{ "outcome": "proposed", "proposal": { "title": "Stepwise floor gate for the companion-removed census: the retained sensitivity floor is not a per-step contraction", "contribution_md": "Return #165's census (research/centered-discrepancy-measurement.md, MSR to x=2^38) stores, for every dyadic level j, four seeded random-sign draws of D_y. Return #839 (job #1629) formed the companion-removed residual R(j) = D_y/2^j + (T1/2^j - C2), normalised it by the rms of those four draws and read the quantity 3*rms_seed(D_seed/2^j) as a per-scale sensitivity floor, concluding that (a) the floor 'improves only about x1.5 per dyadic step' and (b) the 4/25 tolerance of moving-cutoff-parity.md (16) therefore needs about j ~ 48. This route changes the ingredient that claim rests on: it replaces the smooth end-to-end rate with the per-step sequence itself. Re-deriving the whole table from the retained artifact (17/17 checks, 0 CPU-h, no re-measurement) confirms every printed number of #839 exactly - rho_max 1.0514 at j=35, raw ratio 21.816 at j=38, factor 469 at the same j, decay exponents 0.5318 and 0.4448, floors 3.27e-4 / 1.42e-4 / 2.03e-4 - and finds that the floor sequence F(j) = 3*rms_seed(D_seed/2^j) is NOT monotone: the ratio F(j)/F(j+1) is BELOW 1 at five of the eight steps of j=30..38, ranging from x0.64 (a rise) to x3.75. The endpoint average over those nine rows is only x1.315, not x1.5. So 'x1.5 per dyadic step' and the j ~ 48 extrapolation are properties of a geometric-mean model fitted to the retained rows, not of the measured instrument; the extrapolation carries an unstated per-step variance, and the rise steps are exactly where a model error would bite first. The cheapest discriminating experiment is therefore not the full thinned-control re-run but a thinned control restricted to the five rise steps, which is the only place where the two readings (stepwise contraction vs geometric mean) make different predictions.", "prior_art_md": "Online search date 2026-09-17: no online channel reached a source this turn (web_search answered 'No search results found' for every query including the control query 'twin primes'; recorded as a channel failure, never as absence), so no literature claim is made. Local record read: research/centered-discrepancy-measurement.md and its retained artifact centered-discrepancy-measurement.json (schema 1, JMIN 16, JMAX 38, SEGLG 22, GRID 64, SEEDS 4, SEED 219529438, C2 0.6601618158468696; 26 rows, 23 at shift 2); research/moving-cutoff-parity.md (12)-(17), whose consumer (16) is OPEN and which states 'a census of D_y without a specified falsifier ... would repeat the unresolved step'; the shared note N-1629-01-companion-removed-census-statistic.md and return #839, whose numeric table is the object re-derived here; N-1630-01-prior-art-L-adjacent-kill-run.md for the corpus conventions. Exact difference from #839: #839 quotes a single end-to-end rate (x1.5) and derives a required level (j ~ 48) from it; this route keeps #839's statistic and falsifier (rho > 3) untouched and replaces only that rate by the step sequence, showing the sequence is non-monotone and that the endpoint average is x1.315. Nothing here contradicts #839's numbers; the difference is that the extrapolation is now an assumption with a testable step-level consequence rather than a reported measurement.", "uncertainty_md": "The weakest unproved assumption is that the five rise steps are structure and not the sampling noise of a four-draw rms. With only four seeded draws per row, rms_seed has a large relative error (for a normal sample the rms of 4 draws has roughly 35 percent relative standard deviation), which is enough to produce ratios of 0.64 and even 0.70 by chance at some steps. If the entire spread of the step ratios is this sampling error, then the stepwise reading adds nothing over #839's endpoint rate and this route collapses - which is exactly what the restricted thinned control decides, cheaply and in one run. Second unresolved step: even if the ratio sequence is real, mapping a sensitivity floor onto the 4/25 tolerance of (16) still needs the neglected terms of the identity to stay small at the level quoted, which is asserted for j >= 30 only. Third: this note numbers no claim of novelty - the arithmetic is elementary and the object is local, so the value is in the corrected reading of an existing measurement, not in a new theorem." }, "evidence_md": "Evidence is exact, finite and reproducible from the retained artifact at 0 CPU-h: work/job1631/src/job1631-checks.py, 17 checks, 17 PASS, log work/job1631/src/job1631-checks.log. Re-derived and confirmed from centered-discrepancy-measurement.json shift-2 rows: the printed identity D_y/x = (S/x - C2) - (T1/x - C2) - P/x - Epp/x - Eeven/x to 2.15e-14; max rho 1.0514 at j=35; raw |D_y|/x over ctrl rms 21.816 at j=38; raw(38)/rho(38) = 469.1; decay exponent of |R| 0.5318 and of the control rms 0.4448 (the note prints these as positive slopes; fitting ln(value) against j gives the same magnitudes negated, recorded as a reading); raw slope -0.0014; floors at j=34, 37, 38 of 3.268e-4, 1.420e-4, 2.029e-4, the last equal to 1/788.5 of 4/25. NEW in this return: the per-step floor ratios over j=30..38 are 0.77, 3.75, 0.64, 2.99, 0.74, 0.91, 3.42, 0.70 - five of eight are rises - the endpoint average is x1.315, and note that #839's own floor at j=31 (2.35e-3) is HIGHER than at j=30 (1.82e-3), so the non-monotonicity is visible in its printed table without any new computation. The proposed experiment is bounded: a thinned control at the five rise steps only, at the census's own density, is about 40 percent of the 3.5 CPU-h full re-run.", "depends_on": [], "next_step": { "question": "Is the non-monotone ratio sequence of the seeded floor real structure, or the sampling error of an rms over four draws - i.e. does an independent-thinning control reproduce the rise steps 30, 32, 34, 35, 37 with ratios below 1?", "method": "Re-run research/centered-discrepancy-measurement.js (or an equivalent stdlib script) at the five rise levels only (j = 30, 32, 34, 35, 37) plus their successors, adding one thinned control per row: keep each support integer independently with probability p = countActive/countSupport from the census's own columns, recompute D_y and the companion-removed residual under that thinning, and form F_thin(j) = 3*rms_thin(j) exactly as in the retained definition. Report the paired ratio table F(j)/F(j+1) for seed and thin, and the ratio of the two rms values per row. No new mathematics and no support above j = 38.", "success": "The thinned control gives ratios below 1 at the same five steps and its rms agrees with the seeded rms within the four-draw sampling error at every paired row: the non-monotone floor is then a property of the instrument, #839's x1.5 and j ~ 48 become geometric-mean extrapolations with a reported per-step variance, and the stepwise gate is the calibration to quote.", "failure": "The thinned control gives ratios above 1 at the rise steps, or its rms differs from the seeded rms by much more than the four-draw sampling error: the rise steps are sampling noise, the ratio sequence is not a contractible gate, and #839's end-to-end rate should be quoted as the only reading this channel supports - itself a bounded correction to N-1629-01 that is publishable at the same 0.5 CPU-h.", "budget_hours": 0.5, "compute": { "cpu_hours": 1.5, "ram_gb": 2, "disk_gb": 1 }, "required_tools": [ "python3", "centered-discrepancy-measurement" ], "required_sources": [ "centered-discrepancy-measurement-json", "moving-cutoff-parity" ] } }