Skip to content

Design-guidance paper-grade slice — findings report (2026-10-29)

Change: design-guidance-paper-grade-slice (MOTCO 0.12.0). Supersedes at its columns: the pilot-grade rates of the design-guidance pilot (2026-10-28) at n ∈ {80, 200, 500}, three stages, two or three blocks, and at the joint stage effect δ ∈ {0.25, 0.5, 1.0}; the pilot's four-stage, n = 1000 and remaining arm columns stand at pilot grade. Reads: results/design-guidance-slice-2026-10-29/ (report/ and PROVENANCE.txt per profile, figures/, reproduction/).

The slice reruns the pilot report §7 proposal at paper grade: the three-stage core at the sample sizes the planned analyses have, under the production measurement (pooled PLS with the stage-supervised double-CV rank floored at 2, net-displacement orientation, joint magnitude), 500 replicates × 999 permutations (SE ≤ 0.022 on every rate; anchor bound 0.0695 at R = 500), in the pilot's seed family so that every column evaluates the pilot's datasets at its first 100 replicates. It is guidance, not a gate: the predeclared rules are the advisory anchor decision and a column-scoped Type I rule for translation's angle; there is no power target. The second-priority stage-effect arm (§8) was run on the user's decision once the slice's numbers were in.

In one paragraph. Size holds at every column under both predeclared rules. delta is the component a small cohort detects: 0.62 at n = 80 and 0.94 at n = 200 at e = 0.1 with two blocks (0.84 and 0.99 with three), and 0.63 at n = 500 even at e = 0.02 with two blocks. shape is monotone in n, effect and block count: ≤ 0.20 at n = 80 with two blocks, 0.56 at n = 200 and 0.85 at n = 500 at e = 1.0, and 0.48 / 0.86 / 0.97 with three blocks. angle is governed by n and block count with a flat, non-monotone response in effect (0.10–0.15 at n = 80 and 0.16–0.24 at n = 200 with two blocks; 0.36–0.43 at n = 500; 0.24–0.31 / 0.41–0.49 / 0.66–0.74 with three), every reading at a CV-selected rank of 2 or 3. All of it at a stage signal that is saturated (held-out CV AUROC ≥ 0.9998 at every cell), which is what the stage-effect arm relaxes.

1. Configuration and provenance

core magnitude core geometry stage-effect magnitude stage-effect geometry
config design_guidance_core_magnitude_study.json design_guidance_core_geometry_study.json design_guidance_stage_effect_magnitude_study.json design_guidance_stage_effect_geometry_study.json
config sha256 (first 16) 3f4a618d57fc59db bb56f1c6f61ce2e3 2cd406d7d6238ff0 961bac86a2e2a6c8
design columns n {80, 200, 500} × 3 stages × {2, 3} blocks (6) same (6) joint δ {0.25, 0.5, 1.0} at n = 200, 3 stages, 2 blocks (3) same (3)
modes × effects magnitude × {0.02, 0.05, 0.1} orientation, shape, translation × {0.25, 0.5, 1.0} magnitude × {0.02, 0.05, 0.1} orientation, shape, translation × {0.25, 0.5, 1.0}
cells / units 26 / 13,000 62 / 31,000 14 / 7,000 32 / 16,000
SLURM jobs (shards) 929844 + 929980 (40) 929845 (95) 930041 (20) 930042 (46)

