Hypothesis Register

Per-task quality scores are unimodal rather than bimodal.

Outcome. Sarle's bimodality coefficient per task. Unit of analysis: the task.

Refuted by.

• A coefficient above 5/9 on tasks with enough runs to see it.
• Or a run count too small to distinguish the two, which makes it untestable rather than false.

Evidence. none recorded

Cost. a 24-run pilot detects a 2-SD separation 69% of the time and a 3-SD one always

ICC ≥ 0.5, so most variation is task difficulty and pairing pays.

Outcome. The intraclass correlation of quality by cell, with its interval. Unit of analysis: the cell.

Refuted by.

• An interval whose upper bound falls below 0.5.
• The point estimate alone does not do it, which is why the status below is undecided.

Evidence. ICC of judged quality by cell is 0.413, 95% interval [0.00, 0.88], from 56 verdicts over 28 pairs of the 12 race-02 entries. The point estimate is below 0.5; the interval does not exclude either side, so the stated refutation condition is not met.

Cost. reaching ±0.16 on the ICC takes 144 runs

At a fixed run budget the standard error is minimised at one repeat per task.

Outcome. The standard error of the mean at each allocation. Unit of analysis: the allocation.

Refuted by.

• A simulated allocation beating one repeat by more than simulation error.
• Or a budget model where repeats and tasks do not trade one for one.

Evidence. Predicted and simulated SE agree to about 1% across six allocations, and one repeat wins at every ICC, as the algebra requires. Settled without spending a run.

Cost. nothing; it is a claim about the estimator

Runs of one task are exchangeable: no drift, no within-task order effect.

Outcome. Turn duration by within-task position. Unit of analysis: the turn.

Refuted by.

• An order effect that survives dropping any single task.
• Or drift across the run that survives controlling for task composition.

Evidence. Pooled slope −0.47 log-seconds per position (p = 0.012, 20,000 shuffles), but dropping lp-16d590 takes it to −0.10 at p = 0.17 while six other leave-one-outs move it under 0.1. The effect is one bead whose four runs were all gate failures. Exploratory: the outcome is duration, because the loop recorded no quality scores.

Cost. already paid, since it reuses the loop ledger

Fidelity to the original brief falls with the ken ratio, and depth acts only through it.

Requires. The isolated regime. Each turn receives its in-edges and nothing else, with no inherited worktree and no readable history. Under lineage or sharing the ken ratio is 1 for every turn of every shape and this hypothesis is undefined.

Outcome. Of k constraints planted in the setup brief, how many the sink still honours. Unit of analysis: the turn, not the run — so one run yields six observations and runs are blocks.

Predicts. briefed > chain (same six turns, same depth, ken 1.00 against 0.33); bottleneck ~ chain (ken matched at 0.33 while depth differs, 3 against 5); star > bottleneck (depth matched near the shallow end, ken 0.83 against 0.33)

Refuted by.

• bottleneck and chain differ once ken is matched, which would mean depth acts on its own.
• briefed and chain do not differ, which would make the ratio decorative.
• The ordering holds but tracks the sink's in-degree alone, which is cheaper to compute and would make the ratio surplus.
• The run is found not to have been isolated, in which case nothing was tested and the result is void rather than null.

Evidence. none recorded

Cost. 180 turns paired at d = 0.8, and the manipulation is four edges and no extra turns. But the effect size depends on lambda, which H8 measures for one chain run: the gap runs from 0.667 at lambda 0 to 0.079 at lambda 0.95, so H8 should be run first because it prices this.

Turns in one automorphism orbit are exchangeable, so the largest orbit is replication the run already contains.

Requires. The isolated regime. Each turn receives its in-edges and nothing else, with no inherited worktree and no readable history. Under lineage or sharing the ken ratio is 1 for every turn of every shape and this hypothesis is undefined.

Outcome. The standard error of a within-run contrast, measured, against the value deff predicts from orbit size. Unit of analysis: the run, comparing observed SE to predicted SE.

Predicts. SE(star) / SE(chain) = sqrt(1.00 / 1.79) = 0.75 (from orbit sizes 4 and 1 at rho = 0.413, before any data); within-orbit variance does not depend on which member (that is the exchangeability claim itself); between-orbit contrasts stay confounded with position (no shape rescues those)

Refuted by.

• The observed SE ratio departs from the predicted one, so orbit members are not exchangeable after all.
• Within-orbit variance depends on member identity, say w1 always beating w4, which would mean dispatch order leaks and the orbit is not one in practice.

Evidence. none recorded

Cost. no extra turns, being a re-analysis of any run of a shape whose largest orbit exceeds one

Turns with equal ancestry digests are the same experimental condition, across different plans as well as within one.

Requires. The isolated regime. Each turn receives its in-edges and nothing else, with no inherited worktree and no readable history. Under lineage or sharing the ken ratio is 1 for every turn of every shape and this hypothesis is undefined.

Outcome. Within-cell variance of a digest class, pooled across shapes against pooled within one. Unit of analysis: the ancestry class.

