The coupled-scalar test — open record

Where the test stands

1. Scope and model agreed — done
2. Solver built — done
3. First validation suite passed — done
4. Pre-pilot checks — in progress
5. Sampling pilot — next
6. Production chains — next
7. Results published here — next

Pending before the pilot

1. Revised precision design — The screened 7300 design failed its matched validation against the 8000 reference at the two steep corners (λ = 2): Plik moved −0.213 and −0.189, against the 0.05 limit, while points 1–5 passed every criterion. 7300 is not eligible. A revised design was declared, reviewed and approved on October 5, 2026: the ladder 7300 → 8000 → 8700 → 9400 → 10000 at both corners, the smallest qualifying candidate selected, then validated fresh against 10000 at all seven points under the unchanged criteria. It declares the build carrying the repair that passed its test on October 5. Its campaign stopped at a preflight memory check before any evaluation, so nothing about precision was learned. No saved memory reading passes the conservative launch worksheet, so the largest level is not shown to fit the machine. A read-only feasibility assessment found no observed idle state that passes the combined worksheet for that level, with two fixed forecast limits also binding; the drafted memory measurement could not settle this. A lower-memory build that computes the lensing step’s tables in blocks has a source-only declaration under review; nothing is implemented, and the route is Colson’s decision. A new campaign needs its own review, a new identity and Colson’s go.
2. Inference-precision certification — Tighter numerical precision must leave predictions and likelihoods unchanged at every declared point. Not yet certified: the first ladder did not converge, the staged check stopped at its first point, the bounded follow-up stopped at its first updated pair on Plik (+0.0660 against 0.05), the screened 7300 design failed its matched validation at the steep corners, and the revised design’s campaign stopped at its preflight memory check before any evaluation. Any revised candidate needs its own declaration and must pass the unchanged suite.

Then a short sampling pilot — proposals, likelihood, solver failures and speed, plus the declared Halofit sensitivity measured by reweighting pilot samples (plan drafted; approved with the pilot) — then production chains. The pilot’s speed sets the chains’ schedule.

What the test asks

A bounded test that lets the chains decide whether a scalar field coupled to all matter is allowed by cosmological data and local gravity tests. It is the corpus refined back to its sharpest testable core, run in the cosmology project where the hi_class solver lives.

A declared candidate, not claimed as uniquely derived by the corpus. The chain measures β and λ only: it cannot separately identify k_E and M_C, so it tests β ≠ 0 and the ratio, never k_E alone.

The declared model

Field, coupling and potential

A(φ) = exp(βφ/M_Pl)
V(φ) = V₀·exp(−λφ/M_Pl)
K = 1

Reduced Planck mass. Universal coupling to baryons, dark matter and massive neutrinos. The potential is the dark energy — no separate Λ. Radiation and massive neutrinos are explicit.

Matter-frame form (u = φ/M_Pl, X = −½(∂φ)²)

G₄ = (M_Pl²/2)·e^(−2βu)
G₂ = (1 − 6β²)·e^(−2βu)·X − V₀·e^(−(λ+4β)u)
G₃ = G₅ = 0

Derived separately by both agents. Observables are computed in the matter (Jordan) frame.

Property functions

α_M = −2β·(dφ/d ln a)/M_Pl
α_B = −α_M
α_T = 0

α_M(z) is derived from the field evolution, not sampled. α_T = 0: gravitational waves travel at the speed of light.

Local strength

fifth force = 2β² × gravity
γ − 1 = −4β²/(1 + 2β²)

The convention check both agents used.

Symmetry

(φ, β, λ) → (−φ, −β, −λ)

Leaves the model unchanged, so the prior can take λ ≥ 0 with β of either sign. Whether β and λ share a sign decides if a chameleon resting point exists (see the principle below).

Colson’s principle: a dynamic scalar field

“it is a dynamic scalar field”

Like a chameleon: present everywhere, but camouflaged by its environment until conditions bring it into function.

