SRC-m: square-root cancellation of the (D1) m-side pair sum, with an explicit ledger margin

Investment state: active. This describes research progress; claims have separate evidence grades.

Contribution to the goal

If SRC-m holds, the (D1) m-side row stops landing exactly on the demand and starts carrying margin.

Measured basis (this job, on return #5863's own 13 points plus a negative control): the linear m-sum sum_m K = 1^T M 1 is at the generic incoherent pair size 2|H|sqrt(Mdist) — the ratio |K|/(2|H|sqrt(Mdist)) is 0.55..2.18 with log-log slope -0.11 against |H| — while every norm of the pair kernel pays the factor |H| (operator loss 2.6..30.4, slope +0.48; Frobenius slope +0.99; all-ones alignment down to 0.033). So the |H|-gap #5863 attributes to Cauchy-Schwarz in m is intrinsic to the kernel, and only a signed estimate can produce the margin.

Substituted into #5863's recorded ledger identity (per-class = x^{14/25} -> moment 139/100 -> block 400/400, zero margin), the generic exponent x^{17/50} = 2x^{3/50}x^{7/25} for 2|H|sqrt(Mdist) gives moment 117/100 and block 356/400, a margin of 44/400. That is a change of ingredient, not a re-derivation: #5863 normalises against Mdist, this route normalises against the pair sum.

Downstream use, labelled conjectural: a proven SRC-m with an absolute constant would move route 337's block exponent off its zero-margin reading at the route's own exponents (|H| = x^{3/50}, M = x^{14/25}, c = x^{19/20}), and would make job #11882's frequency second moment unnecessary for the m-side rows. It does not by itself close the 7/400, and it says nothing about the t-side dual windows the closed-routes register leaves open.

Prior work and proposed difference

Prior art — job #12234 (route 395 step check)

No new literature search was run: this return makes no new mathematical claim, only a comparison of the held step with returns already on record. The relevant prior art is the route's own, and it is reproduced here so a successor does not have to re-derive it.

Route 395's recorded prior art (from the route record, #5865's search, summarised).

  1. Bilinear Kloosterman / modular-inverse forms in short intervals. Kerr–Shparlinski–Wu–Xi (2023, JLMS 108, arXiv:2204.05038) bound Type-I/II bilinear forms sum_{h<=H} sum_{m<=M} alpha_h beta_m e_c(a h mbar) by Cauchy–Schwarz in m followed by k-fold completion. CS in m places absolute values inside the m-sum and forfeits the signed |H|-cancellation — which is exactly why the corpus's Mode I/II block exponent is 412/400 (#5859, #5865).
  2. Completion / Pólya–Vinogradov. The classical incomplete-sum bound |sum_{m<=M} e_c(am)| <= min(M, sqrt(c) log c) gives nothing below sqrt(c); the completed form replaces the m-sum by a t-sum of length c (or, after localisation, c/M). #714's job1506-T791-dual-length-verdict.md is the corpus's exact accounting of that dual length: mass concentrated at |t| ~ c/M, support all of Z/cZ, tail mass cM/lambda (a fixed fraction), so windowing with exact l2 norms — not a single truncated range — is the legitimate use.
  3. Fouvry–Kowalski–Michel (2014) and Pascadi (Cor. 7.9 / Cor. 8.1 application) are the corpus's other two completion routes; both are recorded as net 0 or as requiring the short-window mechanism (#5865 §"five rows forbid").
  4. AC-ledger context. #5863 measured the linear (D1) m-sum at Theta(Mdist) on 13 points; #5865's r_src statistic and route 395 are the response; #5880 is the route's first execution.

What this return adds to the prior-art picture. Only a calibration point: the 0.15 slope bar that the step inherits from #5880 was set in the log Mdist coordinate, and in the step's own log rho_PV coordinate it is twice as strict (measured ratio ≈ 2.0 on #5880's rows). Nothing in the literature bears on that; it is a property of the design, and the successor must fix the statistic in advance.

Deliberately not searched. No arXiv/OEIS/literature query was issued for "Pólya–Vinogradov threshold sweep" or "single-class vs divisor-slice completion", because the step does not create a claim that would need one, and the cheapest way to refute the step is the recorded comparison above.

Central uncertainty

The weakest unproved step is the premise, not the algebra: that Theta(Mdist) at #5863's 13 points is a shape artefact (all of them sit at Mdist ~ 2.86|H|^2, where the generic size 2|H|sqrt(Mdist) is itself Theta(Mdist)) rather than the true scale at the route's exponents. The same 13 points are consistent with both readings to within a factor ~2, so they cannot decide it.

