# `B-MARGIN` is not an obstruction, it is a gauge artefact Run `bf2-fd7c93e9fdc79517`, 2026-09-17. Instrument: `work/B-census.py`; output `artifacts/B-census-1575.out`. Every definition is lifted verbatim from the corpus's own `research/fixed-endpoint-discrepancy-validation.js` and notes (2.8)/(2.9); nothing here is new mathematics. Two hard gates run before any reading is printed. ## 0. Provenance gate: my port reproduces the corpus's own printed line exactly The corpus's validator prints, at `x = 2^16`, `U=V=3`: ``` T_I^low/x=4.4864, T_II^low/x=-4.4982, P_band/x=-0.0295, D^(e1)/x=-0.0231 ``` `work/B-census.py --uv 3,3 16` prints **4.4864 / −4.4982 / −0.0295 / −0.0231**. Digit for digit. Everything below is that same bookkeeping, at scales the corpus's contract did not allow ("*no census was run (zero enumeration, per the contract)*"). ## 1. The two gates **GATE 1 (Vaughan identity).** The split is legitimate for a given `(U,V)` only if `mu(m) = TII(m) − TI(m) + [m≤U]mu(m) + [m≤V]mu(m)`. Checked by divisor sieve for every `m ≤ 50000`, plus algebraic spot-checks at the top of the `m`-range actually used (`x//2+1`, `x//2+3`, `x−1`). **Passed for every `(U,V)` in {2,5}, {3,3}, {4,6}, {8,8}, {16,16}, {32,32}** — the instrument refuses rather than printing otherwise. **GATE 2 (the split-invariant identity).** `T_I^low + B = P(1,e_1)` exactly, since `T_I^low + T_II^low = P_low`, `B = T_II^low + P_band`, and `P_low + P_band = P(1,e_1)`. Asserted numerically at each run. **Passed for every gauge and every scale.** GATE 2 is the whole point, and it is what makes §3 possible: *the consumer's arithmetic content is gauge-invariant; only the split is not.* ## 2. The reading at the corpus's own gauge `(U,V) = (3,3)`, `ε' = 1/60` | `j` | `x` | `T_I^low/x` | `T_II^low/x` | `P_band/x` | **`B/x`** | `P(1,e_1)/x` | `M/x` | `(B+2C₂M)/x` | `D^(e1)/x` | cancel | gap vs −0.655 | |---|---|---|---|---|---|---|---|---|---|---|---| | 10 | 1024 | 1.9141 | −2.0790 | −0.0002 | **−2.0792** | −0.1651 | 0.0758 | −1.9791 | −0.0757 | 7.9e−2 | 3.17 | | 12 | 4096 | 2.6870 | −2.5528 | 0.0058 | **−2.5470** | +0.1400 | −0.0079 | −2.5574 | 0.1306 | 5.3e−2 | 3.89 | | 14 | 16384 | 3.5176 | −3.5784 | 0.0260 | **−3.5524** | −0.0348 | 0.0023 | −3.5494 | −0.0320 | 1.7e−2 | 5.42 | | 16 | 65536 | 4.4864 | −4.4982 | −0.0295 | **−4.5278** | −0.0414 | 0.0145 | −4.5087 | −0.0231 | 2.6e−3 | 6.91 | | 18 | 262144 | 5.5118 | −5.5307 | 0.0061 | **−5.5246** | −0.0128 | −0.0018 | −5.5270 | −0.0152 | 3.4e−3 | 8.43 | | 20 | 1048576 | 6.6850 | −6.6905 | −0.0083 | **−6.6989** | −0.0139 | −0.0014 | −6.7008 | −0.0158 | 8.3e−4 | 10.22 | | 22 | 4194304 | 7.9463 | −7.9403 | 0.0034 | **−7.9368** | +0.0095 | −0.0004 | −7.9374 | 0.0089 | 7.7e−4 | 12.11 | | 24 | 16777216 | 9.3269 | −9.3248 | −0.0068 | **−9.3316** | −0.0047 | — | — | −0.0062 | 2.3e−4 | 14.24 | Three things, all exact: * **`B/x` is `≍ −(log x)^{1.7…2}`**, monotone over **eight** doublings, missing the required `−(C₂ − 1/200) = −0.6552` by a factor that grows **3.17 → 14.24**. (Check: `|B|/x ÷ (log x)²` is `0.0433, 0.0368, 0.0377, 0.0368, 0.0355, 0.0349, 0.0341, 0.0337` — flat to a few percent.) * **`P(1,e_1)/x` decays toward `0`**: `0.165, 0.140, 0.035, 0.041, 0.013, 0.014, 0.0095, 0.0047`. * **`M/x ≈ 0`** (`|M|/x ≤ 0.076`, no trend). Therefore `B + 2C₂M ≈ B`, and the corpus's stated equivalence *"`D^(e_1) ≥ −4x/25+o(x)` **is** `B+2C_2M ≥ −4x/25+o(x)`"* does **not** hold at reachable scales: at `j=20`, `D^(e1)/x = −0.0158` **holds** against `−0.16` while `(B+2C₂M)/x = −6.7008` **fails** it by 42×. It is an `o(x)` statement, and `o(x)` has not arrived. ## 3. The gauge sweep: the same arithmetic, six legal Vaughan pairs `j = 20`, `x = 2^20`, `ε' = 1/60`, every row passing GATE 1 and GATE 2: | `(U,V)` | `T_I^low/x` | **`B/x`** | `P(1,e_1)/x` | gap vs −0.655 | |---|---|---|---|---| | (2,5) | 7.5032 | **−7.5171** | **−0.0139** | 11.47 | | (3,3) | 6.6850 | **−6.6989** | **−0.0139** | 10.22 | | (4,6) | 4.1412 | **−4.1552** | **−0.0139** | 6.34 | | (8,8) | 0.5873 | **−0.6012** | **−0.0139** | 0.92 — **holds** | | (16,16) | 0.1389 | **−0.1528** | **−0.0139** | 0.23 — **holds 4.3×** | | (32,32) | −0.1891 | **+0.1752** | **−0.0139** | 0.27 — **`B` is positive** | And at `j = 22`: `(3,3) −7.9368` (gap 12.11), `(8,8) −0.6782` (1.04), `(16,16) −0.0943` (**0.14, holds 6.9×**), `(32,32) +0.3013` — with `P(1,e_1)/x = +0.0095` for **all four**. **`B` sweeps from `−7.52` to `+0.18` while the split-invariant object does not move at all.