# Report, job #67: paper "tailcount-transport" written from `paper/proposals/prop-tailcount-transport.md` Manuscript: `tailcount-transport.md` (sha256 in `paper.file`). Written 2026-09-11 by the Benjaminsen handle (claude-fable-5-1, effort max). Conflict of interest as before: the handle is the project author's. ## What the paper is A theorem note from the proposal queue. It states the Tail-Count Transport at the calibration the record gives it (proven, both forms), writes out the alternation-refined relaxation the proposal names as "argued in one sentence and never written out" (Lemma 2), and proves that the alternation-refined certificate equals $G_2$ of the folded tile at every fold (Theorem 2), which the record carried as measured at eight folds and open. It leads with the no-chain result, as the proposal asks, and reproduces the closure from the record. Prior art is the registry's. Nothing on TPC or the exponent. ## What changed against the proposal 1. **The proposal's third upgrade trigger fires.** "Upgrade if the alternation-refined certificate is proven equal to G₂ rather than measured equal at eight folds." Theorem 2 (§4): every window of the old word whose interior class word is a legal two-state walk sits inside a realised maximal dead run of the big tile at the copy whose alignment lies in $A_L(i)$, so some new gap is at least the window sum; with Corollary 1 this gives $G_2(T_q) = M_{\mathrm{alt}}$. The argument is the same containment by which $L$ is a word statistic in the exact-fold-L draft, applied to sums. Rung: proven. Mechanism-checked by `job67-alt-exact-check.js` at folds 7 to 23 (769,282 containments at fold 23, zero violations; window counts agree with the frontier producer's alternation-legal counts at the five shared folds). 2. **The second upgrade trigger fires.** "Upgrade to QUICK-DRAFT if the refined Q_L relaxation is written out as a proof." Lemma 2 (§3): non-emptiness of $A_L(i)$ is equivalent to legality of the interior class word, with the induction in both directions. Rung: proven. 3. **Fold 41, the proposal's first upgrade trigger, fires; its first downgrade trigger does not.** The record priced the ninth fold at 2.4 h + 5.4 h single-threaded and did not run it. A worker-thread port of the served fold-37 engine (`job67-transport-parallel.js`, validated against the served producer at folds 11 to 37, every printed figure agreeing) ran fold 41 in 3,607.8 s wall on 9 threads (28,346 s CPU, 7.9 CPU h of the 18 allowed): $G_2(T_{41}) = 546$ against the ladder's 546; $M_{\mathrm{loose}} = M_{\mathrm{alt}} = 546$; $D(T_{41}) = 8,499,244,879,125 = 39 \cdot D(T_{37})$ exact; $L(T_{37}, 41) = 3$; 0 violations at 91 thresholds, loose and refined, untruncated; max ratio 0.9551 at θ = 72 (a rise of 0.0074 over fold 37, below the 0.012 per step of the preceding five); $N_{\mathrm{new}}(546) = 4 =$ the ladder's `nmax`. Rung: verified. Manuscript §4 "Fold 41", §5. 4. **A small correction to the record's K-cap.** The origin and adversarial records state $K \le 1 + \theta/(3q)$; that is the even case of the run-cost floor. With `a3-05-bound-L.md` Corollary A1, $c_{\min}(m) = 3qm - (q + 2\eta)[m \text{ odd}]$, the cap for both parities is $K \le 1 + (\theta + q + 2)/(3q)$ (§6, Fact 2; §9 item 6). One third of a unit at $\theta = q^2$; nothing downstream moves. 5. **The proposal's HELD trigger does not fire.** The adversarial code is not in the repository, but the frontier producer (in the repository, output embedded) recomputes the adversarial certificate table cell for cell at folds 11 to 31, including the one cell where loose and refined differ; the parallel port written for this paper is a further implementation and agrees at folds 11 to 37 (§9 item 8). 6. **Framing.** With Theorem 2 the honest description of the instrument changes from "a bound that is exact at eight folds by relaxation" to "an identity". The paper says so (§4, "What the theorem says"; §9 item 2) and does not present the eight-fold agreement as evidence about tails. ## Calibration of every headline claim | claim | rung | where | |---|---|---| | Theorem 1, loose form: $N_{\mathrm{new}}(\theta) \le (q-2)N(\theta) + 2\sum_L Q_L^{\mathrm{loose}}(\theta)$ | proven | origin record §1, adversarial (a); §3 | | Theorem 1, refined form with $Q_L^{\mathrm{alt}}$ | proven (Lemma 2 written here) | §3 | | $\sum_L \nu_q(i,L) = q-2$, $D_{\mathrm{new}} = D(q-2)$ | proven | adversarial (a); §2 | | Corollary 1, $G_2(T_q) \le M_{\mathrm{alt}} \le M_{\mathrm{loose}}$ | proven | §3 | | Theorem 2, $G_2(T_q) = M_{\mathrm{alt}}$ | proven (here) | §4; mechanism check folds 7 to 23 | | Zero violations at every threshold, untruncated, folds 11 to 37 | verified | frontier producer OUTPUT; origin producer rerun folds 11 to 23; parallel port folds 11 to 37 | | $M_{\mathrm{loose}}$ strict at fold 29 only (270 vs 258) | verified | three implementations | | Ratio trend 0.8881 to 0.9477 | measured, not modelled | frontier37 §4, §9 | | Fixed-index chain certifies a constant (108/180/240/330) | verified | origin producer part 4B, rerun | | K-cap $K \le 1 + (\theta + q + 2)/(3q)$ | proven | a3-05 Cor A1; §6 | | Histogram not closed (0 of 50 shuffles) | measured, word-level caveat kept | adversarial (f) | | Window frame: $L=1$ term fails, ~39 nats over 233 folds | measured | adversarial (e) | | Chaining on the tile | closed | OUTCOMES row 2026-08-19 | | Two-semiring identification | inferred, unsearched | import-maxplus §1d; §7 | | Holt and Rudd own operator, closure theorem, $(q-2)$ multiplier | registry | PRIOR-ART, proposals-prior-art §2, lit-pdf-holt-rudd; §8 | | Author rung for the return: proven for Theorems 1 and 2; the numerical tables verified | | | ## What I verified and how - Re-ran `research/attack-foldL-03-transport.js 23` (14 s, all self-tests pass): certificates, truths, every-θ ratios at folds 11 to 23, chain table, index table read from the log. - Wrote and ran `job67-alt-exact-check.js` (2.5 s): the mechanism of Theorem 2 at folds 7 to 23, plus the $\nu$ identity at every slot. - Wrote (with a sub-agent) `job67-transport-parallel.js`, a worker-thread port of the served fold-37 engine with block pre- and post-rolls, and validated it against the served producer at folds 11 to 37: every figure agrees, including $D(T_{37})$, the run count 12,338,231,614, the window counts and the ratio 0.9477 at θ = 48. Then ran fold 41 on 9 threads (item 3 above); the log's own checks (truth, D_old, D_new, violations, unclosed runs) all pass. - Read every cited figure to its record; a second sub-agent fact-checked the manuscript's numbers and locators against the served records and refereed the proofs. It confirmed every table cell and every quoted phrase, found the proofs of Lemmas 1 and 2, Theorem 1, Corollary 1 and Theorem 2 sound (with two clauses to add: finiteness of the $L$-sum without assuming a non-qualifying gap, and that a legal window's interior lifts are distinct positions, both now in §3 and §4), and caught seven real errors in my draft, all fixed: the window-frame mechanism was mislocated (the qualifying supply dies at p = 701; the last gap reaching 2p − 2 is at p = 1021, the last kill run of length 2 at p = 421); §7 copied import-maxplus §1d's "closed on the histogram", which contradicts operator-and-pair-count.md and the shuffle test (now §9 item 11); "no upper bound on any maximum gap" over-claimed against 1402.1970 §4's $g(Q) \le g(\bar q\#)$; the Corollary 6.3 page-image read was attributed to the wrong record; "0.88 to 1.19 nats over the whole ladder" is the fourteen-term ladder only; the $\sum_L Q_L$ row mixed loose and refined counts at fold 29; and Theorem 2's relation to the copy theorem of U-FRAME §5a step 2 was unstated (now stated: Theorem 2 is the $m = 1$ case with straddling windows included, and leaves the record's straddling clause measured). - Style: zero em dashes; banned words grepped. ## What I could not verify - The fold-37 producer was not rerun (661.9 s recorded, output embedded and re-verified by `qc/embed.js` per frontier37 §8); its figures are read from the embedded block and reproduced independently by the parallel port. - Page locators into arXiv:1408.6002 are the record's (pp. 25–26 read as page images on 2026-08-19 per the record); not re-read here. - The semiring identification stays inferred; no search was made. ## Sources - Project documents at `/projects/twin-primes/docs/`, snapshot `main`, 2026-09-11: the parent records named in the manuscript header; `research/exact-g2-ladder.js` (ladder and `nmax`); `research/G2-STATE.md` (fold-41 pricing, G₂(41#) = 546). - F. B. Holt and H. Rudd, arXiv:1408.6002v1, via `research/history/staging/lit-pdf-holt-rudd.md` (verbatim quotations there); not re-read. - No local-only sources. ## Files - manuscript `tailcount-transport.md` - `job67-alt-exact-check.js`, `job67-alt-exact-check.log` - `job67-transport-parallel.js`, `job67-calib.log`, `job67-fold31.log`, `job67-fold37.log`, `job67-fold41.log` - `job67-foldL03-23.log` (rerun of the served origin producer) - `recipe.md` ## Transcript Attached, scrubbed as data (JSONL parsed, string values redacted, re-serialised): bearer token, session ids (platform and harness, by prefix), account and organisation identifiers, bridge and tool-result ids, home paths and the encoded working directory, e-mail; plus the two sub-agent transcripts (the parallel-port builder and the fact-checker) concatenated after the main file. No third-party source text was read this run, so nothing was omitted on that ground. Scrub script: `scrub.py` (local, not uploaded; it lists the patterns above).