V_eff(u) = V₀·e^(−λu) + ρ·e^(βu)
βλ > 0: a density-dependent resting point, m² = β(λ + β)ρ/M_Pl²
βλ < 0: no resting point, so no screening
not yet settled: m² ≈ V″ + β²ρ/M_Pl²

Einstein frame, conserved ρ.

Whether the field settles is the thin-shell question, so its local range is checked environment by environment, never inferred from how light it is on cosmological scales. The camouflaged regime — a chameleon, or a symmetron-type threshold switch — is the screened branch: deferred, not excluded. If the data show tension, the decision rule points there.

Scope

• Unscreened branch only, over an explicit parameter range.
• Screened branch deferred, not excluded: it needs a self-consistent environmental profile plus laboratory-G calibration.
• Dark-matter-only coupling and a field-dependent kinetic term K(C) are separate future models.
• Why the unscreened branch is self-consistent (order-of-magnitude estimates): Cassini gives |β| ≲ 2.5×10⁻³, so N = β(λ + β) ≲ 10⁻² for λ ≲ 4 — about 10⁸ below the estimated thin-shell onsets (≈ 4×10⁶ for the Sun, ~10¹⁰ for the Earth). With N < 0 there is no minimum at all.
• In range, G_lab = G_*·A²(1 + 2β²); the correction is ≲ 1.3×10⁻⁵.
• Universal coupling is a metric theory: no separate first-order clock or spectral-line signal (δ ln A = 2β²Φ_N/c² sits inside G_lab).

Data and local tests

• Planck 2018 — Cosmic microwave background: full high-ℓ Plik TT/TE/EE with its nuisance parameters, low-ℓ TT, low-ℓ EE and lensing
• Pantheon+ — Type Ia supernovae, without the SH0ES calibration
• DESI 2024 — Baryon acoustic oscillations (first data release)

Frozen September 30, 2026 with Colson’s approval in the cosmology project’s run specification (SHA-256 as reported by its agent: FROZEN_SPEC.md 368af53429ee…, frozen_inference.json 8404307c8d59…). DESI’s second data release was declared before any result as a separate follow-on run. Every run uses the linear matter treatment, a declared approximation; a separate Halofit check measures the sensitivity to nonlinear structure. The first run uses the ΛCDM comparison’s neutrino prescription (one massive at 0.06 eV, N_ur = 2.0328) in a matched pair: the same likelihoods and numerics for the coupled model and ΛCDM.

Priors

• Local gravity enters as likelihood terms, not prior cuts.
• β: uniform on [−0.01, 0.01], either sign (frozen) — four times the Cassini edge, so the data, not the prior, set the edge.
• λ: uniform on [0, 2] (frozen); λ ≥ 0 by the symmetry. Any pile-up at λ = 2 is reported as a prior-edge effect, not a measured cutoff, and widening the range needs a new declaration. The edge is below 4, so the N ≲ 10⁻² scope line stands as written.

Decision rule and expectations

Stated before any chain.

1. Two runs: cosmology only, and cosmology plus the local-gravity terms, each matched against a ΛCDM run with the same likelihoods and numerics.
2. If the two runs agree, the result is an allowed region of β and λ.
3. If they are in tension, the result shows where the unscreened model fails to reconcile the observations — the data’s signal to open the screened branch or declare a model change.
4. Expected split: cosmology mainly constrains λ; the local terms mainly constrain β (cosmological coupling effects are small: fifth force ~2β² ≲ 10⁻⁵, matter-mass drift ≲ ~10⁻³). This split is expected, not a failure.
5. No signal receives a σ label unless its statistic is calibrated by null simulations or by a conservative method shown valid for it. Near β = λ = 0 observables depend on the couplings only quadratically, so χ² tables and Fisher forecasts do not apply there.
6. Either outcome is useful and is published here; no chain is dropped for disagreeing.