Second: no norm on the m-side can supply the bound (measured here: operator slope +0.48, Frobenius +0.99, alignment 0.033), and the pair-level phase data shows no exploitable structure (coherence at the iid baseline, ratio 0.61..2.48; the only exact resonance family l2h' = l1h carries 3.4%..17.3% in the 4 corpus blocks and is absent from all 9 probe rungs). So SRC-m requires genuine signed cancellation in m, and the standard theory of incomplete Kloosterman sums gives nothing below sqrt(c) — the register's own line ("below c/M the trivial bound already suffices") points the same way.

Third: the x^{17/50} -> 356/400 arithmetic is a substitution into a recorded ledger identity from 13 constructed points, not a theorem; the polylog in SRC-m is unspecified, and the constant 2 in the normalisation is an amplitude convention (|a_h|, |b_h| <= 2), so the ratio is a trend, not a calibrated constant. A computation at larger x could move the measured band without touching the mechanism either way.

Next experiment

Is the square-root-cancellation ratio r_src = |sum_{m in I_m} K(m)| / (2|H| sqrt(Mdist)) flat in the completion ratio rho_PV = Mdist/sqrt(c) when only the modulus moves (M = 100|H|^2 fixed, p_q the only knob), for 1.2 <= rho_PV <= 6, and does the single (l1,l2) class suffice there?

Extend the served src_m_11890.py (sha256 a5ea5365…) rather than rewriting it: keep its exact geometric-series evaluator, the Gate V1/V2 anchor checks and the Gate V3 amplitude diagnostics. Fix mu = 100 so M = 100|H|^2 at |H| in {16, 32, 64}; scale x = round(x0*(c/c0)(M/M0)) so that sqrt(c) << M and A'/(u_i M) = Theta(1) as Gate V3 requires (x must move with c; leaving it fixed collapses mean_m|K|^2 to <1e-5, which #5880 measured); vary only the prime p_q in q = 2 p_q so that rho_PV = Mdist/sqrt(c) traverses {1.2, 1.5, 2.0, 2.6, 3.4, 4.5, 6.0}. At each rung record r_src for the w=0 window, the 4-subwindow RMS ensemble and the 4-(l1,l2)-slice average, plus 2M/c, mean_m|K|^2/|H|^2 and sqrt(c) / M as guards. PRE-REGISTER the statistic before the run: the 4-subwindow RMS (it is the only recorded ensemble that clears the inherited bar at all four #5880 scales; the single-window reading doubles in this coordinate). Keep #5865's short-window family (rho_PV ~ 0.1) as a labelled control only. Do NOT compute the |t| <= c/M dual-completed sum: #714 §1 proves the mass beyond lambdac/M is cM/lambda, a fixed fraction for every fixed lambda, so that truncation is a constant-factor approximation, not a match. If a dual comparison is wanted, use #714 §2's dyadic t-windows, each with its own span and exact per-window l2 norm (span x norm is k-uniform), and compare window-summed physical vs dual.

Required evidence

Unaccepted premises remain conditional.

Evidence behind continued investment

These investigations led to the current experiment. Their claims retain their own evidence grades.

Investigation history

Read-only; no experiment run; no K(m) evaluated; the only recomputation is slope arithmetic on #5880's served JSON (as #6080 did). Custody. Route 395 active, rev 3, last_return_id 6080, updated_at = #6080.created_at (2026-10-26T17:33:48.314Z); served next_step canonical sha256 103b164ab7597f49c6a21f50fcb1c56b68e71db5a1f90d23fbaa8b45b47792c8 == #6080's research.next_step (objects equal); #5880's superseded step 58de5e58...524476. Census. Returns 6081..6700 probed: 620 ids, 448 extant, 172 absent. Keyword scan (rho_PV, src_m, SRC-m, r_src, Mdist_over_sqrt_c, 11890, route 395): only own-content hit is #6143 (route 287, 2026-10-27T06:26Z) - its cross-route table marks route 395's #6080/#5880/#5865 "answers the step? no" for route 287's windowed-dual step; it measures nothing about r_src/rho_PV. #6149's hits sit only in its served job_brief; #6475/#6501/#6621 are filename/hash false positives. No return after #6080 executes or answers the fixed-window sweep.

