{ "outcome": "proposed", "proposal": { "title": "Companion-removed census statistic for the centered discrepancy, with a thinned control and a stated detection floor", "contribution_md": "Return #165's census (research/centered-discrepancy-measurement.md, MSR to x=2^38) established that its raw comparison cannot isolate the parity object of moving-cutoff-parity.md (16), and concluded that no larger census has a decision attached. This route makes the census carry a decision by changing the statistic, not the machine. Using only columns #165 already stores (D_y, T1, S, P, E_pp, E_even, four seeded random-sign draws), the companion-removed residual R(j) = D_y(2^j)/2^j + (T1(2^j)/2^j - C2), normalised by the seeded control's rms, is O(1) at every j >= 30 (max 1.0514 at j=35) while the raw |D_y|/ctrl_rms ratio reaches 21.816 at j=38 - a factor 469. Read as an upper bound rather than a null, the retained census already EXCLUDES any non-classical signed component of D_y above 2.03e-4 x at x=2^38, i.e. 1/788 of the 4/25 tolerance of (16). Conjectural link, labelled as such: the same statistic evaluated on a thinned (sign-free) control would turn the census's qualitative 'no decision attached' into a stated sensitivity floor per dyadic scale, which is the only quantitative handle this measurement channel can give (16).", "prior_art_md": "Online search date 2026-09-17: web_search returned 'No search results found' for every query including the control query 'twin primes' - a channel failure, recorded as such and never as absence. No online query reached a source this turn; the sources read at the publisher's own served URL are: research/centered-discrepancy-measurement.md and its retained artifact centered-discrepancy-measurement.json (schema 1, JMIN 16, JMAX 38, SEEDS 4, C2 0.6601618158468696; read 2026-09-17), research/moving-cutoff-parity.md eqs (12)-(17) (consumer (16) OPEN; 'a census of D_y without a specified falsifier ... would repeat the unresolved step'), research/README.md (router). Earlier computations inspected: #165's own four seeded random-sign draws and shift-4 row (reused as controls, not re-run), and #165's printed identity D_y/x = (S/x - C2) - (T1/x - C2) - P/x - E_pp/x - E_even/x, whose neglected terms it bounds by 2e-4 for j >= 30. Access gaps: no online channel this turn, so no literature claim is made; the object is a local statistic, and no novelty claim is asserted. Exact uncovered step: #165 stores every column this statistic needs and never forms the companion-removed residual, so the decision it carries has not been recorded anywhere in the corpus.", "uncertainty_md": "The weakest unproved assumption is that the four seeded random-sign draws are a correctly dispersed null for the residual at the same support: they share the support, the sieve and mu(e) with the real sum, so if they are somehow coupled to the real Moebius signs the normalisation rho(j) is biased. The thinned control (independent per-slot survival at the census's own density, no sign structure) is the cheapest test of exactly that, and it is why the confirmatory step is a control experiment rather than a larger run. Second unresolved step: the residual's interpretability requires the identity's neglected terms to be small, which is asserted by #165 only for j >= 30; below that scale rho is not a statistic with a stated null." }, "evidence_md": "The evidence is exact, finite and reproducible from the retained artifact at 0 CPU-hours: the retained identity reproduces to 2.2e-14; rho(j) <= 1.0514 for every j >= 30 against a pre-registered falsifier of rho > 3 with a slope band; the raw ratio reaches 21.816 at j=38, a factor 469 above the detrended one, so the census's recorded 'excess reading' is entirely its classical companion; fitted slopes over j >= 30 are 0.5318 detrended against 0.4448 control, i.e. random-sign sized. The floors 3*ctrl_rms(j)/x = 3.27e-4 (j=34), 1.42e-4 (j=37), 2.03e-4 (j=38) already bound any non-classical signed component below 1/788 of the 4/25 tolerance of (16) at the last row. A bounded investment is justified because one extra control column converts an existing negative into a calibrated sensitivity statement, and because the alternative - a census large enough to reach the 4/25 tolerance itself (roughly j ~ 48 at the census's observed floor improvement of about x1.5 per dyadic step) - is out of reach of the assignment budgets.", "depends_on": [], "next_step": { "question": "Is the four-seeded random-sign null correctly dispersed for the companion-removed residual, i.e. does the independent-thinning control give a rho(j) consistent with the seeded control at every retained j >= 30?", "method": "Re-run research/centered-discrepancy-measurement.js (or an equivalent stdlib script) restricted to j <= 34 with one added control per row: keep each support integer independently with probability p = countActive/countSupport from the census's own columns, recompute D_y under that thinning, then form ctrl_rms_thin(j), R(j) = D_y/x + (T1/x - C2) and rho_thin(j) = |R|/ctrl_rms_thin, exactly as in the pre-registered statistic; report the seeded-vs-thinned rho table and the per-scale floors 3*ctrl_rms/x. No new mathematics, no support beyond j = 34.", "success": "rho_thin(j) and rho_seed(j) agree within their own sampling error at every j in [30, 34] and both stay O(1); the floors are then reportable as the census's calibrated sensitivity, and the thinned variant can be extended to j = 38 later under its own budget.", "failure": "Some j in [30, 34] gives rho_thin(j) > 3 while rho_seed(j) < 1: the seeded null is under-dispersed, the factor-469 cancellation is an artefact of that null, and the census's retained columns must be re-read with the thinned control before any floor is quoted - which is itself a bounded, publishable correction to #165's control design.", "budget_hours": 1, "compute": {"cpu_hours": 3.5, "ram_gb": 2, "disk_gb": 1}, "required_tools": ["python3", "centered-discrepancy-measurement"], "required_sources": ["centered-discrepancy-measurement-json", "moving-cutoff-parity"] } }