Predicts. the 13-turn cross-shape cell behaves as one condition (its within-cell variance matches the within-shape variance for the same digest); the six sources form a second cell at deff exactly 1 (one per run, so they are independent)

Refuted by.

• Cross-shape pooling shows more variance than within-shape pooling of the same digest.
• That would mean the runner leaks something the plan does not contain, which is an R15 violation rather than a fact about agents.

Evidence. none recorded

Cost. nothing beyond running the catalogue, which H5 already pays for

What an agent still holds from an ancestor decays geometrically in hops, at a rate lambda belonging to the handoff and the regime rather than to the shape.

Requires. Only that the run states its regime, since lambda belongs to a handoff-and-regime pair rather than to the world. Measuring it under a shared worktree is the control, not a mistake. What it does require absolutely is that no brief anywhere names the residue.

Outcome. Recall of an INCIDENTAL residue at each depth of a chain, scored as set overlap against the harness log, minus the recall a fresh agent achieves on the same question. Unit of analysis: the residue item — k items observed at each of d depths.

Predicts. recall decays geometrically above the floor (straight on a log scale, not a cliff and not flat); lambda is the same on a chain and on a bottleneck (it belongs to the handoff, not the plan); lambda under a shared worktree is near 1 (nothing is lost when nothing is withheld — the control); the floor is well above zero (if it were zero the residue would be unguessable, which no real residue is)

Refuted by.

• Recall does not decay geometrically above the floor, in which case one rate does not describe the handoff and the model is wrong rather than imprecise.
• lambda differs by shape, which would mean it is not a property of the handoff and cannot be carried between designs.
• The floor arm scores as well as the descendants, in which case the probe measures guessability and not retention at all.
• The residue turns out to have been named or implied somewhere in the briefs, in which case it is a copying test and the result is void rather than null.

Evidence. none recorded

Cost. THIRTY-SIX TURNS, not the one this originally claimed. Simulation puts a single six-turn chain at a 95% width of 0.36 on lambda, with 6% of runs failing to identify it at all. Six chains at forty residue items each reach a width of 0.19. Depth beyond six HURTS: recall reaches the floor and the extra points are noise, taking the width from 0.25 at depth 6 to 0.30 at depth 12 and the bias from -0.005 to +0.073.

Six unattended turns reach the defect density of three directed ones, and whether they do is decided by two rates: how much context survives a handoff, and what share of live defects an unattended turn removes.

Requires. A stated regime, as every Unit IV hypothesis does, and the seeds must be invisible: no brief names them, and they must be the kind of fault the task would plausibly contain. It also requires lambda, so H8 runs first — g is estimated by discounting exposure by lambda, and a wrong lambda biases g in the direction that flatters the unattended arm.

Outcome. Defect density — surviving seeded defects per turn of delivered work — for an unattended chain at each length, against the same quantity for a chain a person reads at every handoff. The reported quantity is which of three bands the crossing falls in, not where exactly it falls. Unit of analysis: the seeded defect, observed at each gap from the turn that introduced it.

Predicts. defect density falls with chain length and then flattens (the flattening height is the residue, and it is above zero for any lambda under 1); the directed arm keeps falling where the unattended arm has flattened (a person`s context does not decay, so their later chances stay worth something); the crossing is finite when g and lambda are both moderate and absent when both are low (the two rates substitute for each other, so briefing can rescue a weak gate); giving turns better tools moves g and not lambda (a compiler or a search catches more of what a turn can already see; it does not restore what a handoff dropped)

Refuted by.

• Defect density does not flatten with chain length, in which case there is no residue and the model`s central asymmetry between watched and unwatched work is wrong.
• The unattended arm matches the directed one at equal length, which would mean direction supplies nothing this model can see and the whole decomposition is idle.
• g varies with chain length rather than staying a property of the turn, which would make it not a rate and nothing here would compose.
• The seeded defects turn out to have been announced or inferable, in which case a hunting rate was measured and the result is void rather than null.

Evidence. none recorded

Cost. SEVENTY-TWO TURNS, half of them read by a person, at forty seeded defects across six six-turn chains per arm. THE SIMULATION CHANGED THE OUTCOME RATHER THAN THE PRICE. Far from the feasibility boundary that design calls the three-way verdict right 100% of the time. Near it, the same design is right 96% on the three-way verdict and only 37% on the exchange rate itself, because a swing of 0.05 in g moves the rate from 40 turns to 9. So the outcome is the band and not the number, and a study sized on parameter precision would have bought a figure nobody could act on.

An executable check does not attenuate. Its detection rate is flat in gap, where an unattended turn`s falls geometrically, because a check is re-run rather than remembered.

Requires. A stated regime, and the same absolute condition as H9: no brief names the seeds, and the check-writing turn must not see them. It also requires that the check be genuinely re-run at each later turn rather than summarised, since a summarised check IS a remembered one and would attenuate as the model says everything else does.

Outcome. Detection rate for seeded defects, by the gap between the turn that made the defect and the turn that ran the check, for defects inside the check`s coverage. Unit of analysis: the seeded defect at one gap.