All four profiles: 500 replicates × 999 permutations, base_seed 1028, matched-seed family design-guidance, evaluation.seed 3 (also the RRPP seed), orientation_estimator net_displacement, min_components 2, magnitude_kind joint, the Phase 6 CV knobs (cv1 3, cv2 4, 5 repeats, ≤ 20 components), attribution and gate off, surgery_censoring error. The slice baseline is the pilot's (n = 200, three stages, group_ratio 0.5, p_dmp 0.1, ρ = 0, every per-omic δ 2.0, methylation + expression measured); the arm's configuration baseline is δ = 0.25 (§8). Code revision 8bf8eb7 on feat/design-guidance-paper-grade-slice (clean worktree /home1/tgonza/MOTCO-design-guidance-slice on ing, the main checkout's .venv with PYTHONPATH on the worktree's src; the venv's motco distribution metadata says 0.11.0 and the revision is the version of record); python 3.11.16, numpy 2.3.5, scikit-learn 1.8.0, scipy 1.16.3. Partition 512x1024, 1 CPU and 4 GB per task, 6 h limit, BLAS pinned to one thread, STUDY_N_JOBS unset, --error-policy record. Every record carries n_jobs 1.

Pre-flight: the fast suite at the same tree (ruff, mypy, 1583 passed / 1 skipped); the change's six test modules on ing from the worktree (115 passed); a 2 × 19 smoke of all four profiles (268 / 268 units, 0 failures); 14 units timed at 999 permutations on ing, on a quiet node (8–19 s per unit across the six columns) and on a saturated one (16–33 s) — 999 permutations cost no more than the pilot's 199 did, because the PLS double CV dominates a unit. Shards were sized at the saturated cost plus 20 % at ≤ 2.5 h each. Thirty-five magnitude shards that SLURM placed on an overloaded node sat in NFS wait for 32 minutes without writing a unit; they were cancelled and the same array ids resubmitted elsewhere (job 929980), which changes nothing in the records (shards are keyed by the (cell_id, replicate_index) hash and are resumable). Full details: openspec/changes/design-guidance-paper-grade-slice/notes.md.

2. Unit, failure and reproduction accounting

Profile Array tasks Records (expected) Duplicates / missing / signature mismatches Failed Censored Recorded core-hours (median unit) Wall
core magnitude 40 COMPLETED 0:0 (+ 35 cancelled and resubmitted before writing) 13,000 (13,000) 0 / 0 / 0 0 0 52.7 (11.9 s) 11:42 → 14:27
core geometry 95 COMPLETED 0:0 31,000 (31,000) 0 / 0 / 0 0 0 119.0 (12.1 s) 11:42 → 13:29

Every record carries n_jobs 1, orientation_estimator net_displacement and min_components 2. The slice's 44,000 units cost 171.7 recorded core-hours, between the quiet-node (151) and saturated-node (278) estimates and below the design's 380. The CV selected rank 2 in 35,000 of the 44,000 units and rank 3 in the other 9,000; never more.

Reproduction (scripts/anchor_reproduction.py --by-column, reproduction/):

  • Siblings (siblings.csv): the magnitude and geometry profiles' zero-effect anchors pair at every one of the six columns, 500 of 500 replicates each, 3,000 pairs: generator seeds, selected ranks, observed delta/angle/shape (max |Δ| ≤ 5 × 10⁻¹¹) and p-values identical on all 3,000. The report reads the geometry profile's anchor; the magnitude profile's is this check.
  • Pilot (pilot_core_magnitude.csv, pilot_core_geometry.csv): against the pilot core profiles' merged.jsonl (199 permutations), every slice column pairs replicate indices 0–99 — 600 pairs per profile — with seeds, ranks and observed statistics identical (max |Δ| ≤ 4 × 10⁻¹⁰); p-values are reported as not comparable and excluded from the verdict; the pilot's ten other columns (four stages; n = 1000) are listed with zero pairs. The slice therefore evaluates the pilot's datasets at its first 100 replicates and 400 fresh ones; it supersedes the pilot's rates at these columns and never pools the two.

3. Per-column operating characteristics

Rates are rejection rates at α = 0.05 over 500 available replicates, ± 1 Monte Carlo SE. The figures are figures/design_guidance_{magnitude,orientation,shape,control}.png (scripts/design_guidance_figure.py --grade paper, one panel per block count at three stages, n on a log axis, one line per effect) with the plotted rows in figures/design_guidance_power.csv; the full tables, with the stage signal and selected rank per cell, are results/design-guidance-slice-2026-10-29/tables.md.

