# Session summary, 2026-09-04: one wave toward the wall, four attacks and four red teams, the map moved and neither exponent did

<!-- ledger
id: Q-session-0904-summary
status: ANSWERED
todo: none
question: What did the 2026-09-04 wave (four Opus attacks on wall-facing questions, one orchestrator note, four Opus red teams, all verdicts applied) change, and what is still open?
verdict: Neither exponent moved; item 0's growth half survived a second adversarial pass so the route is a truth gap at rung derived-and-red-teamed-twice; Face 4's "no lower bound on the kappa = 2 sifting limit is known" was a convention failure and published floors at or below 2 are now cited, while the recon's headline that the cap is a method artefact was refuted by its red team; killer 2's coordinate is measured for the first time (argmax mirror-invariant, zero congruence pairs from x = 23, forced) after the multiplicity half turned out to be on the ladder already; no piece of R0 is both legal and parity-exempt with content; L7 is not a transfer; about thirty live-layer sentences corrected across seven files.
-->

*(Orchestrator's summary. Records: nine staging notes dated 0904, all HELD,
every producer bit-honest under `embed.js --check`; `research/history/CHANGELOG.md`
2026-09-04 entry carries every applied edit with its file. Gate at close: §4.)*

## 1. What is still open, first

- The exponent band (2, 4.26645] on G₂ is unchanged. Z₂ has no band below the
  zone width that is not TPC. Nothing today opened a route.
- Item 0 is closed at rung derived-and-red-teamed-twice. Two failed breaks on
  a derivation are two failed breaks, not a proof; only a non-two-class sieve
  input reopens it, and none is known.
- The published floors on β(2) sit at or below 2 (Brady 2017: 1.8196; β(2) ≥ 2
  from the one-sided dimension axiom). Inside the band there is still no
  barrier theorem and no exhibited axiom-only argument. The corpus's LP floor
  3.3152 is a finite-level measurement for one profile and must not be quoted
  as a floor on the class.
- Named follow-ups, none run: the tension between the zone anchor's z-score
  (−0.49 to +2.13 over 43 levels) and `origin-excess.md`'s 21 per cent deficit
  at u = 2; the x = 37 level of the argmax red team (not re-run); the
  translate lemma's consequence for the union bound at the extreme, stated as
  understanding only.
- Owed chores: the one-line X2 fix in `varE-theta2-step.js` and its re-embed
  (item 9); `a3-01-misalignment-ledger.js`:371-373's comment; the
  `qc.js --index` wrapper throws after its checks (the generator run directly
  works).

## 2. What moved, by thread (each line calibrated at its record)

- **Item 0** (`redteam-0904-floor-growth-2.md`, graded by
  `redteam-0904-item0.md`): SURVIVES, same rung. Every number reproduced on a
  third code base. Proof-gap reading REFUTED. Scope widened to every admissible
  level split; crossing (s, u₀)-dependent; onset 1.875e6 at the record's s.
  `REFUTED.md`, `TODO.md`, `README.md`, `G2-STATE.md` corrected; four riders.
- **Face 4** (`recon-0904-sifting-limit-floor.md`, graded by
  `redteam-0904-sifting-limit.md`): literature half CONFIRMED and strengthened
  (Selberg's reciprocal convention with a second source; Brady's class is the
  one Face 4 needs); headline REFUTED on four grounds. Five live files
  corrected, one `SEARCH-CONVENTIONS.md` §1 row added with its gate effect
  measured (nothing existing newly clears).
- **Killer 2's coordinate** (`measure-0904-argmax.md`, graded by
  `redteam-0904-argmax.md`): multiplicity was already on the ladder (brief
  premise stale); positions MEASURED exhaustively at x = 11..37 and reproduced
  at 11..31 by a fourth engine; the null ratio corrected downward by up to 29×
  and then ruled a category error (the statistic is deterministic); four of
  eight sealed forecasts failed. `G2-STATE.md` §3c gains the certificate
  sentence.
- **R0 and parity** (`derive-0904-r0-extension.md`, graded by
  `redteam-0904-r0-extension.md`): the negative written once, conditional on
  Claim 1; the one-coordinate scope slip in three files corrected (it was Tao's
  Example 2); the variance artifact the note said was missing exists; zone
  anchor z-score measured; equivariance transfers by restriction.
- **L7** (`derive-0904-L7-transfer.md`, graded in `redteam-0904-item0.md`):
  not a transfer; both mechanisms priced or dead; OPEN as a statement.
- **Gate**: six aged absence verifications re-dated after re-verification.

## 3. Method notes

- A brief premise can be stale even when it comes from a red-teamed object
  read: "multiplicity unknown" was on the ladder in three producers. Grep the
  `.js` producers, not only `QUESTIONS.md`, before calling anything unasked.
- A recon's headline needs its red team before any live sentence: the day's
  largest claim ("method artefact at rung MEASURED") died on four grounds
  within three hours of being written.
- Deleting one dangling annotation in a producer costs three re-binds and a
  custody rider. Put annotations on the store's line or not at all.

## 4. Gate and custody

`node research/qc.js` at zero findings after each integration step;
`research/QUESTIONS.md` regenerated (472 questions, 499 notes, 0 unindexed);
every new id on its item's `Ledger:` line. Full gate run at close (see the
commit). No agent ran a git command beyond read-only status; no agent edited an
existing file.