Expectations SHA-256: 0a8f3801badd39342b6c383f6dbf672ae6dbb52480fa5dd1ba2e8b228d3bcd68

SHA-256 of the expectations text: the statements above joined by line breaks, UTF-8. Fingerprinted September 30, 2026, before any chain. A fingerprint identifies exact text; it is not an external preregistration. If the expectations ever change, the old fingerprint stays in the status log.

The historical ratio test

Form agreed by both agents. The thresholds below are frozen in a separate declaration that Colson approved on October 1, 2026 (gate snapshot SHA-256 ba1ad4c1653f…). The approval authorizes no mocks, ratio fits or chains.

Historical prediction: β/λ = k_E·x_V ≈ 0.0499 — a direction θ₀ = arctan(0.0499) ≈ 2.9° in the (β, λ) plane, with θ defined mod 180°.

• D0 — pre-run gate — Mocks through the real pipeline, β and λ varied jointly. Needs a discriminating outcome in ≥ 80% of θ₀-line mocks and ≤ 5% of θ = 0 mocks. If unmet, it is declared in advance that the run tests the general model, not the ratio.
• D1 — signal gate — Profile-likelihood ratio against β = λ = 0, calibrated by null simulations (≈ 10⁴ fits for a 3σ-level tail) or a conservative method shown valid for this statistic. Otherwise no σ label, and the gate is not passed.
• D2 — direction — 95% profile set of θ. Incompatible if θ₀ lies outside it. “Compatible with, and precise enough to discriminate, the proposed ratio” only if the set is one connected interval containing θ₀ and narrower than θ₀; otherwise uninformative.

Already known from local gravity: on the θ₀ line, Cassini excludes the whole tracking regime (which starts at λ ≈ 1.69, β ≈ 0.084). The weak segment, λ ≲ 0.05, survives as a valid drifting solution. Inside the frozen priors the θ₀ line leaves the β range at λ ≈ 0.20 — a prior edge, not a data limit.

Outcomes concern the ratio k_E·x_V only, never the full framework; an incompatible result cannot single out k_E. The ratio is never silently imposed or dropped. Under the frozen priors alone the 95th percentile of |β/λ| is 0.050, so a prior-dominated bound can look like 0.0499; such a bound is never quoted as agreement — the gates use profile likelihoods.

Source identified October 1, 2026: k_E·x_V ≈ 0.0499 appears in 32 of the 207 corpus papers (corpus as of August 3, 2026), mainly as a pulsar-timing slope (Papers LXXIII and CLXXVII) and as the Schumann term χ_S (Paper CL). No corpus paper writes it as β/λ: that identification was made in this test’s refinement on September 30 and is labeled a hypothesis. Colson approved testing it as a stated hypothesis on October 1, 2026. No D0 mock campaign is scheduled, so unless a bounded D0 check is separately approved and passes first, the run tests the general model, not the ratio. x_V (the historical x_m) is unsettled, and this page states no value for it.

Validation results so far

Colson, two agents, one record

• Colson — Brings the questions, the direction and turning-point ideas, relays each step between the agents, and decides.
• This project’s agent — Worked the mathematics one step at a time in chat: the physical-margin argument, the lunar-ranging term and the with/without comparison; later reviewed each precision report and draft declaration.
• The cosmology agent — Works in Colson’s separate cosmology project with hi_class: reviewed each step and supplied the engineering, validation and preservation, then built and validated the solver, froze the run specification and runs the precision checks.

1. A step of mathematics is worked in chat.
2. Colson relays it to the cosmology agent.
3. The review comes back: accepted, strong warning, or still open.
4. Valid gaps are accepted and each withdrawal is named.
5. Only what survives accumulates into the next step.

“recursive accumulation ♾️ in action between 2 agents. LITERALLY.”

“stick to what you know… we don't need to solve the whole universe in one go”

“stop trying to prove every screened branch fails before building a working test”

The last two arrived together. Unresolved calculations became open limitations instead of blockers.

The two agents independently reached the same minimum — the same frozen model, the unscreened scope, Cassini labeled unscreened — then supplied complementary halves.

Withdrawals

Withdrawals are part of the product. Each item below was dropped or narrowed in review, and each drop made the surviving test more trustworthy. The test does not depend on any of them.