Predicts. in-coverage detection is flat in gap (a check written at turn 2 is as sharp at turn 6 as at turn 3); out-of-coverage detection decays at the lambda H8 measured (the control, and it is in the same run); the integrator`s own defects are detected at the same rate as anyone else`s (a check has somebody after the last turn, which is the part WP3`s model could not have); some passing artefacts are wrong in a way the check asserts (unsoundness is not zero and this measures it directly)

Refuted by.

• In-coverage detection falls with gap, which would mean a check is remembered rather than re-run and the whole distinction collapses.
• In-coverage and out-of-coverage detection decay alike, in which case coverage is not the right partition and the model is wrong rather than imprecise.
• Detection is flat because it is flat at zero — a check nothing trips is not a gate, and the loop ledger`s probe ceiling is what that looks like.
• The checks turn out to have been written against the seeded defects, in which case the seeds were announced by the back door and the result is void rather than null.

Evidence. none recorded

Cost. TWELVE TURNS on top of H9, being two six-turn chains with a check-writing turn, since the control arm is H9`s own out-of-coverage defects and is already paid for. The cheapest hypothesis on the site since H7, and the one with the largest consequence if it holds: a gate would be the only known way to put a floor under an unattended run.

Which verification strategy wins is decided by whether a check can be written more reliably than the work can be done twice: specify-first below an unsoundness of about the ungated density, build-twice above it, and the crossing moves with error correlation.

Requires. A stated regime, and a held-out scorer: the check used to score the artefact must be written by somebody outside all three arms, or specify-first is graded by its own homework. It also requires that the three arms be paired on task, since the run-to-run variance here is the ICC of 0.413 and unpaired arms at this size would resolve nothing.

Outcome. Surviving defect density of the integrated artefact under three assignments of the same six turns: ungated, specify-first, and build-twice with disagreement as the detector. Unit of analysis: the run, paired on task across the three assignments.

Predicts. the two versions in build-twice fail together more often than independence predicts (Knight and Leveson found this for people; one model sampled twice has less reason to differ); specify-first leaves defects the check certifies as passing (that is the term with no other source); specify-first`s advantage shrinks as the briefing improves (context and verification are substitutes, so raising lambda lowers the optimal coverage); neither beats the other everywhere (if one did, the theory has one strategy in it and not three)

Refuted by.

• One strategy wins at every setting, which would mean the crossing this theory is built around does not exist.
• The two versions in build-twice fail independently, which would contradict the only empirical anchor the strategy has and make its floor p squared after all.
• Specify-first leaves no certified defects at all, in which case unsoundness is zero, checks are free, and the interesting half of the model is idle.
• Density does not differ across arms by more than the run-to-run variation, in which case six turns cannot resolve this and the design is underpowered rather than the theory wrong.

Evidence. none recorded

Cost. EIGHTEEN TURNS PER TASK, three six-turn arms, and the pairing is what makes that affordable: at the measured rho of 0.413 the paired design needs about 40% fewer tasks than three independent arms would. The held-out scorer is the real expense and it is human or it is worthless.

A turn boundary buys independence and nothing else buys it. Subagents inside one turn share the root`s context, so they fan out at lambda 1 and their errors correlate; two turns that cannot see each other`s reasoning do not. So dividing work by ISSUE belongs inside a turn, and dividing it by GROUP is what a turn boundary is for.

Requires. Equal TOTAL budget across arms, not equal per-turn budget, or the six-turn arm gets six times the money and the comparison is about spend. The solo turn must be told it may delegate, since a turn that does not fan out is a chain of one and tests nothing. And standard`s isolation must be demonstrated, or its independence is assumed rather than shown.

Outcome. The REDUNDANCY between a run`s two checks — the share of seeded defects that both kill, of those either kills — for the same task done as one turn with subagents and as six turns in isolation, at equal total budget. Unit of analysis: the run, paired on task across the two shapes.

Predicts. solo`s two checks are more redundant than standard`s (they were written in one context, which is rho near 1 with checks in place of implementations); solo is not worse on the held-out checks (more context and a concentrated budget should if anything help — the claim is about correlation, not quality); measured coverage is similar and redundancy is not (if both move together the outcome is confounded with effort and the design says nothing); isolation is reported INAPPLICABLE for solo rather than clean (a one-turn shape has no handoffs, so it never had the opportunity to leak)

Refuted by.

• solo`s checks are no more redundant than standard`s, in which case a turn boundary buys no independence and the graph programme is measuring something that does not exist.
• standard beats solo on the held-out checks at equal budget WITH equal redundancy, which would mean something other than independence is doing the work and this decomposition has not found it.
• Redundancy tracks measured coverage across both arms, in which case the outcome is a proxy for effort and the contrast is confounded.
• The solo arm cannot spend its budget — a single turn hits the step ceiling before the work is done — in which case the arms are not matched and the result is void rather than null.

Evidence. none recorded

Cost. SIXTY DOLLARS AND TWELVE TURNS PER TASK PER REPLICATION, two arms of thirty. The bank already scores both arms for free, so the only new expense is the runs. It is the cheapest test of the programme`s central premise that exists: the whole apparatus assumes a turn boundary is worth something, and nothing has ever checked.