3.1 Magnitude / delta

column e = 0.02 e = 0.05 e = 0.1
n = 80, 2 blocks 0.036 ± 0.008 0.234 ± 0.019 0.620 ± 0.022
n = 80, 3 blocks 0.112 ± 0.014 0.512 ± 0.022 0.836 ± 0.017
n = 200, 2 blocks 0.228 ± 0.019 0.694 ± 0.021 0.936 ± 0.011
n = 200, 3 blocks 0.430 ± 0.022 0.878 ± 0.015 0.992 ± 0.004
n = 500, 2 blocks 0.632 ± 0.022 0.962 ± 0.009 0.998 ± 0.002
n = 500, 3 blocks 0.834 ± 0.017 0.994 ± 0.003 1.000 ± 0.000

Monotone in n, effect and block count at every column. The pilot's values at these 18 cells are reproduced within 0.04 (mean difference +0.004), inside the pilot's SE of ≤ 0.05.

3.2 Orientation / angle and shape / shape

column angle e = 0.25 angle e = 0.5 angle e = 1.0 shape e = 0.25 shape e = 0.5 shape e = 1.0
n = 80, 2 blocks 0.100 ± 0.013 (rank 3) 0.110 ± 0.014 (3) 0.150 ± 0.016 (3) 0.044 ± 0.009 0.104 ± 0.014 0.204 ± 0.018
n = 80, 3 blocks 0.238 ± 0.019 (2) 0.312 ± 0.021 (2) 0.274 ± 0.020 (3) 0.184 ± 0.017 0.316 ± 0.021 0.484 ± 0.022
n = 200, 2 blocks 0.158 ± 0.016 (2) 0.242 ± 0.019 (3) 0.184 ± 0.017 (3) 0.236 ± 0.019 0.418 ± 0.022 0.562 ± 0.022
n = 200, 3 blocks 0.406 ± 0.022 (2) 0.494 ± 0.022 (2) 0.492 ± 0.022 (2) 0.486 ± 0.022 0.668 ± 0.021 0.856 ± 0.016
n = 500, 2 blocks 0.364 ± 0.022 (2) 0.426 ± 0.022 (2) 0.358 ± 0.021 (3) 0.548 ± 0.022 0.750 ± 0.019 0.848 ± 0.016
n = 500, 3 blocks 0.662 ± 0.021 (2) 0.738 ± 0.020 (2) 0.710 ± 0.020 (2) 0.780 ± 0.019 0.892 ± 0.014 0.970 ± 0.008

The number in parentheses is the median CV-selected rank at that cell (the range is 2–3 at every cell of the slice). shape is monotone in n, effect and block count. angle rises with n and with the third block but not with the effect: at every column the e = 1.0 rate is within 0.08 of the e = 0.25 rate and below the e = 0.5 rate at four of six columns, the pattern the pilot saw at pilot grade, now resolved at SE ≤ 0.022. Against the pilot the orientation rates are reproduced within 0.11 (mean +0.044, the largest differences at n = 200 and 500 with three blocks at e = 1.0, where the pilot's 100 replicates read 0.38 and 0.61 and the slice reads 0.49 and 0.71) and the shape rates within 0.12 (mean +0.016).