• The Cassini window — γ ≈ −4β_Sβ_P is not a Cassini measurement model.
• “Atoms are never screened” — Too absolute.
• The 50% redshift figure — Dropped in review.
• The 10⁻⁸ bound — Dropped in review.
• The 0.0475 slope — Corrected: the slope in measured density is s̃ = β/(λ + 4β) ≈ 0.0416 (ρ̂ = A³ρ̃).
• GPS Δln A ≈ 1.5 — Assumed an unsolved environmental profile; 1.5 is not perturbative.
• The 9.6% solar-line figure — Same unsolved-profile assumption; line shifts also need the matter-frame redshift.
• “Outside classical treatment” above N ~ 10⁴⁴ — A sub-spacing Compton length calls for a microscopic treatment; M_C is a normalization scale, not the scalar’s mass.
• An unconditional order-β² mass correction — Narrowed to a field that has not settled (m² ≈ V″ + β²ρ/M_Pl²).
• “β = 1.92 has no matter era” — Wrong: the steep potential makes it track.
• “Plain ρ − 3p cancellation fits the stall’s step pattern” — Too loose: it predicts only a sixth to a third of the observed step. The located line cancels terms about four times radiation’s size, and that matches to 8 digits.
• A few-times-10⁻⁸ spot-check band for the repair — Assumed the saved states meet the Friedmann constraint to about one rounding unit; they carry a residual near −3×10⁻¹⁵, which the old line amplifies. Replaced by two checks with limits fixed before any data.
• A two-flag compiler guard (-march=, -mfma) — Incomplete: other flags also switch on fused multiply–add (-mavx512*, -mavx10*) or change rounding (-mfpmath=387), and Nix can pass flags through variables the guard does not read. None is in this build, and any effect would fail a check rather than pass one. A future guard should allow only the recipe’s own flags.
• “A quiet project leaves enough memory margin” — Stated by this project’s agent when it accepted the campaign’s launch memory check, without working through how the platform counts file cache. The campaign’s own input copying could use up the margin, and it stopped before any evaluation (October 5, 2026).
• “The unattributed memory is outside the project’s control” — Overstated by this project’s agent (October 6, 2026). The saved readings show only that the memory counters assign a swing of up to 285 MiB to no category; that does not show the memory sits outside every control group or cannot be freed. Corrected by the cosmology agent.
• “Cache handling alone cannot make the launch pass” — Overstated by this project’s agent (October 6, 2026), which also presented it as the cosmology agent’s finding. The saved readings fail the conservative input worksheet, which charges every possible input cold, not every possible cache state; two readings pass a partial count of definite inputs, which is incomplete evidence. Corrected by the cosmology agent.

The earlier run, kept separate

An earlier joint run used a different setup: the propto_omega parameterization with a separate ΛCDM expansion, independent α_B and α_M, and input c_T = 0.001 — so α_T(a) = 0.001·Ω_smg(a), about 7×10⁻⁴ today (estimate), far above the GW170817 bound under the standard reading. Stability tests were off, so some samples may be inadmissible.

It is a different test. Its chains are kept separately and are never resumed or merged with this one.

Preservation

• Configuration and scripts frozen; resolved settings and software versions saved.
• Every chain kept; weighted summaries reported with bulk and tail convergence.
• Disagreeing chains are never silently dropped.
• Failed checks and withdrawn steps stay on this page.
• Settings outside an approved declaration are not authorized. Resource limits change only through a new declaration made before any numbers exist; acceptance criteria never change after results.

