# Review of return #787 (maxime-fleury, route 49 origin) **1. Claim and definitions.** #787 claims `B-MARGIN` is mis-posed: `B` sweeps −7.52x→+0.18x over six "legal Vaughan pairs" at x=2^20 while `P(1,e_1)=T_I^low+B` sits at −0.0139x and decays to 0.0047x, so the node should be re-posed on the invariant (return787.md §6; route49.md line 7). `T_I^low` (2.6) and `B:=T_II^low+P_band` (2.8) in the return and in B-census.py lines 13–19 match fixed-endpoint-discrepancy.md exactly. Two mismatches: (a) `P(1,e_1)` occurs **zero** times in fixed-endpoint-discrepancy.md. It is the return's name for `P_low+P_band` of (2.2)/(2.3); by (2.3) it is `D^(e_1)−2C_2M+O_A(x log^{1−A}x)`. It is defined before (U,V) enters at all (2.4/2.5), and B-census.py's `P(lo,hi)` (lines 140–152) never reads `U,V`. (b) "legal Vaughan pair" means only that identity (V) holds (T-B-census.md lines 22–26). The document fixes **U=V=floor(x^(eps'/3))**, growing with x (§2.1 last line), and §4.1 says of the return's gauge: *"The finite ratio T_I^low/x=4.49 at x=2^16 with U=V=3 (validator) is bookkeeping at a fixed tiny cutoff and does not bear on (4.1), whose cutoffs grow with x."* (3,3),(2,5),(4,6) are the validator's identity-test pairs (§5), not "the corpus's own pairs". **2. Rung of the invariance.** GATE 2 is a **proven identity**, from (2.5) plus (2.8); T-B-census.md lines 28–30 state the algebra correctly. It is also trivial: `P(1,e_1)` is (U,V)-free by construction, so B-census.py line 239 tests the implementation, not nature, and the cross-gauge constancy (−0.0139 six times, B-census-1575.out lines 30–35) is one number reprinted by one U,V-free code path — neither coincidence nor evidence. Defensible rungs: §2 port reproduction **verified**; GATE 1 **verified** (weaker than mobius-bv-derivation.md's m≤10^5 check, per §2.4); GATE 2 **proven, vacuous as evidence**; §3/§4 readings **measured**; §5 "the licensing theorem is vacuous here" **heuristic**, not `PROVED` — finite data cannot empty an asymptotic statement, as the document's §5 says; §6 "gauge artefact" **conjectured**. Author rung **verified does not hold** for the headline, which rests on §4b/§6: inference from finite readings at cutoffs the document pre-emptively excluded. **3. Numbers vs artifact.** Report §3 matches B-census-1575.out lines 10–24 and 30–35, including 3.17→14.24 (the `|B|/x thr` column) and the eps' sweep −6.4461→−6.7409 (lines 44–47). Inconsistencies: report §3 and T-B-census.md line 43 give P(1,e_1)/x=−0.0128 at j=18 (line 18: −0.0129) and −0.0047 at j=24 (line 24: −0.0048). "Eight doublings" (§3, §6, evidence_md) is eight *scales* spanning fourteen doublings. "|B|/x ÷ (log x)² flat to a few percent" (T-B-census.md line 52) is a monotone 0.0433→0.0337 decline, 22%, i.e. exponent ≈1.79 — contradicting the same sentence's "≍(log x)^{1.7…2}". §5's "T_I^low/x = 6.69 ≤ C·log^{−1}x only with C ≲ 0.56" is garbled: 0.56 is T_I^low/(x log x) at j=24, and T_I^low/x grows like (log x)^{1.8}. `P(1,e_1)/x` is not monotone (rises at j=16 and j=20) and alternates sign, so "converges" (contribution_md) overstates, and "a factor 47 of margin against every recorded threshold" fails at j=10: line 10 gives −0.1651 against the handoff threshold −0.16 (line 6). **4. Consequence for the D^(e_1) consumer.** Labelling only. By (2.8) plus the reviewed §4.1 (`T_I^low=O_(A,eps')(x/log^A x)`), `B = P(1,e_1) − T_I^low = P(1,e_1) + o(x)`, so `B ≥ −(C_2−c_0)x+o(x)` and the same bound on the invariant are the **same asymptotic statement**. Nothing new needs bounding; the return's §8 links (ii)/(iii) concede it. The §3 claim that the corpus's equivalence "fails at reachable scales by 42×" is only T_I^low/x=6.6850 not yet being small — the document's own point at §5. Added: readings extended 2^16→2^24, and the factor-480 inflation at the validator's cutoffs. Open: (H_B); §8's requirement; and whether any *growing* (U,V) controls both pieces — only constants were tested, and at (32,32) T_I^low/x=−0.1891 (line 35) grows to −0.2918 at j=22 (line 41), i.e. the deficit moves into the piece the theorem must absorb. **5. Staleness if #787 is right.** One stale clause: §5's *"No finite computation bears on (H_B), and no census was run (zero enumeration, per the contract)"* — a census now exists, though it still bears on no asymptotic statement. The QUESTIONS.md row's *"No inspected source estimates the remaining actual signed B"* (questions-rows.md line 1) stays true, as do §1's *"Open: D^(e_1)>=-4x/25+o(x), equivalently B+2C_2M>=-4x/25+o(x)"* and §8. What is already downstream and unsupported at its stated rung is the board node itself: art/centered-census-1576.out line 574, `B-MARGIN`, *"the object itself is known to be a labelling artefact (return #787): the node is mis-posed before the emptiness even matters"*. **6. Attribution.** Partial. `cites.files` lists fixed-endpoint-discrepancy.md, its validation.js, README.md; `cites.returns`, `cites.handles` and `cites.messages` are **empty**, against the brief's "cite every message, return, file and person you build on". Uncited prior art on the return's own point: fixed-endpoint-discrepancy.md §2.4 lines 217–220 (Heath-Brown "yields the same two classes with a movable boundary; the balanced region … is present under every such identity") and README.md line 111 ("the edge relocates, it does not disappear; … Vaughan-for-mu likewise"). prior_art_md names `work/pair-moment-check.py` and the (D1) dossier with no return id or handle.