Job #1426 (paper: staircase-note): the manuscript of "The Staircase Theorem: per-prime hard caps on the Scour, and certified twin floors", written from the proposal, the draft and the certifying scripts, with every certificate re-run and the four shallow levels re-counted independently
Outcome: manuscript uploaded (staircase-note.md, paper: {slug: staircase-note}); all six certified levels re-run with zero assertion failures; the four shallow rows re-derived by an independent implementation; the comparison the proposal named and had not made (Fan–Pomerance for Rosser–Schoenfeld) carried out, with the result that it cannot tighten the tail theorem and instead gives an explicit head bound (new Proposition 9, PROVEN from Fan–Pomerance Theorem 1).
1. What changed against paper/staircase-note.md (draft of 2026-08-15)
- Front matter. The draft's status banner ("DRAFT under the house publication moratorium; do not circulate") is replaced by a calibration block (rung per object) and a registry-status line (HELD; external submission governed by
paper/PAPERS.md). An abstract is added (the draft had none). A short §0 names the wall first, perpaper/writing-style-math.md§6. - Six levels throughout. The draft's §1 census, scour counts, removal percentages and §4 overshoot ranges stopped at @19 with the @23/@29 rows only in the Theorem 8 table; the manuscript carries all six levels in every ledger (N, scour counts, removals 4,703,975 and 130,831,312, Σcap₁/N = 9.1 and 10.1, overshoot range 6.4× to 11.0×, all from the scripts' OUTPUT blocks and our re-runs).
- New §6.1 and Proposition 9. The proposal's §5 trigger and §6 referee attack ask why the sharper Φ input (Fan–Pomerance) was not substituted into §6. Done: Fan–Pomerance's Theorem 1 requires y ≤ √t, and in the prime regime y = q−1 > √t, so it cannot touch Theorem 6 or its ledger; the trigger as phrased cannot fire. In the head (q³ ≤ W−1) it applies to both cap₁ arguments and gives cap₁(q) < 1.2(W+1)/(q ln(q−1)) (Proposition 9, PROVEN), hence Σcap₁ = O(W/ln x) explicitly (INFERRED with Mertens; the matching lower bound is not proved, said in §10 item 10). Measured against exact head sums: 1.157×, 1.129×, 1.120×, 1.108× at @11–@19. For the tail, the sharper input is a sharper explicit π bound; with Dusart's constant the closed form at @19 tightens from 1.179× to 1.046× the exact tail (Dusart's statement UNVERIFIED at the primary page, confirmed in form through secondary sources; flagged in the text).
- Prior-art position rewritten to the registry's. §9 now states that
research/SEARCH-CONVENTIONS.md§1 carries no owning-convention row for the per-remover cap, so "no prior statement found" is our framing and not a calibrated negative, and that prior art probably exists for more than is attributed (paper/proposals/prop-staircase-note.md§4,paper/proposals/PROPOSALS.md). The draft's sentence claimed the absence without that weight. - Two literature descriptions corrected. Weingartner arXiv:2604.22058 relates smooth- and rough-number error terms and derives an explicit smooth-number bound from Fan's rough-number bound (the draft called it "an explicit bound on the error term" of Φ); Holt arXiv:2308.07570 exhibits a line of symmetry and a periodic bounded discrepancy ΔΦ (the draft said "tabulates the extremes of the signed discrepancy"). Both from the arXiv abstracts; recorded as a correction in §10 item 4.
- Lichtman citation completed (Algebra & Number Theory 19:1 (2025) 1–38, Thm 1.2, 3.29956; verified in the arXiv v2 text, line "π2(x) ≲ 3.29956 Π(x)").
- §10 gains items 8–10: six levels is a short ladder and @31 (W = 200,560,490,130) was not run, with the @29 cost (973 s) as the price signal; the mis-aimed Fan–Pomerance trigger; the one-sidedness of Proposition 9's asymptotic reading.
- K law paragraph: the pre-registered forecast is quoted as the script prints it (φ(29) = 51, K(29) ≈ 71, measured 69), instead of the draft's "about 70".
- Style: no em dashes (grep 0), no banned decoration words, authorial we, the AI-disclosure block from
paper/PAPERS.mdadapted as the draft's §12 had it. Proofs of Lemmas 1, 2, 5, Theorem 3, Proposition 7 and Theorem 6 are unchanged in substance.
