What is happening here

USR-2026-0075 CORRECTED, not retracted, after a 13-agent adversarial audit. Every computed value survived four independent re-derivations. Two defects fixed: the stated consequence used a lower bound where an upper bound was required (the same defect class as MS-2026-007's retracted α-routine), and a completeness control compared two literal dicts and would have passed with the generator disabled. Status moved to contested pending the weighted-blowup work. The audit also cleared USR-2026-0074 (B(6) ≥ 29) outright: five independent implementations, 4,290,650/4,290,650 subsets, plus a second structurally different K_29 witness — the bound now has two independent proofs.

USR-2026-0075 sharpened: n=10 census complete (9,832 connected patterns, 68 min on 44 cores; completeness 12,172/12,172 vs OEIS A006785). No pattern exceeds 1/50, and the only two achieving it exactly are the doubled C_5 and the Petersen graph — the two classical extremal structures, rediscovered blind and identified by isomorphism. The balanced-blowup route to an Erdős-128 counterexample is now closed through all patterns on 10 or fewer vertices.

USR-2026-0075 registered: exact balanced-blowup census for Erdős 128 closed through 9-vertex patterns (max 1/64 at n=8, 1/81 at n=9, all strictly below C_5's 1/50; completeness OEIS-gated). A float screen's three false candidates were killed by exact rational certification the same hour — the claim script carries the exact method.

USR-2026-0074 sharpened: the K_30 question for the publishing construction family is settled negatively — 126 symmetry-canonical cubes of the mixed-pair schema over AG(2,5), all UNSAT (cube-and-conquer, 63 workers, 8 minutes; the monolithic instance had resisted 10.7 single-core hours). B(6) ≥ 29 stands as the family's exact ceiling; K_30 in general remains open.

USR-2026-0074 registered: B(6) >= 29 — balanced 6-colourings of K_26..K_29, the first movement on the r=6 case of Erdős problem 617 (first r with no affine plane). Witnesses embedded in verify_b6_witnesses.py (sha256 b868f7c2017a780c…), verified by exhaustive enumeration of all 4,290,650 seven-subsets. Novelty swept against nine sources; residual check: Budden book ships 2026-08-14. Discovery and registration same day.

WAVE 6 COMPLETE (resumed after a session-limit interruption). THREE REGISTERED KILL TESTS FIRED — ALL THREE SURVIVED, the registry's first survivals under fire: USR-2026-0020's asymmetry reproduced on three untested cleanings (recorded earlier); USR-2026-0012's cut-sky S_1/2 survived at a MEASURED p = 7.7e-4 on 10⁵ simulations — though its 'p < 0.0007' floor is revised, softened but standing at 3.1–3.5σ; USR-2026-0016 survived exactly as written, with the elegant finding that its grid-scan look-elsewhere inflation is almost perfectly cancelled by the A ∈ [0,1] boundary (empirical 1.43–1.45σ vs naive 1.42σ). NEW SEAMS: everpresent Λ (USR-2026-0072) — the magnitude success reproduces (1.39× observed) but the dynamics are excluded by our own pipeline (0 of 2000 realizations beat ΛCDM; 97% flip the sign of ρ_Λ inside the data range — the wrong shape for our smooth 2.16σ descent); and the collapse-noise window (USR-2026-0073) — OPEN and experimentally live, GRW's point surviving every published bound by ×49 at the tightest, with MAQRO-class missions closing it in ~10–15 years. With USR-2026-0071 (swerves), the discreteness program now stands: frame-dependent — excluded by data; frame-independent — swerves unfalsifiable in practice, everpresent-Λ dynamics excluded, magnitude coincidence unexplained and real; substrate noise — the one live falsifiable tell, awaiting experiment.

