at zero alpha, depth by registration with alpha (M12): a survey reads the definition partition only and is a Register record; registration is the author's --yes naming the ids; the four checks and their evidence decide on the measurement partition; the cells screened are counted beside every question registered from them (N6).

Consequences

• whatif gains close_probability, universe_rho, affordability and the two tables; --archive and --register flags.
• question and loop gain --config; the console's --register repeats; make console renders both Registers.
• The first programme's conclusion (docs/PROGRAMME-CONCLUSION.md) stays a draft until programme 2 has run, and is then joined by programme 2's own, written after the search and not before (M12.8).
• M12.1 (universes as records), M12.2 (the grid), M12.3–M12.7 follow; none registers anything.

Considered options

• Reopen programme 1 with a larger falsifier count. Rejected: a count raised after three verdicts is a count chosen with knowledge of them, which is the thing every gate in the lab exists to prevent; and ADR-0033 says terminal.
• A new repository. Rejected: the apparatus, the archive, the rights records and the ADRs are the accumulated value; a second Register is one file and keeps the two records distinguishable by construction.
• One Register, two programmes. Rejected: LabClosed is a terminal record for its Register, and the falsifier count is evaluated over that Register's verdicts; mixing programmes would either bypass the closure or count the first programme's nulls against the second.