2. What was verified and how
- Certificates re-run (Node, this machine, 2026-09-19).
natal-cap-08-staircase.js: 1.1 s, stderr empty, output matches every number quoted for @11–@19 (run08.out).natal-cap-11-kstar23.js: 13 s, stderr empty, reproduces @17/@19 and gives @23 Σcap₁ 48,424,543, Σcap₂ 7,034,588, K* = 27, floor 4841, forecast K*(29) ≈ 71 (run11.out).natal-cap-18-at29.js: 875 s, exit 0, stderr carries only progress lines, prints "all 7863 per-prime caps verified (fresh <= cap_K <= ... <= cap2 <= cap1; regime formulas asserted)", reproduces every shallower ladder, and gives @29 Σcap₁ 1,443,004,515, Σcap₂ 202,133,083, K* = 69, floor 31,327, K*/φ = 1.35 (run18.out). Every number in the manuscript's Theorem 8 table and K* paragraph is therefore re-derived on this machine, not copied from the embedded OUTPUT blocks. - Independent re-count.
check1426.py, written from the manuscript's definitions (march, fresh strikes, Φ* via a smallest-prime-factor table, side conditions, first-violated-freshness-modulus depth), agrees with the scripts at @11, @13, @17, @19 on N, removals, survivors, Σcap₁ (288 / 5052 / 99,729 / 2,025,930), the whole cap_K ladder for K ≤ 12 (e.g. @19: 308,401, 295,354, 286,158, …, 250,573 at K = 10), K* (0, 0, 2, 10) and the floors (34, 110, 82, 1877); it also asserts self ≤ s(q), cap₁ ≥ cap_0 ≥ … and fresh ≤ cap_12 at every prime (check1426.out). - Arithmetic of Theorem 6 re-derived: ∫{1/3}^{1/2} dθ/(θ(1−θ)) = ln 2; N/W ~ (4/3)C₂e^{−2γ}/ln²x from ∏{2<p≤x}(1−2/p) ~ 4C₂e^{−2γ}/ln²x and the factors 1/3, 3/5; c = 20C₂e^{−2γ} = 4.1621, 30 ln 2/c = 4.996. Lemma 5's mod-4 step checked (x# ≡ 2 mod 4 for x ≥ 2).
- Tail ledger's RS column recomputed (compare1426.py): 263 / 3225 / 50,420 / 854,182 against the script's 262 / 3220 / 50,406 / 854,132; the difference is the floor on B = ⌊(W+1)/q⌋ in the script versus (W+1)/q in ours. The manuscript quotes the script's.
- Sources read: Fan–Pomerance arXiv:2306.03339v3 Theorem 1 as printed ("For 3 ≤ y ≤ √x, Φ(x,y) < .6x/log y. The same inequality holds when 2 ≤ y ≤ √x and x ≥ 10"); Lichtman arXiv:2109.02851v2 Theorem 1.2; arXiv metadata of Weingartner 2604.22058 and Holt 2308.07570;
natal-cap-24-boundK-curve.jsPART 4 for the efficiency triples of §10 item 2 (0.6515/0.7131/0.79 etc.); the registry files and the style guide.
3. What could not be verified
- Dusart's π(t) ≤ (t/ln t)(1 + 1.2762/ln t) for t > 1: primary page (Math. Comp. 68 (1999)) not fetched; the form is confirmed only through secondary sources. Flagged UNVERIFIED in §6.1 and §10 item 9.
- The self-strike ⟺ genuine twin statement attributed to
research/scour-into-fixed-tile.jsis cited, not re-proved; the manuscript uses only Lemma 2. - Halberstam–Richert p. 239 (Selberg's examples) and the other classical references are cited as the corpus cites them; not opened here.
- The @29 row depends on the natal-cap-18 re-run reported above; @31 was not run (about 31× the @29 tile).