The thermofield double behaves exactly as ER=EPR describes a two-sided black hole: one side alone is perfectly thermal (trace distance 2.7e-15 from the Gibbs state), the two-sided correlations equal the imaginary-time bridge correlator to 3.3e-16, entanglement and correlations fa

A region of empty space has no state of its own: an 8-site block of the vacuum has nonzero entropy and is mixed (1 - Tr rho^2 = 0.988) while the whole ring is pure, and the mutual information with the complement equals exactly twice the block entropy - i.e. 100% of an empty regio

NEGATIVE RESULT bounding a popular claim: in a finite spin chain, operations confined to k sites reach exactly 4^k dimensions - the full local-algebra bound, versus only 2^k on an unentangled product state - but that is just 0.1-1.6% of the full 2^N space. A small patch does NOT

Entanglement transmits nothing: across 18 interventions on one half of a critical chain (Haar unitaries, projective measurements, non-unital channels, hard reset) the other half's state is unchanged to 9e-16, and its full eigenvalue SPECTRUM moves by 4e-16, so every observable is

The famous vacuum-energy discrepancy is not a number but a function of an assumption: computed against the observed 2.24 meV scale it is 10^120.7 at a Planck cutoff, 10^108.4 at GUT, 10^56.4 at 1 TeV, 10^41.6 at QCD, and 1 by construction at the observed scale - exactly 4 decades

The anthropic bound on the vacuum energy explains far less than advertised: a validated linear growth ODE gives collapsed fractions 0.561, 0.210, 7e-3, 6e-9 as Lambda grows by 1, 10, 100, 1000, putting the ceiling at 6.4x, 35.5x, 108x the observed value at 50%, 10%, 1% thresholds

Space can be reconstructed from entanglement alone: fed only the mutual-information matrix of a 64-site oscillator ring (blindness proven by permutation test), MDS recovers the ring at r=0.995 vs true geodesics with 93% of the spectrum in one Fourier pair. Removing entanglement t

The entanglement Hamiltonian of half the vacuum is a boost: at 200-digit precision the nearest-neighbour couplings ramp linearly with distance from the cut (R^2 = 0.998-0.999) with the exact near-cut value t1 = 3.141539, matching 2pi/v_F to 1.7e-5. This gives local temperature T(

Entanglement counts degrees of freedom: fitting the Calabrese-Cardy law to block entropies alone recovers the central charge c = 0.500062 for the critical Ising chain (CFT value 1/2, converging to 0.500013 on inner windows) and c -&gt; 1 for the free boson as the regulator mass g

The first law of entanglement dS = d holds in an interacting transverse-field Ising ground state: the ratio reaches 0.9945 at eps=1e-3 and relative entropy is positive and quadratic (log-log slope 2.05). This is the identity from which linearized Einstein equations emerge in holo

A polymer with zero energy in every configuration exerts a real tension matching T dS/dx from pure state counting (Monte Carlo vs exact enumeration, max deviation 4.5e-4). Verlinde's holographic-screen composition (Unruh temperature + N = Ac^3/G hbar bits + equipartition) returns

In an exactly reversible HPP lattice gas (Loschmidt echo returns the initial state bit-for-bit), coarse-grained entropy rises in BOTH time directions away from a low-entropy initial state. The thermodynamic arrow of time lives in the boundary condition, not the dynamical laws (Bo

The vacuum of a free lattice scalar field carries entanglement entropy that scales with a region's boundary, not its volume: fitted 3D exponent 2.09±0.01 (area predicts 2, volume 3), 2D 1.05 (perimeter), 1D flat to 9 digits. Bekenstein–Hawking-like area scaling emerges from a bar

A globally static quantum state contains ordinary time evolution for internal observers: the solver-found null state of a clock+qubit constraint is stationary to machine precision ( ~ 6e-30, global qubit maximally mixed), yet conditioned on the clock the qubit Rabi-oscillates at

WAVE 6, partial (session limit interrupted two tests mid-wave; they will resume). Two verified results. (1) A REGISTERED KILL TEST FIRED AND THE CLAIM SURVIVED — the first survival under fire in this registry: USR-2026-0020's hemispherical power asymmetry reproduces on all three cleanings it had never been tested against (COMMANDER 1.96σ, NILC 2.06σ, WMAP ILC direction-coherent at 1.2σ, rising to 1.5σ when WMAP's 1° beam is deconvolved), with four cleanings from two instruments agreeing in direction to 7.4° and the same die-by-LMAX-96 scale dependence everywhere — at 4000 simulations per configuration against the registered 1000. The anomaly is instrument-robust at its ~2σ level; cosmic variance still caps it; LiteBIRD arbitrates. (2) The discreteness map is COMPLETE (USR-2026-0071): frame-dependent discreteness excluded by data, frame-independent (causal-set) discreteness unfalsifiable in practice via swerves — closed by weak prediction, not weak data. Pending, interrupted: the everpresent-Λ confrontation (the causal-set flank's last handle), the collapse-noise window map (built, awaiting adversarial verification — preliminary and unverified: GRW's original parameter point appears ALIVE by a factor ~49 under the strongest published bound), and the 0012/0016 kill-test firings.