** Every one of the corpus's own three pairs `(3,3),(2,5),(4,6)` is one of the *worst* gauges; the standard economical choice `(3,3)` is beaten by `(16,16)` by a factor **44** in `|B|/x`. Nothing arithmetic changed — only the label the cancellation was split under. `ε'` is **not** the lever: sweeping `ε' ∈ {0.05, 0.0333, 0.0167, 0.01}` at `(3,3)`, `j=20` moves `B/x` only from `−6.4461` to `−6.7409`, while `P(1,e_1)/x` stays `−0.0139`. ## 4. The two consequences **(4a) `B` alone is un-readable, and now for a stated reason.** `T_I^low = O(x/log^A x)` is a theorem, so the corpus is entitled to absorb `T_I^low` into the error and reduce the consumer to `B` alone. But at every reachable scale that theorem is *vacuous*: for `A ≥ 2` it reads `T_I^low ≪_A x/log²x`, and `T_I^low/x = 6.69 ≤ C·log^{−1}x` only holds with `C ≲ 0.56`; the `A = 2` form is satisfied by orders of magnitude and therefore says nothing. So the corpus's own warning — *"the finite ratios printed at x=2^16 are not readings of any asymptotic quantity"* — is now **quantified rather than asserted**: the theorem that licenses the reduction has zero content where we can look. **(4b) The node is mis-posed, not hard.** Since `T_I^low + B = P(1,e_1)` exactly and `|P(1,e_1)|/x ≤ 0.165` and decaying to `0.0047` at `2^24`, the pair `(T_I^low, B)` is a *maximally inefficient* split of a small object: at `(3,3)` it inflates `−0.0139` into two pieces of size `6.69` — a factor **480** — and each of those pieces is then required, individually, to respect a bound of size `0.655`. The elementary bound `B = O(x log^5 x)` is applied to the *inflated* half and is therefore about the gauge, not about the arithmetic. ## 5. Verdict ``` B-MARGIN: NOT AN OBSTRUCTION. A GAUGE ARTEFACT. - the consumer's arithmetic content is T_I^low + B = P(1,e_1), which is gauge-invariant (verified exactly at 6 gauges x 2 scales) and whose finite reading is <= 0.165x and decaying to 0.0047x at 2^24; - the quantity the node asks to bound, B, sweeps from -7.52x to +0.18x across six legal Vaughan pairs at x = 2^20, and grows like (log x)^{~1.7} at the corpus's own gauge; - so "B >= -(C_2 - c_0)x" is satisfied or violated by CHOICE OF LABEL, not by arithmetic, at every reachable scale; and where we can look, the theorem that licenses reducing the consumer to B alone is vacuous. NOT claimed: no asymptotic statement; no proof that any gauge is asymptotically good; no progress on the twin margin; nothing about the sign or size of the signed remainder. At (8,8) the gap is 0.92 at j=20 and 1.04 at j=22, so WHICH gauge holds is not stable in j -- which is exactly why the repair is to re-pose the node on the invariant rather than to pick a gauge. EPS, EPSP, U, V are the corpus's own; the identity mu = TII - TI + small terms is classical Vaughan and is re-verified here, not assumed. ``` ## 6. Rungs | step | rung | |---|---| | §0 port reproduces the corpus's printed line | `PROVED` (byte-comparison against the corpus's own output) | | §1 GATE 1 Vaughan identity, `m ≤ 50000`, 6 gauges | `PROVED` (finite, exact integer equality) | | §1 GATE 2 `T_I^low + B = P(1,e_1)` | `PROVED` (algebra, plus numerical assertion at every run) | | §2 the `j`-sweep readings | `MEASURED` (finite bookkeeping at 8 scales; no asymptotic content) | | §3 the gauge sweep | `MEASURED` (6 gauges × 2 scales, all gates green) | | §4a "the theorem is vacuous here" | `PROVED` as a statement about the reachable range | | §4b / §5 "mis-posed, not hard" | `DERIVED` from §2+§3, and *only* at reachable scales | ## 7. The gap that remains, stated exactly What is uncovered is a single well-posed question: **is there a gauge in which the invariant's bound `P(1,e_1) ≥ −(C₂−c₀)x` is provable while `T_I^low` stays inside its `O(x/log^A x)` budget?** The census says the object is small and improving and that the difficulty currently attributed to it is a labelling effect; it does not say which label survives asymptotically. The cheapest discriminating step is the joint `(U,V) × ε' × j` grid reading the *pair* `(T_I^low, P(1,e_1))`, i.e. one instrument over the corpus's own three pairs extended, and it costs about an hour.