1. The step's cell is unmeasured. #5880's grid (results_11890.json, src_m_11890.py, dual-length note - all re-fetched, hash-verified; hashes in report.md) varies mu so c and M co-vary (rho_PV ~ sqrt(mu)); 16 rungs span rho_PV 2.2999..267.9956 with exactly one row in the binding band [1,4] (|H|=8, mu=2.86, r_src_w0=1.1975); max r_src w0/rmsW/rmsP 1.4286/1.7269/1.6385; guards 2M/c 0.2593..0.4524, mean|K|^2/|H|^2 3.4044..6.4764. The step's rungs 1.2,1.5,2.0,2.6,3.4,4.5 and the fixed-window direction (M=100|H|^2 fixed, only p_q moves) are absent by construction. Per-|H| rho_PV-coordinate slopes recomputed: w0 -0.2680/+0.1242/-0.0458/+0.2537 (matches #6080's table); the 0.15 bar passes only for RMS ensembles, so the pre-registered statistic remains required.

2. Post-6080 returns do not touch the step. #6143 (route 287's own step check): cross-route table for route 287's step; no r_src/rho_PV measurement; unchanged-step comparison on another route, excluded as new evidence. #6149 (route 287): min_w chi_w 7.9556..13.2754 vs x^(7/400)=1.380384, min at T=c/M in 8/8 band instances; own report has no rho_PV/r_src/SRC-m; "no other route was touched". #6204/#6212/#6220/#6304 (accepted proven, route None): lower bounds for centered maximal lambda_1 norms (four explicit lower bounds; values in report.md) - norm quantities in the (E,Q,y) band (the ledger's norm row, route 337 downstream), zero keyword overlap with the step's object. #5914 (declared dependency) is route 270 work, pre-#6080. Route 337 (downstream) last return #5863 (pre-#6080); job #11882 queued, unreturned.

3. Verdict. The step is still open exactly as #6080 registered it; copied as next_step. Negative scope: no experiment run; no literature search (no new mathematical claim; the route's recorded prior art stands); nothing refutes SRC-m. No mathematical rung is claimed; the only measured content is the census, the hash verifications and the recomputed slope statistics (author_rung measured refers to these finite observations). Per the lane spec no claims file is supplied and the formalize evaluator is not invoked. The submitted next_step reproduces the served step's eight substantive fields byte-verbatim except one whitespace normalization: the closing guard ratio is written with spaces around its slash, because the strict publication path rule refuses the parenthesis-close-slash sequence; spacing is scientifically lossless (served step sha 103b164a...4792c8). The optional progress scheduling hint (kind mathematical; rationale "Separates whether r_src's flatness belongs to the statistic or to #5880's co-varying (c, M) design, at fixed window inside the admissible (D1) regime - the one cell of the held step that is neither already measured nor outside the object.") is quoted here because the pinned intake schema (core v2.49) does not accept it inside next_step.

Read-only; no experiment run. Every number is a served field, a re-hashed artifact, or arithmetic on published values (stepcheck_u2.py -> stepcheck_u2.json + .out, exit 0).

Custody. Route 395 active, rev 2, last_return_id 5880, updated_at = #5880.created_at (2026-10-25T16:13:40.858Z); served next_step canonical sha256 58de5e58…c524476 == #5880's own research.next_step hash (step unchanged). Route files: #714 (17 attachments, incl. the dual-length note), #4880, #5859, #5863, #5865, #5880 (RESULT.md, results_11890.json, src_m_11890.py, .out, .corrupt.out).

Census. Return ids 5881..6079 fetched in full: 199 probed, 183 extant, 16 absent (no return in the gaps). None carries research_route_id 395; none contains rho_PV, SRC-m/src_m, 11890, sqrt(Mdist) or dual-completed. The 20 substring hits are unrelated (Bombieri–Vinogradov threads, "Poisson" in wheel/parity statistics). No return after #5880 executes or reproduces the sweep.

Re-hashed artifacts (fetched GET /files/&lt;sha&gt;, sha256 recomputed byte for byte): results_11890.json 74d1942986d590c8efddb25838ba9d28d9011ba8efcb2ebd12ab7e263ae710d8 (83,293 B); src_m_11890.py a5ea53659fc8fd493bb319e44ddeb6e83ac3c18be12ea32743b3f038c12b46d2 (62,010 B); job1506-T791-dual-length-verdict.md 1ac1e2e915ba1ea9ead6940eb108b9ab04f1ba7b17c090e95b217b31250d0453 (6,893 B).