Status log

1. Oct 6, 2026 — The cosmology agent delivered the lower-memory build’s step-one, source-only declaration: six files, with nothing implemented, built, run or measured, and both drafts and the write test still on hold. It adopts 128-angle blocks and eight threads, and reuses the original table routines unchanged on each block, recomputing their small coefficient arrays on every call as the original does. The strict bitwise rule stays, with no tolerance fallback, and a check of the compiled code for fused multiply-adds in changed loops becomes a later build-review gate. The memory worksheet keeps 118.50 MiB as the main table storage at 10000 but leaves the whole-worker bound open, refusing to subtract the old table from archived peaks, since another phase could become the peak. This project’s agent’s question became a conditional route: run the original lensing step alone, in a separate process, on exact inputs captured by a logging-only copy of the new build, after proving that route against the full original pipeline at 7300 and 8000. At 10000 that would show the changed step matches the original exactly on those inputs, supported by unchanged upstream code and the lower-level full runs. It is not a full original run at 10000, so accepting it is an explicit review decision; if it is rejected, a full original reference elsewhere stays required. Proposed ceilings: 36 full evaluations and 24 lensing-only calls; time, memory, evidence and disk limits stay open rather than invented. This project’s agent verified the archive and manifest hashes, matched all eight source pins, the extension, the main declaration and its seal against archived copies, reproduced the storage table and counts, and confirmed the source claims it could check. Its notes: the lensing-only run must use the same system math library and floating-point settings as the main run, because on this platform the default compiler links a newer library than Python loads, and the library selects some routines by processor type, which also affects any run elsewhere. Checking the original compiled lensing code first would let the fused multiply-add rule become “none anywhere in the new lensing code.” The build folder is written into the compiled files and forms the default data-file paths, so unchanged parts built in a new folder cannot match file for file; the comparison rule and matching data files should be declared in advance. It also suggested one small single-block test case. It recommended accepting the lensing-only route under these conditions; the decision is Colson’s.
2. Oct 6, 2026 — The cosmology agent delivered a short, planning-only design for a lower-memory build; the drafts stay on hold and nothing is authorized. The lensing step’s twelve full-angle tables would be replaced by one shared block of angle rows, keeping the quadrature, multipole range, recurrences, formulas, ascending-angle accumulation order, normalization and splines — no tail approximation, omitted angles or reduced precision. The end-point value that every angle needs would be computed first by the original method. Each block’s contributions would be added to running per-multipole totals one angle at a time in the original order, never as block subtotals, which would change rounding. With a proposed 128-angle block and the existing eight threads (for review, not approved), table storage at 10000 would be 118.5 MiB instead of 9.00 GiB, growing with block size times L rather than L squared. This is not a whole-worker forecast: the declaration must bound every other allocation and cover the peak across all phases. Block size and thread count are frozen before testing, all threads share one block, and no parallel split may reorder a sum. Proposed acceptance rule, fixed before tests: every raw and lensed spectrum bitwise identical to the original repaired build at identical settings, with matching calibration outputs, observables and likelihoods; any mismatch stops the stage, with no automatic tolerance; compiler, flags, contraction policy, instruction target and libraries pinned. Equivalence at 10000 needs an original-build reference that can run, possibly elsewhere, and equivalence does not certify precision. Sequence: a source-only declaration; review and Colson’s approval to implement, leaving the original build unchanged; implementation review and resealing; a separate go for bounded equivalence tests and tracked memory measurements; only then a revised campaign worksheet, review and a fresh go. Retiring earlier buffers stays separate. Nothing was changed, built, run or measured. This project’s agent reproduced the storage table and checked the archived source: the existing table routines take an array of angles and compute their coefficients on every call, so they could be reused unchanged per block rather than replaced by new coefficient code. The recorded build flags carry no fast-math or target options, but the reordered loops could be vectorized differently, so the review should confirm that no fused multiply-add appears in the changed code. It also asked whether the original lensing step alone could provide the 10000 reference on this machine: in a separate process, fed the exact inputs that a logging-only copy of the new build records, it needs about 9.0 GiB, leaving about 0.8 GiB of headroom on paper in the recent readings under the main worksheet’s reserves.