Translation (negative control), all statistics at every effect

column e delta angle shape column-scoped verdict (angle at e = 1.0 vs 0.0695)
n = 80, 2 blocks 0.25 0.020 ± 0.006 0.074 ± 0.012 0.020 ± 0.006
0.5 0.034 ± 0.008 0.078 ± 0.012 0.020 ± 0.006
1.0 0.032 ± 0.008 0.040 ± 0.009 0.024 ± 0.007 holds
n = 80, 3 blocks 0.25 0.044 ± 0.009 0.064 ± 0.011 0.062 ± 0.011
0.5 0.058 ± 0.010 0.056 ± 0.010 0.044 ± 0.009
1.0 0.046 ± 0.009 0.028 ± 0.007 0.048 ± 0.010 holds
n = 200, 2 blocks 0.25 0.042 ± 0.009 0.054 ± 0.010 0.040 ± 0.009
0.5 0.058 ± 0.010 0.054 ± 0.010 0.050 ± 0.010
1.0 0.050 ± 0.010 0.052 ± 0.010 0.032 ± 0.008 holds
n = 200, 3 blocks 0.25 0.046 ± 0.009 0.052 ± 0.010 0.042 ± 0.009
0.5 0.062 ± 0.011 0.050 ± 0.010 0.056 ± 0.010
1.0 0.048 ± 0.010 0.044 ± 0.009 0.050 ± 0.010 holds
n = 500, 2 blocks 0.25 0.042 ± 0.009 0.054 ± 0.010 0.068 ± 0.011
0.5 0.042 ± 0.009 0.048 ± 0.010 0.050 ± 0.010
1.0 0.056 ± 0.010 0.048 ± 0.010 0.042 ± 0.009 holds
n = 500, 3 blocks 0.25 0.050 ± 0.010 0.048 ± 0.010 0.054 ± 0.010
0.5 0.036 ± 0.008 0.046 ± 0.009 0.052 ± 0.010
1.0 0.050 ± 0.010 0.032 ± 0.008 0.052 ± 0.010 holds

Translation runs at all three geometry effects because a configuration has one effect list (design D2); the predeclared rule reads e = 1.0 and the two smaller offsets are descriptive. No translation rate exceeds 0.078 anywhere. The two readings above the bound are angle at the two smaller offsets at n = 80 with two blocks (0.074 and 0.078, SE 0.012); at the same column the e = 1.0 rate is 0.040 and the anchor's 0.036.

4. Stage signal and selected rank

The held-out CV stage AUROC (design_point_stage_signal.csv, the number that compares with a real analysis) is ≥ 0.9998 in mean and 1.000 at the q33–q67 range at every cell of both profiles: the stage signal is saturated at the core stage effect δ = 2.0, as in the pilot. The CV selects rank 2 or 3 at every cell (median 2 at 36 of the 42 geometry cells and 3 at the others, the n = 80 / two-block column and the e ≥ 0.5 orientation cells at n = 200 and 500 with two blocks among them; never above 3). With a saturated stage signal and two or three retained components, orientation is measured in a two- or three-dimensional latent space whose second axis carries little signal, and angle power is governed by how much of the group difference that second axis resolves — by n and the block count — rather than by the size of the orientation surgery (§9).

5. Size at every column

Anchor decision (anchor_decision.json; zero-effect anchor, R = 500, bound 0.0695; the geometry profile's anchor, the magnitude profile's being its bit-exact reproduction):

column delta angle shape verdict
n = 80, 2 blocks 0.022 ± 0.007 0.036 ± 0.008 0.010 ± 0.004 size_holds
n = 80, 3 blocks 0.052 ± 0.010 0.032 ± 0.008 0.046 ± 0.009 size_holds
n = 200, 2 blocks 0.040 ± 0.009 0.060 ± 0.011 0.034 ± 0.008 size_holds
n = 200, 3 blocks 0.038 ± 0.009 0.060 ± 0.011 0.062 ± 0.011 size_holds
n = 500, 2 blocks 0.040 ± 0.009 0.050 ± 0.010 0.040 ± 0.009 size_holds
n = 500, 3 blocks 0.042 ± 0.009 0.046 ± 0.009 0.040 ± 0.009 size_holds

Every statistic holds at every column; the largest anchor rate is 0.062 (shape, n = 200, three blocks) against 0.0695. Chance predicts about 0.4 excesses over these 18 tests; none occurred. The pilot's two excesses (both 0.10 against 0.0936 at R = 100) were at columns outside the slice.