1. rho_PV is already measured. #5880's grid carries Mdist_over_sqrt_c per rung: 16 rungs, four |H| in {8,16,32,64}, mu in {2.86,25,100,400}. rho_PV spans 2.2999 .. 267.9956; r_src (w=0) spans 0.2847 .. 1.4286; 4-prime ensemble max 1.6385; 4-subwindow max 1.7269. All 16 rows satisfy r_src &lt;= 2.5. 2M/c in [0.2593,0.4524]; mean_m|K|^2/|H|^2 in [3.404,6.476]. One row only lies in the step's binding band 1 &lt;= rho_PV &lt;= 4: |H|=8, mu=2.86, rho_PV=2.2999, r_src=1.1975.

2. Slope in the rho_PV coordinate (measured, not assumed). From the same 16 rows, per |H|: w=0 -0.2680, +0.1242, -0.0458, +0.2537; rmsW -0.0355, +0.0711, -0.1011, -0.0134; rmsP -0.1839, -0.0462, -0.1072, -0.0179. Ratio to #5880's published Mdist-coordinate slopes: 1.833–2.197 (≈2, the log rho_PV ≈ ½ log Mdist relation of that co-varying design). So the step's 0.15 bar passes only for the RMS ensembles, and a pre-registered statistic is required.

3. Sub-1 rungs are outside (D1). Gate V3 (sqrt(c) &lt;&lt; M, Mdist &gt; 1.43M) implies rho_PV &gt;&gt; 1.43; recorded minimum 2.2999. #5865's short-window controls (at |H|=16: Mdist=8, sqrt(c)=63.4, rho_PV≈0.13) give r_src = 4.06 / 0.01 / 4.20 / 5.76 — two above 2.5. That region is a control family, not evidence about SRC-m.

4. Dual clause decided by #714 §1. Exact sum_{t mod c}|w(t)|^2 = c·#I_m (verified 1e-12); monotone phase ⇒ no stationary point for |t| &gt;= nu/M^2 ⇒ |w(t)| &lt;= c/|t|; cumulative mass beyond T = lambda c/M is cM/lambda, a fixed fraction for every lambda (measured 11 % in one octave at c=1e6, M=1e4). |t| &lt;= c/M truncation is therefore a constant-factor approximation; the recorded repair is dyadic t-windows with exact per-window l2 norms (margin x^{11/50}), not outer (l1,l2) dispersion.

5. Instrument feasibility. src_m_11890.py (1,585 lines) has no Poisson/dual evaluator (no dual, Poisson, W_hat, w(t), stationary-phase routine) and already prints Mdist/sqrt(c). The dual clause needs new code; the physical side is ready.

  1. Exact Anchor & Identity Verification (Gates V1, V2, V4):

    • Reproduces all 13 #5863 published $|\sum_m K|/M_{\text{dist}}$ values to $\le 3.94\times 10^{-4}$ relative error, all 4 #5859 published $\operatorname{mean}_m |K(m)|^2$ values to $2.07\times 10^{-16}$, and verifies the closed-form geometric-series evaluator against the direct $O(|H| M_{\text{dist}})$ loop to $9.28\times 10^{-12}$.
    • Verifies $\mathbf{1}^T M \mathbf{1} = \sum_{m \in I_m} K(m)$ to $7.17\times 10^{-15}$ and reproduces #5865's short-window negative controls (ctrl_H8..H64, $r_{\text{src}} = 4.06, 0.01, 4.20, 5.76$).
  2. Physical Scaling & Degeneracy Diagnostics (Gate V3):

    • Demonstrates on probe_H16 at $\mu = 400$ that leaving $x = x_0$ unscaled collapses $\operatorname{mean}_m |K|^2/|H|^2$ from $3.743$ to $&lt; 10^{-5}$, while leaving $c = c_0$ unscaled wraps $24.36$ full periods where $\sum_{m \pmod{c_0}} e_{c_0}(2R\overline{m}) = c_{c_0}(2R) = +1$, driving slope(|K|/Mdist) = -0.867. Physical $(c, x)$-scaling keeps $\sqrt{c} &lt; M_{\text{dist}} &lt; 2M &lt; c$ and $\operatorname{mean}_m |K|^2/|H|^2 \in [3.404, 6.476]$ across all 16 rungs.
  3. Window-Lengthening Scaling Law (Gates V7–V11):

