The lab's closing statement — what three programmes established

Dated 2026-09-23. Adopted by the author on 2026-09-24, as written, on the author's instruction ("Let's adopt this closing statement"). Written from the three Registers — register.jsonl (34 records, head c2706a09716b), programme-2.jsonl (37, head 23cf22aa6586) and programme-3.jsonl (14, head e92fda14a7f1) — beside docs/programme.html and the three adopted conclusions. Every number here is a record's; the Register does not choose, and this statement recommends nothing.

1. The scoreboard, three programmes

Across the three: 77,952 cells screened at zero alpha; six questions registered by the author; five verdicts reached by the machine against a floor declared before each run — four null, one supported; one question refused at measurement before a verdict; 0.55 of the 1.20 alpha declared across the axes spent, the three reserves of 0.10 never looked at (ADR-0006); 120 names archived under a rights record each; every partition cut from a frozen calendar (ADR-0038); every verdict at bounded costs (M5.0 closed by decision); nothing traded, no order placed, no account used (ADR-0044).

The verdicts, each beside its origin and its measured numbers:

2. What the Register establishes

• Five mechanisms were put to the measurement partition against a declared floor and four were not there. Three in the first programme (Q-003, Q-004, Q-005), one in the second (Q2-002): each refused by the floor, three of them by beats-null or leave-one-out as well. The falsifier fired on the third (ADR-0033) and closed the first lab.
• One mechanism cleared the four checks the lab then had, and the record beside it says the entry added nothing. Q2-001: EV +0.213 net R over 6,946 trades, 412.7 a year, beats-null, plateau, floor and leave-one-out passed (#22); always-long at the winner's geometry and gate on the same partition makes +0.214 (#24). The verdict stands as reached; the diagnostic beside it is why the lab now has a fifth check (ADR-0043) and judges the sweep on the margin (ADR-0045), binding from the next registration and never re-judging the one that showed the gap.
• The same mechanism, chosen by margin, beat always-long and missed the floor. Q2-002: +0.109 net R over 12,732 trades, four of five checks passed, the floor refused by 0.041 R, measured margin +0.075 against the screen's +0.201 (#33). One winner rule asked is there money here? and got yes for the wrong reason; the other asked does the entry earn it? and got yes, not enough.
• A screen's statistic did not predict transfer, twice. All twenty candidates of each survey passed the fifth check on the definition partition with a positive margin in every era; the one that was measured lost its whole margin, and the one registered in the third programme never reached the check.
• Three Drafts, two grids, six questions, one refusal: the pipeline refused a dead world and accepted a planted edge before every one of them (S3, S10), on both engines, and does so on every build of the console. No verdict was reached while a control was wrong.

3. What is not established

• Anything about any mechanism beyond the six questions. 77,946 of the 77,952 cells screened were never registered; a screen is not a verdict and a cell not registered says nothing.
• Whether a margin the screen finds transfers — the question ADR-0046 opened the third programme to ask. Its one question was refused at measurement; the apparatus to ask it stands ready and the numbers to afford it were not declared.
• Anything at measured costs, in a forward window, or live. Every verdict is at bounded costs; the one Strategy that reached FORWARD (Q2-001's winner) stands there as a record, because the execution host is deferred outside this lab (ADR-0044); APPROVED -> LIVE was never taken.
• Anything about the reserves, the fundamental axis, crypto, intraday or a second device — sealed or parked, each with its reason.
• Any universe below large caps (M12.1e), and any name the source does not serve: survivorship stays a named bias on every universe record.

4. The findings that outlast the programmes

The lab's product is its refusals. Six questions, one supported, and the supported one carries beside it the record that undid its meaning. Every refusal — a coin flip, a floor missed by 0.041, a plateau the sweep could not hold, a pooled effect carried by one name, a winner with fewer trades than the plan required — is in a Register with its evidence, and none was edited away. What was learned was learned from what the guards refused, and it is written where the next programme must read it.

Every gate must be shown passable before alpha moves, and "shown" means counted the way the guard counts. The first programme learned to freeze the calendar, measure ρ, require the seed and add a kind only by ADR. The second learned that the screen's statistic must be the verdict's. The third learned that a promise made from the template's signals is not the guard's count of the winner's trades, at the cost of 0.15 alpha and a refusal the guard was built to give. Each lesson is a guard, a stamp or a bound now, with a test; none is a note.

The always-long return at a geometry is a property of the era, not of the entry. At hold 20 inside the up regime it was +0.016 on 1993–2003 and +0.214 on 2003–2019 (#9, #24 of programme 2). No statistic on the definition partition saw it coming. A lab that compares an entry only to random entry can be right about chance and wrong about the market it is in; the fifth check exists because the fourth verdict showed that.