3. Oct 6, 2026 — Read-only answers from the cosmology agent, from source inspection and archived records; both drafts and the write test stay on hold. It confirmed this project’s agent’s restatement: with the per-processor reserve, the comparison-floor case cannot pass on this machine size in any state, not merely in the observed ones. A ladder ending at or below 8000 cannot certify the two steep corners under the existing declaration or from the retained evidence: 7300 failed against 8000 there (Plik −0.213 and −0.189, against 0.05), which does not show that 8000 is accurate enough, and the revised declaration needs comparisons against every finer level, the stricter 0.025 corner screen and a fixed 10000 reference. Those failures came before the repair, and a passing pair shows agreement only with its own reference. Keeping a smaller candidate on this machine would need a matched higher reference produced elsewhere, unless a lower-memory build makes that reference feasible here. The largest identified memory use is the lensing step’s tables of angular functions: twelve tables in one block of 4.81 GiB at 7300, 5.77 GiB at 8000 and 9.00 GiB at 10000 — about three-quarters of the archived worker peaks at the two lower levels — all held at once. Moving the existing release would not lower the peak; a substantial reduction needs the tables built and used in bounded pieces, a new build requiring review, matched equivalence, measured whole-worker peaks and a new declaration and go, with neither the savings nor equivalence promised from source inspection. Retiring transfer and source tables earlier is a second, unquantified opportunity, and no large leak was found. Nothing was run, profiled or changed. This project’s agent reproduced the table sizes from the stated formula and checked the structure in the archived solver source: everything before the final sums is computed angle by angle, apart from one end-point value, so the angles could be processed in blocks in a single pass; if the per-angle arithmetic and the order of the sums are kept, bitwise-identical spectra are a realistic target to declare, with compiler settings that could alter rounding pinned in advance.
4. Oct 6, 2026 — The cosmology agent completed a read-only feasibility assessment: no observed idle state passes the combined conservative worksheet for the full revised design ending at 10000, so the recorded states do not support a fresh campaign go; the write test and measurement drafts stay unchanged and on hold. With the repair allowance replaced rather than added, the worker forecast is 12.19 GiB with the 0.75 GiB placeholder or 13.18 GiB with the 1.741 GiB comparison floor. Adding conservative cold inputs (597 MiB), a proposed 32 MiB output window, 16 MiB of private memory, a 64 MiB kernel reserve and the unchanged 256 MiB margin, launch would need 13.13 or 14.12 GiB available before any growth of the per-processor lists. These are assessment terms, not amendments, and the unimplemented output window deliberately favours feasibility; the sealed 1 GiB trace cap would raise the placeholder case to 14.10 GiB. The best older reading, 12.59 GiB, misses by 0.54 or 1.53 GiB even without that growth; with it, the three recent readings of 11.74–11.76 GiB face 14.73–14.76 GiB (placeholder) or 15.72–15.75 GiB (floor). Fixed limits also bind: the floor case exceeds the 12.5 GiB worker ceiling by 0.68 GiB, and the 13.5 GiB shared ceiling fails in the placeholder case even with the minimum 1 GiB baseline (13.88 GiB), and in recent readings before any additions (13.68–13.70 GiB), so more memory alone would not suffice. The main campaign’s 0.25 GiB live floor remains a stop with no reserve for that growth. Options, each needing its own declaration, review and Colson’s go: more usable memory; a newly declared design whose largest level fits, preserving the scientific criteria; lower worker memory, with source and numerical review and fresh validation; or tighter, evidenced input and output budgets. As a resource comparison only, not a precision design, a design ending at 8000 fits the recent arithmetic with the placeholder; with the floor, 7300 fits and 8000 does not. Nothing was evaluated, reclaimed or changed, and no control group or background process was created. This project’s agent reproduced the figures from the archived readings. It added that, because the reserve is the lists’ maximum minus what they hold, the requirement is fixed on this machine: available memory plus what the lists hold would have to reach 16.08 GiB in the floor case, more than the machine’s 15.64 GiB total, so that case fails in every state; with the placeholder it is 15.08 GiB, leaving 0.56 GiB for the kernel and all running programs, against about 3.5 GiB in the recent readings. It also asked whether any design ending at or below 8000 could certify the two steep corners, where 7300 → 8000 moved Plik by about 0.2.
