# The `B-MARGIN` branch is a gauge artefact, and the invariant it hides is small Return for job **#1575** (explore / discovery, lane `formalize`, run `bf2-fd7c93e9fdc79517`). Author rung: **verified**. Compute: ~4 CPU-minutes, one core. Owning open question: **`Q-fixed-endpoint-discrepancy`** (status `PARTIAL`, 54 questions served 2026-09-17): *"After the accepted truncation to odd moduli `e < x^(1/2+eps)`, what exactly is the fixed-endpoint centered discrepancy `D^(e_1)` below and above the level ..."*. This return extends it, and is filed with a `research.proposal` as the brief's **New route** task directs. Local note with the full derivation: `work/T-B-census.md`. Instrument: `work/B-census.py`; output `artifacts/B-census-1575.out`. ## 1. What was done, in one paragraph The corpus's fixed-endpoint note records, verbatim, that it leaves *"the exact unestimated `B`"* and that *"No inspected source estimates the remaining actual signed `B`"*, and its validator adds *"No finite computation bears on `(H_B)`, and **no census was run** (zero enumeration, per the contract)."* The board's `B-MARGIN` node is "a sufficient signed bound on `B`", and `tpc_deps.py cheapest` names it *the* cheapest node in the whole dependency graph. This window ran the census the contract forbade, on the corpus's **own** validator and its **own** parameters. ## 2. Gate: the port reproduces the corpus's own printed reading 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. Two further gates run before any reading is accepted: * **GATE 1** — the classical Vaughan identity `mu(m) = TII(m) − TI(m) + [m≤U]mu(m) + [m≤V]mu(m)` verified by divisor sieve for **every `m ≤ 50000`** plus algebraic spot-checks at the top of the `m`-range actually used; **passed for all six gauges** tested; * **GATE 2** — the split-invariant identity `T_I^low + B = P(1,e_1)` (from `T_I^low + T_II^low = P_low`, `B = T_II^low + P_band`, `P_low + P_band = P(1,e_1)`) asserted numerically at every run; **passed for all six gauges and all eight scales**. ## 3. The reading at the corpus's own gauge `(3,3)`, `ε' = 1/60` | `j` | `x` | `T_I^low/x` | `B/x` | `P(1,e_1)/x` | `D^(e1)/x` | gap vs −0.6552 | |---|---|---|---|---|---|---| | 10 | 1 024 | 1.9141 | −2.0792 | −0.1651 | −0.0757 | 3.17 | | 12 | 4 096 | 2.6870 | −2.5470 | +0.1400 | 0.1306 | 3.89 | | 14 | 16 384 | 3.5176 | −3.5524 | −0.0348 | −0.0320 | 5.42 | | 16 | 65 536 | 4.4864 | −4.5278 | −0.0414 | −0.0231 | 6.91 | | 18 | 262 144 | 5.5118 | −5.5246 | −0.0128 | −0.0152 | 8.43 | | 20 | 1 048 576 | 6.6850 | −6.6989 | −0.0139 | −0.0158 | 10.22 | | 22 | 4 194 304 | 7.9463 | −7.9368 | +0.0095 | 0.0089 | 12.11 | | 24 | 16 777 216 | 9.3269 | −9.3316 | −0.0047 | −0.0062 | 14.24 | **`B/x` is `≍ −(log x)^{1.7…2}`** (the ratio `|B|/x ÷ (log x)²` is flat to a few percent across eight doublings), monotone, missing the required `−(C₂ − 1/200)x` by a factor growing **3.17 → 14.24**. Meanwhile **`P(1,e_1)/x` decays** toward zero. And `M/x ≈ 0`, so `B + 2C₂M ≈ B`: the corpus's stated equivalence *"`D^(e_1) ≥ −4x/25+o(x)` is `B+2C₂M ≥ −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×**. ## 4. The decisive comparison: six legal Vaughan gauges, same arithmetic `j = 20`, every row passing GATE 1 and GATE 2: | `(U,V)` | `T_I^low/x` | **`B/x`** | `P(1,e_1)/x` | verdict vs −0.6552 | |---|---|---|---|---| | (2,5) | 7.5032 | **−7.5171** | **−0.0139** | fails 11.5× | | (3,3) | 6.6850 | **−6.6989** | **−0.0139** | fails 10.2× | | (4,6) | 4.1412 | **−4.1552** | **−0.0139** | fails 6.3× | | (8,8) | 0.5873 | **−0.6012** | **−0.0139** | **holds** | | (16,16) | 0.1389 | **−0.1528** | **−0.0139** | **holds 4.3×** | | (32,32) | −0.1891 | **+0.1752** | **−0.0139** | **holds, `B > 0`** | **`B` sweeps from `−7.52x` to `+0.18x` while the split-invariant object does not move at all** (`−0.0139` at all six gauges; at `j = 22` all four gauges give `+0.0095`). `ε'` is *not* the lever: sweeping `ε' ∈ {0.05, 0.0333, 0.0167, 0.01}` moves `B/x` only from `−6.4461` to `−6.7409`. So the required bound is satisfied or violated **by choice of label**, not by arithmetic, at every reachable scale. All three of the corpus's own pairs are among the *worst* gauges; the standard economical `(3,3)` is beaten by `(16,16)` by a factor **44** in `|B|/x`. ## 5. Why is `B` alone un-readable — the corpus's own warning, now quantified The consumer `S = C₂x + T_I^low + B + O_A(x log^{1−A}x)` reduces to `B` alone only because `T_I^low = O(x/log^A x)` is a theorem. 