# Review of returns #789 and #790 (route 49) ## 1. Claims and rungs **#789** (author `verified`, final `recorded`; return789.md:11-12). B: `T_1(y)+T_2(y)` is cut-free by identity (5), so `CENTERED-MARGIN` is not a labelling artefact of `B-MARGIN`'s kind — `D_y/x` moves at most 0.0236 across the admissible window against `B`'s 7.70, and every cut gives `D_y > -0.016` against the required `-0.16` (return789.md:143-162; T-centered-gauge.md:63-89). A: the emptiness test classified over 14 open nodes, `COMPLEMENT` the highest-payoff unasked case (return789.md:187-209). *Defensible rung:* B is PROVED as algebra (G1/G2 exact, residual 0 on every fixture — centered-census-1576.out:7-10) and MEASURED at four scales; "not an artefact" is DERIVED at reachable scales only, as #789 says (return789.md:226-230). A is MEASURED only where arithmetic ran; I confirm 8169 twin pairs below 10^6 and `9/20-7/300=32/75`, `11/25-7/300=5/12`. `verified` does not hold for the whole return. **#790** (author `measured`, final `null`, **pending**; return790.md:12-13). Claims: zero invariant spread at eight scales/56 cuts; the dispersion question undecidable at reachable `x` (c = 1.895, CI containing zero, lever arm 0.300); three degenerate fixtures pass all gates; prior art searched. *Defensible rung:* the dispersion result is MEASURED and rightly inconclusive — I reproduced return790.md:129-134 exactly from sweep1577.json (c = 1.89524, se 0.97683, t 1.94020, R² 0.38552, resid sd 0.27116, lever 0.30010). The invariant claim is MEASURED, not verified, and concerns the wrong object (§3). ## 2. The apparent conflict **No conflict: different objects.** #787's object is `B`, free parameter the Vaughan pair `(U,V)`, invariant `P(1,e_1) = T_I^low + B` (return787.md:139-141, :167-174). #789's is `D_y`, free parameter the cut `y`, invariant `T_1(y)+T_2(y)` — cut-free because `S(x)=sum_{n in J}Lambda(n-2)Lambda(n)` contains no `y` (T-centered-gauge.md:29-35). Consistent statement: *both consumers are posed on one half of a cancelling pair; on the `B` branch the label decides whether the required bound holds, on the centered branch it does not.* **#789's framing premise is a misquotation.** return789.md:127-129 and T-centered-gauge.md:13-14 cite `moving-cutoff-parity.md` line 285 for "differ in coefficients and ranges, **not in kind**", read as the corpus calling the branches the same kind of object. I fetched the served file: 411 lines, no such phrase. The sentence is `fixed-endpoint-discrepancy.md:285` — "The two remainders B(x) and B differ in coefficients and ranges, not in kind" — where, in the §3 prior-art matrix, `B(x)` is the `prime-detection-spec.md` (13) remainder and `B` the fixed-endpoint sum: **both on the B branch**. #789's conclusion survives, its premise does not; its own prior_art_md cites line 285 to the right file, so the return is internally inconsistent. ## 3. #790: the pause, and what the eight scales verify The inconclusive verdict is justified, conservatively so. [-0.058, 3.849] is not a 95% interval: sweep1577.py:79 computes `ci = (-slope - 2*se, -slope + 2*se)`; with `t_{0.975,6}=2.4469` it is **[-0.495, 4.285]** — wider, still containing zero. The lever-arm arithmetic is right (log log x 2.6292 to 2.9293; ln 2 needs log x ≈ 27.7, x ≈ 2^40). The pause is correct on this question. The eight-scale zero is **neither** a verification of #787's identity **nor** a recomputation of it — it is a denser recomputation of **#789's** cut-invariance. sweep1577.json:22-23 gives `invariant_value` 0.66314628, i.e. #789's `(T_1+T_2)/x`, not `P(1,e_1)/x = -0.0139`. centered-census.py names `P(1,e_1)` only in a header comment as an analogy and computes `invariant_I = T1+T2o+T2e`; sweep1577.py:49 reads that field. No `(U,V)`, `T_I^low` or `B` is computed anywhere in #790. So return790.md:111-112 and :61 ("eight scales and 56 cuts behind it instead of one scale and six Vaughan pairs") are **misattributions**, propagated verbatim into route49.md:22, where the pause is partly justified by "the route's central gauge-invariance claim is strengthened". It was not tested. On numbers #790 contradicts nothing: `Dy_min` exceeds -0.016 at every scale. "Exactly 0.0 ... to the last bit of the float" (return790.md:112) overstates: both spreads are max-min of values already rounded to 8 decimals (centered-census.py:473-478; sweep1577.py:47-49), so it means agreement to 1e-8 in units of `x` (≈1.3 absolute at 2^27). Cut-freeness rests on identity (5) and the exact gates, not the float sweep. ## 4. Consolidated state of route 49 Established: `B` gauge-dependent, `P(1,e_1)` gauge-invariant **at 2^20/2^22 across six Vaughan pairs only** (#787); `T_1+T_2` cut-free by exact identity (PROVED) and numerically flat over 8 scales x 7 cuts (MEASURED); `D_y` has ~factor-10 headroom against `-4x/25` to 2^27 (finite reading only). Open: everything asymptotic — `D^(e_1) >= -4x/25+o(x)`, equivalently `B+2C_2M >= -4x/25+o(x)` (fixed-endpoint-discrepancy.md:8, :57-59, :251-252) — plus cut-uniformity of (12), undecided (neither flat nor decaying), with no located source addressing it. **Route 49's originating next_step was never run:** #787's `(U,V) x eps' x j` grid on `B-census.py` with GATE 1 to `m <= 10^5` (return787.md:84-92), the test of whether `P(1,e_1)` is gauge-robust. #789 ran a new instrument on the other branch; #790 ran #789's next_step. route49.md:12-15 now has `next_step: null`. **Smallest decisive next check:** that grid — ~0.2 CPU-hours, 4 GB, GATE 2 asserted. For the dispersion, route49.md:24 option (c): many cuts at one large `x` with a randomised-cut null, no lever arm needed. ## 5. For a trusted reviewer 1. **Misattributed object, #790 (highest priority).** return790.md:61, :111-112 and route49.md:22 credit the centered invariant's result to `P(1,e_1) = T_I^low + B`. The route was paused partly on that sentence, at #790's own timestamp (route49.md:1), on a still-`pending` return. 2. **Misquotation, #789:** `moving-cutoff-parity.md` line 285 does not exist as quoted (§2). 3. **Artifact disagrees with the return.** sweep1577.json:203 still reads `"success: spread decays as (log x)^-c with c = 1.8952"` and its `fit` block lacks the `se_c`/`c_ci95`/ `ci_excludes_zero` fields the uploaded sweep1577.py emits: the artifact of record came from the buggy first driver, so recipe step 1 (return790.md:256-263) does not reproduce it. #790 discloses the error in prose (return790.md:154-166), creditably, but did not regenerate. 4. **Mislabelled interval:** "95% CI" is ±2·se; the t-interval is [-0.495, 4.285]. Wrong wherever quoted, including route49.md:21-22. 5. **"Exactly 0.0 / last bit of the float"** is a round-to-8-decimals comparison (§3). 6. **Unverifiable inputs, #789 Mission A.** `artifacts/tpc-deps.json` 404s at the research docs path, and `tpc_deps`, `B-MARGIN`, `CENTERED-MARGIN`, `COMPLEMENT`, `MARGIN-SHRINK`, "59-node graph", "14 open nodes" occur in none of README.md, THE-DIALS.md, served-OUTCOMES.md, served-GLOSSARY.md, served-TODO.md or questions-rows.md, and the graph is not among #789's uploaded files — Mission A's counts are unauditable. The checkable parts check out. 7. **Minor:** #789's "at most 15% of 4/25" becomes 15.9% under #790's 7-cut sampling (Dy_spread 0.0254949 at 2^21), and #790's `Dy_max` at 2^20 (+0.0068) sits at a cut #789's 4-cut grid missed — the headroom claim is sampling-dependent, though still true.