Column-scoped Type I rule (acceptance_report.csv, type_i_control[translation@1.0, angle, <column>], upper bound at α 0.05 and k 2): translation's angle at e = 1.0 is met at all six columns of the geometry profile, at 0.028–0.052 against 0.0695 (the table in §3.2). In the magnitude profile the same entry reports every column unavailable, as predeclared: translation runs only in its Type I baseline phase, and the entry is declared so that the sibling acceptance blocks are identical.

Type I baseline phase (the two cells keyed on their own ids, a fresh 500-replicate draw at the baseline column, two-sided entry at tolerance 2): the shared anchor rejects delta 0.048, angle 0.050, shape 0.042 (identical in both profiles — the anchor cell is the same cell); translation at the profile's largest effect rejects 0.040 / 0.038 / 0.028 at e = 0.1 (magnitude profile) and 0.056 / 0.056 / 0.060 at e = 1.0 (geometry profile). The two-sided entry flags the magnitude profile's translation shape 0.028 as not met — a conservative rate two empirical SEs below α, the known property of the two-sided form, not an excess; every one of these rates is under the upper bound.

6. Reproduction checks

Stated in §2: the sibling profiles reproduce each other's anchors at every column with identical p-values (3,000 / 3,000), and each slice profile reproduces the pilot's datasets at replicate indices 0–99 at every column (600 / 600 per profile, seeds and observed statistics identical, p-values not comparable across 999 and 199 permutations and excluded from the verdict). Pilot and slice replicates are never pooled.

7. Design-guidance statement

For a two-group, three-stage design measured as the production pipeline measures it, at the core stage effect (δ = 2.0; a saturated stage signal) and with the structural-analogue blocks:

  • Size. Every statistic holds its nominal size at every sample size and block count (anchor ≤ 0.062 against 0.0695), and the location-only control holds the angle bound at every column at e = 1.0. A rejection beside a location difference still warrants care at n = 80 with two blocks, where the smaller location offsets gave angle 0.074–0.078.
  • delta (magnitude) is the component a small cohort detects. With two blocks: e = 0.1 at 0.62 (n = 80), 0.94 (n = 200) and 1.00 (n = 500); e = 0.05 at 0.23, 0.69 and 0.96; e = 0.02 at 0.04, 0.23 and 0.63. A third block raises every rate: 0.84 / 0.99 / 1.00 at e = 0.1, 0.51 / 0.88 / 0.99 at e = 0.05, 0.11 / 0.43 / 0.83 at e = 0.02.
  • shape needs n or blocks. With two blocks it stays ≤ 0.20 at n = 80, reaches 0.42–0.56 at n = 200 (e = 0.5–1.0) and 0.75–0.85 at n = 500; with three blocks 0.32–0.48 at n = 80, 0.67–0.86 at n = 200 and 0.89–0.97 at n = 500. A shape non-rejection at n = 80 is uninformative.
  • angle (orientation), every reading at a selected rank of 2–3: 0.10–0.15 at n = 80 and 0.16–0.24 at n = 200 with two blocks (a non-rejection is uninformative there), 0.36–0.43 at n = 500 with two blocks; 0.24–0.31, 0.41–0.49 and 0.66–0.74 with three blocks. The rate does not increase with the orientation effect between e = 0.25 and 1.0; what moves it is n and the block count. Report the cohort's selected rank beside any angle result.
  • Blocks. Measuring the third block raises delta, shape and angle power at every n — by 0.2–0.3 at the mid-range cells — at no cost in size; block count is confounded with feature count (§9).

None of this is claimed below the core stage effect; the stage-effect arm (§8) measures how the statement degrades as δ falls toward the SEA-AD regime.

8. Stage-effect arm

Run on the user's decision on 2026-10-29, once the geometry slice's numbers were in (jobs 930041 and 930042, 20 + 46 shards submitted at 13:42 with the overloaded node excluded; both arrays COMPLETED 0:0; 7,000 and 16,000 records; 0 duplicates, missing records, signature mismatches, failures or censored surgeries; 68.1 recorded core-hours at a median 10.5–10.7 s per unit; PROVENANCE.txt per profile). The arm is the slice profile at n = 200, three stages and two blocks with the joint per-omic stage effect delta_methyl = delta_expr = delta_protein at δ ∈ {0.25, 0.5, 1.0} as explicit design points. δ = 0.25 is its configuration baseline — so no column repeats the slice, whose n = 200 / two-block column is the δ = 2.0 reference in the table — and its Type I baseline phase is a fresh 500-replicate draw at δ = 0.25.