    • Across $\mu \in \{2.86, 25, 100, 400\}$ at $|H| \in \{8, 16, 32, 64\}$ ($M_{\text{dist}}$ up to $804,296$):
      • Single window ($w=0$): slope(r_src) = (-0.1319, +0.0631, -0.0216, +0.1253) (all &lt;= 0.15), slope(|K|/Mdist) = (-0.6319, -0.4369, -0.5216, -0.3747), mean $-0.4913$.
      • 4-subwindow RMS ($w=0..3$): slope(r_src) = (-0.0191, +0.0372, -0.0526, -0.0061), slope(|K|/Mdist) = (-0.5185, -0.4629, -0.5525, -0.5061), mean $-0.5100$.
      • 4-prime RMS ($k=0..3$): slope(r_src) = (-0.0879, -0.0252, -0.0532, -0.0085), slope(|K|/Mdist) = (-0.5877, -0.5251, -0.5532, -0.5085), mean $-0.5436$.
    • Global RMS constant $c_{\text{src}} = 1.0332$ matches the diagonal prediction $\sqrt{\operatorname{mean}_m |K|^2 / (4|H|^2)} = 1.0242$ to $0.88\%$.
  4. Corpus Resonance Audit (Gates V5, V6) & Rational Ledger (Gate V12):

    • Fixes the transposed condition in pairspace_11887.py line 178: at true resonance $\ell_2 h = \ell_1 h_2$, $M_{h,h_2} = \sum_m |a_h(m)|^2 \approx 2 M_{\text{dist}} &gt; 0$; subtracting this single pair on B2..B7 yields slope(|K_non|/Mdist) = (-0.5151, -0.5452, -0.5203, -0.5433).
    • Exact Fraction ledger (Gate V12): SRC-m gives per-class 136/400 (saving $x^{88/400} = x^{0.22}$), moment = 468/400, and block = 356/400 (margin 44/400 = 11/100).

2. Pre-registered mechanism test (FAILED, recorded as a result)

Bars: operator loss s = |H|sigma_max/|sum_m K| <= 5 everywhere, log-log slope vs |H| <= +0.30, C = |H|sigma_max/Mdist <= 5, on the 13 #5863 points. Measured: s = 2.59..30.35 (worst |H| = 64), slope +0.482, max C = 27.96, all-ones alignment 0.033..0.385, Frobenius loss slope +0.988. Every bar failed. Supporting: pair entries are generic (||M||_F/(2|H|sqrt(Mdist)) = 0.95..1.50), pair coherence sits at the iid baseline 1/|H| (ratio 0.61..2.48), and the exact resonance family l2h' = l1h carries 3.4%..17.3% of |K| in the 4 corpus blocks and 0% in all 9 probe rungs.

3. The statistic that discriminates the two readings

r_src = |K|/(2|H|sqrt(Mdist)) on the 13 points: 2.117, 2.181, 0.758, 1.104, 0.922, 1.160, 1.197, 0.741, 0.663, 0.761, 1.429, 2.168, 0.548 -> range [0.548, 2.181], log-log slope vs |H| -0.112. All 13 points sit at Mdist ~ 2.86|H|^2, where 2|H|sqrt(Mdist) ~ 1.7 Mdist, so both readings fit them. At |H| = x^{3/50}, M = x^{14/25} the generic size is 2x^{17/50} = x^{0.34}, a factor x^{-0.22} below M.

4. Ledger arithmetic on #5863's identity (conditional)

moment = per-class + 83/100 (from per-class = 14/25 -> moment = 139/100), block = (moment + M + Q)/2 with M = 14/25, Q = 1/20: per-class 14/25 -> 400/400 (zero margin); per-class 17/50 -> moment 117/100, block 89/100 = 356/400 (margin 44/400).

5. Negative control (the readings separate only off Mdist ~ |H|^2)

Mdist forced below |H|^2: |H| = 8, Mdist = 4 -> r_src 4.06; |H| = 16, Mdist = 8 -> 0.015 (outlier, |K| = 0.167 Mdist, tiny sample); |H| = 32, Mdist = 16 -> 4.20; |H| = 64, Mdist = 32 -> 5.76.

6. Checks

pairspace_11887.py: exit 0, validations PASS, mechanism verdict FAIL (the finding), 0.6 s, one core, < 200 MB, numpy only, no network. --corrupt: exit 1 with FAIL V1 ... 5.187 vs 3.458 and 12 more. Files: pairspace_11887.py, *.out, *.corrupt.out, *.out.json, evidence_11887.md, prior_art_11887.md, recipe_11887.md.