Numbers are declared for a shape. The alpha split that bought two 15-cell families on grid-001 bought none of grid-002's seventeen 30-cell candidates; a floor affordable at k = 4 is not a floor affordable at k = 15. The Registers record what each set of numbers could buy and do not say what the numbers should have been; that is the author's, and it was the author's each time.

Breadth at zero alpha is cheap and honest; depth is priced and rare. Two surveys of 38,976 cells each, computed on a spot burst in 47 and about 97 minutes and torn down the same day, screened everything the closed enum could express on four universes. From them the author registered three questions. The ratio is the design, and the design held: no cell was measured that was not registered, no question registered that was not screened or drafted, no alpha spent that the ledger did not charge first.

What was found was found on the way. The limit fill model, the exhaustive fill audit and the missed-trade census (ADR-0040–0042); a target family that could not compile; an archive that had grown under a reproduction; an engine sha dirtied by the loop's own writes; a sentence that denied its own gate; question ids that would have numbered one programme's question as another's; a power promise that counted the wrong thing. None was a verdict. Each is a record, a guard or a test, and the S4 audit of 2026-09-21 found no normative text contradicting another.

5. What it cost

Alpha: 0.55 of 1.20 declared across three programmes' axes, the reserves untouched. Data: the licensed archive at the source's free tier, 120 names, no bar in the repository. Compute: two spot bursts, torn down with every resource verified absent. Time: fourteen days from the first question to the third stop. Money: inside the cap recorded at M0.14, with no execution host and no subscription; no amount is written in any Register or page, by construction.

6. What would have to be true to continue

Stated as conditions, not as a recommendation, because each is a decision the Registers do not make:

• a configuration declared for the shape of the grid it will buy from — the floor, the split and the falsifier count, the author's three sets of numbers — after python -m occams whatif shows what it affords;
• a measured spread, the first act of any programme that trades, never a recommended figure (M5.0);
• an execution host, priced and credentialed outside this lab, before any Strategy leaves FORWARD (ADR-0044);
• a publication decision recorded either way (R7), whether or not any of the above is taken.

7. The decision

None is taken here. Three programmes are stopped by record and concluded, every conclusion adopted; nothing prepares, registers, surveys or measures in any of the three Registers again. The publication decision is the author's and is the last step in Definition of done; the publication check (make prepublish) passes on every page and Register, and is the gate's check, not the decision.

Drafted by the agent on 2026-09-23 on the author's instruction, from the three Registers as they stand and the three adopted conclusions. Adopted by the author on 2026-09-24, as written; the publication decision (R7) is the one step of Definition of done that remains.

Correction, 2026-09-24 (appended; nothing above is edited): §2 cites record #22 for Q2-001's four passed checks. The resolved record is #19; #22 is the Strategy's transition to FORWARD. Found while indexing the Registers for docs/INTEGRITY.md.

Note added 2026-10-10. Nothing above has been edited.

In October 2026 an outside reviewer looked at the statistics behind the five checks and found they were easier to pass than we had claimed. We reran every point the review made, and it held up. The fixes went out in three releases, 1.0.1 and 1.1.0 on 3 October and 2.0.0 on 4 October. The checks are stricter now. The three Registers this statement is written from have not changed by a single line.

That raised an obvious question: if the six questions had been judged under the stricter rules, what would the answers have been? To find out, each question was first rerun in the exact code that judged it at the time, to confirm the original numbers came back. Five of the six did. Q3-001 never reached the engine, so there was nothing to rerun. Each of the five was then judged again with the current code, on the same data and the same seed. The results are kept in register/diagnostics.jsonl and written up in plain terms in docs/RESCORE-2026-10.md.

The one that matters is Q2-001, the only supported verdict this lab ever produced. Under today's rules it fails. The best cell in its sweep is now a different one from the one recorded (cell [2, 0] rather than [4, 0]): 7,642 trades averaging +0.091 R each, which is +0.054 R better than simply being long at the same times. Once the uncertainty in that average is allowed for, the lower bound is +0.002 R, against a floor of 0.15 R. And the fifth check, which asks whether the entry signal adds anything over just being long, gives p = 0.031 where 0.01 is required. Nothing that failed a question originally passes it now, and no question comes out better than its original verdict.

What this does not change. The Q2-001 verdict stays "supported" in the Register, because that is what the rules in force at the time said, and sealed records are never rewritten. This note sits beside it. The falsifier count, the alpha spent and every number in §1 are as they were. The lab is still closed: nothing will be registered or measured in these three Registers again. And this note is not a verdict. It is a reading of the old questions under new rules, written down so that nobody has to take the one supported verdict at face value.

One known gap remains. On the multi-day engine, when the names in a universe move together with the market, the beats-null check still passes a little more often than it should at the 5% level. Every question in this lab was run at the 1% level, where it behaves. The gap is recorded in ADR-0048.