δ CV stage AUROC (anchor; orientation e = 1.0) selected rank, median (range): anchor; orientation e = 1.0 delta e 0.02 / 0.05 / 0.1 angle e 0.25 / 0.5 / 1.0 shape e 0.25 / 0.5 / 1.0 anchor delta / angle / shape translation angle e = 1.0
0.25 0.997; 0.950 11 (6–19); 16 (5–19) 0.050 / 0.056 / 0.084 0.064 / 0.088 / 0.126 0.048 / 0.054 / 0.046 0.052 / 0.058 / 0.062 0.054 holds
0.5 1.000; 1.000 4 (3–6); 9 (6–19) 0.068 / 0.110 / 0.310 0.074 / 0.110 / 0.462 0.036 / 0.052 / 0.066 0.046 / 0.050 / 0.038 0.048 holds
1.0 1.000; 1.000 3 (2–3); 4 (3–5) 0.058 / 0.196 / 0.508 0.074 / 0.070 / 0.470 0.070 / 0.100 / 0.168 0.028 / 0.048 / 0.028 0.048 holds
2.0 (slice, §3) 1.000; 1.000 2 (2–3); 3 (2–3) 0.228 / 0.694 / 0.936 0.158 / 0.242 / 0.184 0.236 / 0.418 / 0.562 0.040 / 0.060 / 0.034 0.052 holds

Rates ± ≤ 0.022 (R = 500); the full tables with every cell's AUROC and rank are in tables.md. The median rank at the angle cells is 12 / 13 / 16 at δ = 0.25, 4 / 5 / 9 at δ = 0.5 and 3 / 3 / 4 at δ = 1.0 for e = 0.25 / 0.5 / 1.0.

  • Size holds at every δ. Anchor ≤ 0.062 (the δ = 0.25 shape) against 0.0695 and size_holds at all three points; translation's angle at e = 1.0 meets the column-scoped rule at 0.048–0.054; no translation rate at any effect exceeds 0.062. The Type I baseline phase at δ = 0.25 gives anchor 0.026 / 0.062 / 0.036 and translation 0.036 / 0.056 / 0.034 (e = 0.1, magnitude profile) and 0.028 / 0.054 / 0.060 (e = 1.0, geometry profile) — all under the upper bound; the two-sided entry flags the conservative 0.026 and 0.028 rates, not an excess.
  • The stage signal stays strong down to δ = 0.25 (held-out CV AUROC 0.997, against 0.64–0.70 in the SEA-AD cohorts and 0.55 / 0.73 at the pilot's δ = 0.05 / 0.1), so the arm covers the upper part of the transition; the orientation surgery at e = 1.0 is the only cell that erodes it (0.950 at δ = 0.25).
  • The CV rank inflates as δ falls: a median 11 components at δ = 0.25 (up to 19), 4 at δ = 0.5, 3 at δ = 1.0 and 2 at δ = 2.0; at the orientation e = 1.0 cells 16, 9, 4 and 3.
  • delta and shape lose their power with δ. At e = 0.1 delta falls from 0.94 (δ = 2.0) to 0.51, 0.31 and the null (0.08); e ≤ 0.05 is at the null for δ ≤ 0.5 and 0.20 at δ = 1.0 and e = 0.05. shape is at the null for δ ≤ 0.5 and reaches 0.17 at δ = 1.0 and e = 1.0 against 0.56 at δ = 2.0.
  • angle does not follow δ monotonically. At e ≤ 0.5 it sits at the null for every δ ≤ 1.0 (0.06–0.11), but at e = 1.0 it reads 0.46 and 0.47 at δ = 0.5 and 1.0 — above the 0.18 at δ = 2.0 — with ranks 9 and 4 against 3. This is the rank confound of §9 in the other direction: a weaker stage signal makes the CV retain more components, which gives the largest orientation surgery room to be resolved, while a saturated signal collapses the latent space to two or three stage-driven axes. At δ = 0.25 (rank 16) the e = 1.0 rate is 0.13.
  • Against the pilot arms (100 × 199) the arm's rates are within 0.06 (delta), 0.04 (angle) and 0.03 (shape) at every cell and the median ranks agree; the arm supersedes them at these three columns.