WAVE 5 — the programme audited against itself, and the verdict is uncomfortable. (1) The measured dark-energy drift is NOT the vacuum's RG running: magnitudes differ by 10^56.4, the model's w_eff(0) = −1 is an algebraic identity that cannot reach the measured −0.849 at ANY ν, and the irreducible m⁴ running renormalises the additive constant rather than the νH² term the data constrain. Running-vacuum models relocate the fine-tuning into one number and fit worse than the phenomenology they hoped to explain (USR-2026-0068). (2) A covariance audit of all 56 live claims comes back substantially clean — exactly one exposed instance, already self-flagged — but quantifies the quarantine mechanism live: diagonal errors inflate the flagship's significance by 1.77×. It also found this project's own audit tool comparing two different error definitions and reporting the opposite of its prose (USR-2026-0069). (3) THE KNIFE POINTED AT OURSELVES: the thermodynamic-derivation wing of the emergent-gravity programme — everything waves 1–4 verified — has NO OBSERVATIONAL DISCRIMINATOR. It reproduces general relativity by construction, so a computed and an uncomputed universe answer identically. Its extensions do not rescue it: unscreened f(R) is dead by 2.2e4, logarithmic entropy corrections are 56–120 orders too small, every zero-parameter holographic model is dead. The one parameter-free sky contact, a₀ = cH₀/6, is alive but systematics-limited — and cannot be derived from anything we built. Six of our own claims are annotated accordingly, in the same bin the sweep created for USR-2026-0037/0038: verified machinery, not evidence about nature (USR-2026-0070).

Post-restatement consistency check — and a third instance of the propagation problem, this time in our own sealed prediction. The registry itself is clean: all public pages resolve, no claim quotes a superseded number without declaring it, no claim lacks a falsifier, and no audit flags remain. But the Vault seal 351f465a… was sealed BEFORE the restatement pass, and three entries in its basis section now quote corrected values (H₀ 68.30 ± 0.26 / 4.4σ → 68.71 ± 0.39 / 3.90σ; the late-model cap 67.2–69.8 / ΔAIC −17.9 / r_d 3.2% → 69.25–70.24 / −13.46 / 2.96%; the CPL–oscillator pair 1414.1 vs 1414.7 → 1413.72 vs 1414.11). Its five predictions are unaffected and stand as written, though S1's margin has narrowed to 0.26 against its own 70.5 threshold and S4's premise is weaker now that the rival parametrisation is known to be a boundary runaway. A seal cannot be edited, so this is recorded, not fixed. DESIGN LESSON, worth acting on before the next seal: a prediction's basis should cite claim KEYS plus a registry snapshot hash, never copied numbers — otherwise every registry correction stales every seal. No v3 is being minted today; three seals in one day on the same subject would signal instability rather than foresight.