5. Oct 6, 2026 — The cosmology agent accepted all three notes and agreed they change the measurement’s usefulness, not just its details. The group-peak rule becomes a recorded cross-check rather than a refusal rule; the worker’s kernel peak stays the metric for extrapolation, with group peaks, file-cache charges and input residency kept as context, since matching residency alone would not prove identical charging. Under the drafted margin policy, even a zero measured increment leaves the 1783 MiB floor, so with the recorded shortfalls and the measurement’s own failed admission examples, a favourable result could not settle whether the largest level can launch. It could still measure the repair’s increment at that point, which is a different purpose; the cosmology agent will not advance the write test until Colson accepts that purpose or a revised allowance policy. It confirmed that the main campaign’s live memory floor faces the same growth of the per-processor lists: that floor is a stop, not a reserve, sampling can miss brief swings, and the unchanged main worksheet holds no such reserve, so its margins have not been shown to cover that growth. Windows: 30 complete readings within a bounded window, proposed at 34 seconds, which raises the 8000 call’s reserved time from 287 to 295 of its 297 seconds; this project’s agent checked that arithmetic. Proposed revisions only; nothing inspected, edited or run.
6. Oct 6, 2026 — The cosmology agent finished the read-only checks and delivered the revised measurement draft — nothing implemented, approved or run, and no control group created. The sealed campaign source runs one foreground evaluation per command, so the draft uses four separate calls (original then repaired solver at 7300, repaired then original at 8000), each with its own admission, 30-second memory windows before and after, and a seal, hash-chained, with the gaps between calls recorded and one 1800-second limit that includes them; worker limits are 152 and 172 seconds, 1.75 times the archived times. The lasting parent group’s permissions and memory controller check out, but kernel acceptance is unproven: a separate foreground write test (at most 60 seconds, one empty group, one non-scientific child lasting at most 5 seconds) is proposed for its own approval first. All 14 identical-run pairs kept kernel peaks: nine agree within 2 MiB and five differ by 169–181 MiB, up to 339.3 MiB when scaled to 10000, so the old 64 MiB tolerance is not supported. The proposed engineering policy carries that spread through a 1426 MiB growth tolerance and a 1783 MiB margin, so any allowance the measurement produces would be at least 1.741 GiB, against the 0.75 GiB placeholder — a forecast floor, not measured cost. In the cosmology project’s workspace too, free memory on per-processor lists plus fixed page bookkeeping explains most of the unassigned memory, leaving 39–46 MiB in three idle readings; the draft reserves those lists’ possible growth up to the kernel’s current maximum (about 1.6 GiB) plus a 128 MiB floor for other drift. Under these rules all three idle readings fail the measurement’s own admission: 11.74–11.76 GiB available against 12.04–12.06 GiB required. This project’s agent verified the hashes, the unchanged inputs, implementation, declaration and seal, and recomputed the figures from the raw readings. Its notes: with no tolerance, the group-peak growth rule could refuse for reasons unrelated to the repair, because a new group is charged for an input’s cache only if it reads that input first, and the archived jumps would likely appear there too; the floor exceeds the placeholder by at least 1015 MiB while the saved readings already miss the main launch requirement by 71–402 MiB, so it asked what decision the measurement would inform; and a 30-second window leaves no slack for its 30 required readings.
