The third programme asks whether a margin the screen finds transfers

Programme 2 ran and stopped on the author's record (#36) with two verdicts and one lesson in two halves. Q2-001, its winner chosen by EV, cleared the floor on the geometry where being long in the up regime paid the same and the entry added nothing (#24, margin −0.001). Q2-002, its winner chosen by margin under ADR-0045, beat always-long at the corrected alpha and missed the floor (#33, #35). The screen had ranked both families by their margin over always-long on the definition partition; the readiness pass of 2026-09-18 showed every one of the twenty candidates passing the fifth check there with a positive margin in every era, the one that lost its whole margin among them. Programme 2's conclusion (adopted 2026-09-20) says what that leaves open: the definition partition's eras give no warning; whether a margin the screen finds transfers is a question the second programme could not ask cleanly, because its screen and its verdict ranked by different statistics until its last question.

That is the third programme's question, and ADR-0033's rule applies as it did to the second: a new question is a new programme.

Decision (the author's, 2026-09-20)

Programme 3 runs in this repository beside the first two, and never in their Registers.

• Its own Register and queue: register/programme-3.jsonl and register/programme-3-queue.jsonl. Programme 2's Register ended at the author's stop and is never written again; the console and the programme page render all three.
• Its own configuration: configs/programme-3.toml, gitignored. It carries the author's floor convention, falsifier count, alpha split and partition split for this programme, declared after python -m occams whatif configs/programme-3.toml --archive archive --register register/programme-3.jsonl has shown the falsifier arithmetic and the universe affordability table. Nothing in this ADR proposes a value, and programme 2's values are not a default: the author copies them or declares others, and either is a recorded act (M13.0).
• The same four universes, re-declared as records in the new Register from programme 2's latest record under each name — members, rule and bias unchanged, the flat bars named — unless and until the author supersedes one. Each is frozen on its own calendar at the programme's first act after the configuration exists (ADR-0038).
• A grid that supersedes grid-001: surveys/grid-002.toml, the same closed enum on the corrected engine — the resting limit fills at the open (ADR-0040) and the fill audit is exhaustive (ADR-0041), so breakout_low is measurable for the first time — with the same geometry; a grid is never edited, and this one names what it supersedes.
• The apparatus as it stands, unchanged: nine entry kinds by ADR, the survey at zero alpha on the definition partition, gate readiness under the fifth check with the margin in every era (survey candidates --fifth-check), registration by the author's --yes, five checks on the measurement partition, the winner and the plateau and leave-one-out on the margin over always-long (ADR-0043, ADR-0045), the shrinkage and the eras recorded after every verdict, a stop that is a record and a conclusion written from the Register after it (M12.8). Every command takes --register and --config; the programmes share code and nothing else.
• What the third programme reads that the second could not: two measured priors. The readiness table and the conclusion cite every Shrinkage record beside the screen; programme 3's tools read programme 2's Register for them, read-only, so that the one statistic the screen and the verdict now share is compared with what it did the last two times.

Consequences

• The M13 phase in TASKS-v4.md, mirroring M12's rows with the machinery already built: declared (M13.0, the author's numbers), universes and calendars (M13.1), the grid (M13.2), the survey on the burst (M13.3), the page (M13.4), candidates and registration (M13.5), the loop (M13.6), the programme page (M13.7), the stop and the conclusion (M13.8).
• make console and make programme take the third Register; the prepublish check covers it as it covers the others.
• Programme 3's first spend of money is the burst that computes its survey, priced by the last one at about the cost of a coffee; its first spend of alpha is the author's --yes.
• Programme 2's Strategy at FORWARD (#22) is not carried over: ADR-0044 deferred the host, and a Strategy in one programme's Register is not a question in another's.

Considered options

• Reopen programme 2 with the unspent 0.05 on each axis. Rejected: its Register ended at the author's stop, terminal by construction (M12.8), and 0.05 buys no family the survey holds.
• A grid two inside programme 2's Register. Rejected for the same reason, and because the question changed: the second programme asked what its grid held; the third asks whether what a screen finds transfers, and needs its own falsifier declared for that.
• A programme that trades. Rejected as this programme: ADR-0044 deferred the execution host outside the lab, and M5.0 closed by decision with the cost basis bounded; a programme that trades begins with the host and a measured spread, and this one begins with neither.
• The same four universes, or new ones. The four are re-declared so the priors from programme 2 compare like with like; the author may add a universe by declaration at any time, and M12.1e keeps the lab above large caps.