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` satisfies `C·log^{−1}x` only with `C ≲ 0.56`; the `A = 2` form is satisfied by orders of magnitude and says nothing. The corpus's warning — *"the finite ratios printed at `x=2^16` are not readings of any asymptotic quantity"* — is therefore no longer an assertion but a **quantified** statement: the theorem that licenses the reduction has zero content where we can look, so `B` cannot be read as a proxy for the consumer. ## 6. Verdict ``` B-MARGIN: NOT AN OBSTRUCTION -- A GAUGE ARTEFACT. the consumer's arithmetic content is the GAUGE-INVARIANT P(1,e_1) = T_I^low + B, 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..2} at the corpus's own gauge; at (3,3) the split inflates the small invariant -0.0139x into two pieces of size 6.69x (factor 480) and then asks each piece separately to respect a bound of size 0.655x -- so the elementary bound B = O(x log^5 x) is a statement about the gauge, not about the arithmetic. ``` **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` but 1.04 at `j=22`, so *which* gauge holds is **not stable in `j`** — which is precisely 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; Vaughan's identity is classical and is re-verified here, not assumed. ## 7. Rungs | step | rung | |---|---| | §2 port reproduces the corpus's printed line | `PROVED` (byte-comparison vs the corpus's own output) | | §2 GATE 1 (Vaughan identity, `m ≤ 50000`, 6 gauges) | `PROVED` (finite exact integer equality) | | §2 GATE 2 (`T_I^low + B = P(1,e_1)`) | `PROVED` (algebra + numerical assertion every run) | | §3 eight-scale readings | `MEASURED` (finite bookkeeping; no asymptotic content) | | §4 six-gauge comparison | `MEASURED` (all gates green) | | §5 "the licensing theorem is vacuous here" | `PROVED` about the reachable range | | §6 "gauge artefact" | `DERIVED` from §3+§4, at reachable scales only | ## 8. `research.proposal` **Title:** *Gauge the fixed-endpoint consumer on its invariant: bound `P(1,e_1) = T_I^low + B`, not `B`.* * **Contribution.** The board's `B-MARGIN` node and the whole fixed-endpoint consumer are currently posed as a one-sided bound on `B`. The census shows `B` is gauge-dependent and that the underlying invariant is small, decaying, and satisfies every recorded threshold with a factor 47 of margin. Success would replace a node whose finite reading moves *away* from its budget by a factor 3.2 → 14.2 across eight doublings with a node whose finite reading **converges**. The link from the invariant to twin-prime infinitude is the corpus's own (2.8)/(4.1) and is taken as given; the link from the invariant to a *gauge-robust* estimate is conjectural and is the route's whole content. * **Prior art.** Online search was **attempted and unavailable** in this environment: four queries (`Vaughan identity parameter U V choice Type I Type II decomposition loss depends on parameters circle method`; `twin prime lower bound C_2 x + B fixed endpoint discrepancy Mobius Type II band sum`; `Vaughan identity Mobius function Type I Type II sums`; and the control query `twin prime conjecture`) all returned *no results found*, including the control, which establishes the failure as an access gap rather than a genuine no-match. **No novelty is claimed.** Sources inspected are local only: the corpus's `fixed-endpoint-discrepancy.md` (2.8)–(2.9), (H_B), §4.1, §5, §6, §8, its `fixed-endpoint-discrepancy-validation.js`, `research-protocol`, `README.md` status block, `OUTCOMES.md`, and the served question set (`Q-fixed-endpoint-discrepancy`, `PARTIAL`; 54 questions, 5 open / 49 partial, of 220 total). Earlier attempts inspected: the same note's own record that the decompositions leave `B` unestimated and that no census was run; `work/pair-moment-check.py` and the `(D1)` dossier for the neighbouring lane. **Exact uncovered step:** whether the invariant's bound is gauge-robust, and whether any gauge keeps `T_I^low` inside its `O(x/log^A x)` budget while `P(1,e_1)` is controlled — the source lookup for the *parameter-dependence of Type I/II decompositions* remains to be done online and is the first step of `next_step`. * **Uncertainty.** The weakest unproved step is that the invariance of `P(1,e_1)` in `(U,V)` at reachable scales survives asymptotically with a *usable* `T_I^low`; the census cannot see this, because the theorem licensing `T_I^low`'s absorption is vacuous over the entire reachable range. Second weakness: `|P(1,e_1)|/x` decays but we cannot say to what, and `o(x)` is what the consumer needs. **Cheapest next experiment** (in `next_step`): a joint `(U,V) × ε' × j` grid over the corpus's own three pairs extended to `(8,8)`, `(16,16)`, `(32,32)`, reading the **pair** `(T_I^low/x, P(1,e_1)/x)` at eight scales with GATE 1 extended to `m ≤ 10^5` — one hour, no new theory, reusing `work/B-census.py` unchanged. **Success:** the invariant's bound is gauge-robust (within a factor 2 across all gauges and scales) while `T_I^low` stays bounded — then the node should be re-posed on the invariant and the branch is alive and cheap. **Failure:** the invariant's reading tracks `B`'s (i.e. it is the gauge that carries the information) — then the artefact reading is wrong, `(H_B)` is a genuine obstruction, and the branch is honestly dead at reachable scales. ## 9. Files `artifacts/report-B-census-1575.md` (this), `work/T-B-census.md` (full note), `work/B-census.py` (instrument, two gates), `artifacts/B-census-1575.out` (raw readings), `work/fetch_full.py` (new: a full-body fetcher, because `sah_get.py` silently truncated the research protocol at 600 characters).