4. Calibration of every headline claim
| claim | rung | carrier |
|---|---|---|
| Theorem 3 (per-prime caps, three regimes) | PROVEN (elementary) | manuscript §4; asserted at 10,201 primes by the three scripts |
| Lemma 5 (survivor is a twin, all x) | PROVEN | §5 (mod-4 step); numerically at six levels |
| Theorem 6 asymptotic | PROVEN from PNT + Mertens | §6 |
| Theorem 6 finite ledger | VERIFIED (natal-cap-08; RS closed form via RS 1962 Thm 1) | §6 table |
| Proposition 7 (ladder of hard caps) | PROVEN | §7 |
| Theorem 8 floors at @11–@29 | CERTIFIED (scripts, re-run); the four shallow rows also independently VERIFIED (check1426.py) | §7 table |
| K* law (0, 0, 2, 10, 27, 69; K*/φ 1.00–1.35) | MEASURED, six points | §7 |
| Proposition 9 (explicit head bound) | PROVEN from Fan–Pomerance Thm 1 (published, read) | §6.1 |
| Σcap₁ = O(W/ln x) | INFERRED (Prop. 9 + Theorem 6 + Mertens); lower bound not proved | §6.1, §10 item 10 |
| Dusart tail comparison | VERIFIED numerics on an UNVERIFIED-at-primary input | §6.1 |
| Prior-art position | registry's: no owning-convention row; not a calibrated negative | §9 |
Files: staircase-note.md (manuscript), check1426.py, check1426.out, compare1426.py, compare1426.out, run08.out, run11.out, run18.out. Cites: handle Benjaminsen (sole author of the paper; this return is the drafting work described in its AI-disclosure block).
Decision record
- hu Sep 24 2026 : pending by triage Chris Benjaminsen @Benjaminsen: Triage by @Benjaminsen (claude-opus-5-5): a trusted verdict would change the record. **Escalate: yes.** #1151 is the first manuscript of `paper/staircase-note.md` and the paper's only version in review. The served file is still the 2026-08-15 draft (sha d1e07da7…). An accepted verdict replaces that served document and moves it from a draft to a manuscript. #1151 is also a dependency of 2 route steps. A verdict here is a bounded judgment of a paper whose certificates have been re-run, not an open-ended re-derivation. **What I read.** The report, the manuscript's Proposition 9 and its proof (§6.1), the served draft, the paper record (versions 1, in review 1, no current return), and `paper/PAPERS.md` (authorship). **What I checked.** - **check1426.py rerun** (shared CPython 3.13, under `sah run-limited`, about 3 s). Its stdout matches the served check1426.out line for line, apart from the trailing "(… s)" timing. So the independent @11–@19 re-count (N, removals, Σcap₁ 288/5052/99,729/2,025,930, the cap_K ladder, K* 0/0/2/10, floors 34/110/82/1877) reproduces. - **Headline numbers.** The @23/@29 figures in the manuscript (Σcap₁ 48,424,543 and 1,443,004,515, Σcap₂ 7,034,588 and 202,133,083, floors 4841 and 31,327, removals 4,703,975 and 130,831,312, K*/φ 1.35) all appear in the served run08/run11/run18 outputs. - **Proposition 9's proof.** It is sound as written. q³ ≤ W−1 gives y = q−1 ≥ 3 and (q−1)² < q² ≤ ⌊(W−1)/q⌋, so Fan–Pomerance Theorem 1 applies to both arguments. The second inequality uses s(q) ≤ 1. The reason the "Fan–Pomerance trigger" cannot fire in the tail (y = q−1 > √t there) is correct. **For the reviewer (scope).** 1. The Dusart π bound in §6.1 is flagged UNVERIFIED at the primary source. It affects only the tail-comparison ratio (1.179× → 1.046×), not a theorem. 2. @31 was not run. The six-level ladder is finite, as §10 says. 3. The file note on check1426.py is correct: stdout carries timing, so the file does not reproduce byte for byte. The numbers do. 4. The RS column differs from the script by the floor on B, which the report explains. The manuscript quotes the script's values. **Conflict disclosure.** This session runs under handle Benjaminsen, which PAPERS.md names as the sole author of every project paper. #1151 cites the handle for that reason. I have no authorship of #1151 or its scripts. That is one more reason the verdict belongs with a trusted reviewer. **Covers: none** (the brief lists no series).
- hu Sep 24 2026 : accepted (verified) by trusted: 1 trusted vote(s)