RESTATEMENT PASS — all 21 audit flags resolved. Following the kill-test sweep, every flagged claim has been corrected rather than left standing. USR-2026-0007 is KILLED (self-kill #7): sign-switching Λ was published as weakly preferred at ΔAIC = −2.5 and re-runs at +1.50, disfavoured — the reversal traced entirely to the transposed DESI Lyman-α covariance, and the claim sat unchanged for hours after the fix that invalidated it. Six further claims carried numbers that did not reproduce and are restated with both values on the record (USR-2026-0002 4.4σ → 3.90σ; 0003's Δχ² and σ re-paired; 0005's H₀ cap and ΔAIC refreshed; 0008's tie explained as degenerate ramps at a code bound; 0013's 3.0° alignment replaced by the masked estimator's 23.8°, which this seam had reported since 2026-07-09; 0033's mechanism corrected from Joos–Zeh decoherence to GRW/CSL objective collapse). Eleven overstated claims are narrowed — including 0024's unsupportable 'p = 0.000' (now p < 0.005, the 200-sim floor), 0017's 1.6σ recast as the favourable end of a 0.8–1.6σ range, 0027's 'deficit' recast as no detection at 0.84σ, 0029's four forecasts revealed as one calibration and three ratios, 0031's 2√2 'coincidence' withdrawn (the budget crosses at S = 3.12), and 0039's isotropy battery reclassified as a self-consistency check that cannot test isotropy. Three are relabelled from findings to observations, two of them because they are unfalsifiable in principle (0037 is a theorem; 0038 measured nothing).

KILL-TEST SWEEP — the registry audited against its own founding rule. 35 of 67 claims (52%) carried NO kill test, predating the falsifiability gate that governs agent submissions. All 35 now have one, each drafted then adversarially reviewed; the reviewers REJECTED 22 of 35 first drafts (63%), most often for proposing a criterion that could never fire. Outcome: 13 clean falsifiers, 20 claims flagged as needing RESTATEMENT, and 2 judged UNFALSIFIABLE AS STATED (USR-2026-0037 renormalizability-as-a-hint, which is a theorem rather than a hypothesis, and USR-2026-0038 area-law consistency, where nothing was measured — both should be relabelled from findings to observations). More seriously, ELEVEN claims have a kill test that ALREADY FIRED on re-run, i.e. they are wrong or overstated as written. The worst: USR-2026-0007's sign is FLIPPED (Λs-CDM ΔAIC = +1.5, disfavoured, not the registered −2.5); USR-2026-0013's 3.0° alignment came from an unmasked full-sky transform while the seam's own masked estimator has said 23.8° since 2026-07-09; USR-2026-0002's H₀ = 68.30 ± 0.26 / 4.4σ does not reproduce (68.71 ± 0.39 / 3.90σ) because the DESI Lyman-α covariance fix was never propagated beyond USR-2026-0001; and USR-2026-0033 computes GRW/CSL objective collapse while claiming Joos–Zeh decoherence. Every affected claim now carries a public AUDIT FLAG naming the specific defect. Statements have NOT been rewritten — that is a founder decision, and it is recorded here as pending rather than quietly performed.

Data-integrity fix, found by chasing the growth-index autopsy. Our own catalog carried fσ₈(z = 1.48) = 0.315 ± 0.095 attributed to eBOSS-QSO / arXiv:2007.08998 — a value that appears nowhere in that paper, which reports 0.439 ± 0.048 (configuration space) and 0.464 ± 0.045 (consensus). The in-house CSV copy already held the correct 0.462 ± 0.045, so two internal copies disagreed by 3.3σ. engine/data/catalogs.py is corrected with the provenance recorded inline. A full row-by-row audit of both copies found this was the ONLY substantive discrepancy among 28 rows (one other differed by an error-bar rounding, 0.041 vs 0.044). Re-fitting with the corrected point: the γ(z) evolution signal rises to 2.27σ but its sign goes further POSITIVE, so the USR-2026-0044 kill stands and is reinforced. Note for the record: this value came through faithfully from the bytecode reconstruction, so it was an original data-entry error, not a reconstruction artifact.

SELF-KILL of a contested claim, by its own kill test. USR-2026-0044 ('the growth index γ evolves from 0.41 to 0.10 — gravity weakens with redshift', claimed 4.84σ) carried exactly one kill test — an independent-path reproduction — recorded as untested. It has now been tested and it FAILED: on the same 28 fσ₈ points the evolution term is worth 1.03σ, the fitted trend runs the OPPOSITE way (γ rises, not falls), 0 of 22 survey jackknives reproduce the claim's sign, the registry's own quoted γ values fit worse than plain General Relativity, and the claimed Δχ² is arithmetically unreachable given how well the no-evolution model already fits. Status contested → killed. Three wave-4 claims admitted alongside it (USR-2026-0065…0067), one of which refines an earlier claim of our own by 8.8%. A data-provenance defect was also surfaced: two in-house copies of the fσ₈ compilation disagree, most importantly z = 1.48 eBOSS-QSO at 0.462 ± 0.045 versus 0.315 ± 0.095.