Reproduction. The two arm profiles' anchors reproduce each other at all three columns, 1,500 / 1,500 pairs with identical p-values (reproduction/siblings_stage_effect.csv). Against the pilot arms profiles the committed --by-column tool pairs nothing (pilot_stage_effect_*.csv, 0 pairs): it keys a column on its full design point, and the pilot arms' points carry seven axes (the other one-factor arms at their baseline values) against the arm's three. The seed identity holds nonetheless — the profile tests assert it at enumeration for every arm cell (replicates 0–99) — and a projected join of the records on (δ, replicate index), restricted to the pilot columns whose other axes sit at the baseline, confirms it on the data: 300 / 300 pairs per profile (100 at each δ), with generator seeds, selected ranks and observed delta/angle/shape identical on every pair (reproduction/pilot_stage_effect_*_projected.csv; p-values not comparable across 999 and 199 permutations). An option to project the column key onto named axes is the natural follow-up for the tool.

Guidance below the core stage effect (n = 200, three stages, two blocks): at δ = 0.25 no component has power at any effect in the enumerated ranges except a weak e = 1.0 orientation reading (0.13 at rank 16); at δ = 0.5–1.0 only delta at e = 0.1 (0.31–0.51) and angle at e = 1.0 (0.46–0.47, at ranks 9 and 4) leave the null; shape needs the core stage effect. A real design's held-out stage AUROC places it on this axis: 1.000 reads as δ ≥ 0.5, 0.997 as δ ≈ 0.25, and the SEA-AD level (0.64–0.70) sits below the arm, where the pilot measured no power for any component at δ ≤ 0.1.

9. Limitations

  • Translation at three effects. A configuration has one effect list, so translation ran at e ∈ {0.25, 0.5, 1.0}; the predeclared rule judges e = 1.0 only, and the two smaller offsets are descriptive controls (above the bound for angle at n = 80 with two blocks). A per-mode effect list in the engine is the documented alternative.
  • Block count is confounded with feature count (498 measured features with two blocks against 498 plus the protein block with three): the block effect is a measurement contrast on the same datasets, not a claim about protein data.
  • Structural analogue. The blocks are InterSIM-calibrated structural analogues of methylation, expression and protein data; the rates transfer to a real design only through the stage signal and the feature counts, not through any biology.
  • Saturated stage signal. At δ = 2.0 the held-out CV stage AUROC is 1.000 everywhere, far above the SEA-AD cohorts (0.64–0.70); the statement in §7 holds for a strongly staged design, and §8 is the only part of this report that speaks to weaker staging.
  • The rank confound of orientation power. angle is measured in the CV-selected latent space (rank 2–3 here) and its power follows n and the block count rather than the orientation effect; a flat or non-monotone response in e is a property of the measurement, not evidence that larger orientation differences are harder to detect. The latent-rank ladder (2026-09-08) and the pilot report §4 document the mechanism; every angle reading above carries its rank.
  • Scope. Three stages, ρ = 0, equal groups, p_dmp 0.1, n ≤ 500. The pilot's four-stage, n = 1000, group-ratio, stage-proportion, p_dmp and ρ columns stand at pilot grade (100 × 199) and are not rerun here. Group-specific dispersion is not in the generator (change B of the pilot's follow-ups).