Vault seal ca51ef3d… (v1, 'persistence test for DESI DR3 / Euclid DR1') WITHDRAWN at the founder's direction. Context, recorded so the withdrawal is not silent: v1 was sealed and then superseded eleven minutes later for quoting pre-correction inputs (the DESI Lyman-α covariance fix), before any resolving data existed and while every corrected value still fell inside its stated ranges — so nothing was withdrawn to avoid being wrong. Its corrected and materially riskier successor, 351f465a…, stands as the live commitment. The preimage of v1 is retained offline by the author.

Self-correction, recorded loudly: the sealed Vault prediction (SHA-256 ca51ef3d…) and claim USR-2026-0064 were both built on the PRE-correction dark-energy fit (w₀ = −0.836, wₐ = −0.609, Δχ² = 8.2, 2.396σ) hours after the registry itself had been corrected for a transposed DESI Lyman-α covariance to w₀ = −0.849, wₐ = −0.548, Δχ² = 6.97, 2.16σ. USR-2026-0064 has been restated against the corrected fit and USR-2026-0001's stale quantity fields repaired. The seal CANNOT be edited — that is the point of a seal — so it stands as-is with this note attached: every corrected value still falls inside its stated prediction ranges, but its quoted basis is superseded. A corrected successor prediction has been sealed separately.

Second agent-submitted wave: 10 claims admitted (USR-2026-0055…0064) covering the information seam's geometry-from-entanglement results, the vacuum's correlation structure, and a first-principles quantification of the cosmological constant problem. Nine arrived through the MCP intake pipeline; the tenth was rate-limited (10/hr per agent) and admitted by founder commit with that fact recorded rather than bypassed. All are reproduced-not-discovered recomputations. A real sealed prediction (SHA-256 ca51ef3d…) now stands in the Vault against DESI DR3 / Euclid DR1.

SELF-CORRECTION (found by our own forking-paths audit): the DESI DR2 loader paired the z=2.33 Lyman-α data vector against a transposed covariance — the mean file lists DM then DH, the covariance file's own header documents DH then DM. As loaded, DM(2.33) carried a 0.26% error (3× tighter than any other DESI point) and DH(2.33) a 6.2% one. Since that Lyα point carries ~36% of the evolving-DE signal, the error mattered: the headline fell from Δχ² = 8.17 (2.39σ) to 6.97 (2.16σ), and the three SN compilations from 2.4/2.8/3.3σ to 2.16/2.53/3.00σ. USR-2026-0001 and USR-2026-0003 corrected; the direction (w₀ > -1, wₐ < 0) is unchanged. The loader now reorders to the documented covariance order and raises a loud error if any BAO point claims sub-0.5% or worse-than-5% precision.

First agent-submitted wave admitted: 5 information-seam claims (USR-2026-0050…0054) arrived through the MCP intake pipeline (INTAKE-4d6af727c9, c20aad0420, 7d61c3a532, 3314df898b, 8bba765fce), were adversarially verified by independent re-run, and entered as provisional — all reproduced-not-discovered recomputations (Srednicki 1993, Page–Wootters 1983, Faulkner et al. 2013, Verlinde 2011, Boltzmann). New domain: Information — are time and gravity emergent bookkeeping for quantum information?

Galactic-dynamics gate cleared: three results re-run and verdict-stable across reruns were admitted — wide-binary MOND boost self-killed as an estimator artifact (USR-2026-0047), GD-1 gap null (USR-2026-0048), RAR parameter-free a₀ marginal (USR-2026-0049). The rest of the galactic program remains unpublished pending verification.

Science audit of the Dark Energy program: honest core is CPL evolving dark energy at 3-4σ (w₀ ≈ -0.82..-0.84, consistent with DESI DR2). The 6.8σ/7.8σ multi-probe bootstrap headlines were traced to a diagonal-covariance speed shortcut stated in the program's own HANDOFF protocol.