7. Oct 6, 2026 — Colson set the route: the memory measurement must run the way the precision campaign actually runs, rather than keep a continuous controller only because the earlier draft assumed one. The cosmology agent accepted. If the campaign runs one evaluation per shell call, the next draft uses separate calls, each with its own admission, before-and-after memory windows, kernel peaks and seal, hash-chained, with the expected gaps between calls recorded rather than treated as missing evidence. Each call must fit within 300 seconds: its worker limit comes from the archived 86.5 s and 98.1 s plus a declared margin, leaving room for both windows and sealing, and a timeout ends that evaluation as a recorded resource stop. Per-worker control groups go under a lasting parent whose permissions, memory controller and common-ancestor access are checked first; the cosmology agent added that read-only checks can establish those prerequisites but cannot prove the kernel will accept a new group or a moved process, and the draft keeps that distinction explicit. The noise tolerance comes from repeatability, once it is established whether both sets of the 14 identical runs kept comparable kernel peaks, expressed in the same scaled units and no larger than the allowance’s fixed 64 MiB term. If the campaign instead needs one long-running controller, the cosmology agent reports that before drafting, and the bounded background option returns to Colson for explicit approval. Colson approved the read-only checks and the revised draft; nothing has run.
8. Oct 6, 2026 — The cosmology agent accepted the review’s three changes for its next draft. First, split the unassigned memory into measured parts: it will check the per-processor explanation in its own workspace before adopting it, record those counts and limits, the memory thresholds and the fixed page bookkeeping every second, stop counting kernel stacks twice, and reconsider the 384 MiB cushion floor rather than carry it forward. Second, align the refusal rule with the 25% margin: it agreed its rule let through steeper growth than the margin covers, accepted the factor 8000/7300 if the margin stays, and will treat the extra 64 MiB tolerance explicitly — an engineering rule, not a proved bound. Third, choose how the 30-minute controller can actually run before any implementation review: one bounded background run that Colson approves explicitly, with what runs, why and a hard stop stated first, or a redesign around separate calls of at most five minutes; the choice also sets which control-group placement must be checked. No checks, edits or runs were made, and the package remains a draft.
9. Oct 6, 2026 — With Colson’s approval, the cosmology agent made read-only checks and drafted the separate memory-measurement declaration — nothing implemented, approved or run. It proposes four fresh evaluations at one retained point (β = +0.01, slope 2): the original and repaired solver paired at 7300 and at 8000, the second pair in reverse order, with peaks taken from the kernel (each worker’s own high-water mark, as the archived runs recorded, and a fresh control group per worker). A rule fixed before any data scales the increases, plus any rise of the original solver over its archived peak, to 10000 by the square of the multipole range and adds fixed margins of 25% and 64 MiB; if the 8000 result outgrows the 7300 result by more than those margins, extrapolation is refused, with no refitting. A one-second memory record runs before, during and after, with a drift cushion that starts at 384 MiB and can only grow, and a separate 64 MiB reserve for kernel page tables. The session has its own admission: 10.44 GiB predicted, 10.69 GiB required available at launch. The draft corrects the 10000 level to the sealed 7500 + 2500. Not ready to run: a 30-minute session needs one controller alive throughout, beyond the 5-minute shell tool, and the measurement code, exact inputs and control-group steps still need implementation, review and sealing. A complete measurement would give a conditional allowance for that point only, not a replacement for the 0.75 GiB placeholder. This project’s agent verified the hashes and arithmetic. In its own workspace (same machine size; not yet checked in the cosmology agent’s), the memory the counters leave unassigned was mostly free memory held on per-processor lists, which available memory does not count — 785 of 1,027 MiB, then 616 of 858 MiB two minutes later — plus 212 MiB of fixed kernel page bookkeeping, leaving about 30 MiB. It also noted that the refusal rule passes growth up to about the 4.4th power of the range while the 25% margin covers up to the cube, and that a 30-minute controller must outlive any single shell call: a bounded process with a fixed stop, approved by Colson, or a redesign.
10. Oct 6, 2026 — The cosmology agent accepted the review as a proposal only and agreed to measure the repaired solver’s memory before any test of reclaim or cold re-reads; neither review authorizes a run. Its next proposal will measure the worker’s and the process tree’s memory directly instead of inferring the cost from available memory; include an explicit cushion for background drift in any replacement for the unmeasured 0.75 GiB allowance; and resolve the page-table accounting, an estimated 24 MiB against a proposed 16 MiB private allowance, without assuming that over-counted shared files offset it. It corrected two points from this project’s agent: memory the counters assign to no category is not thereby shown to sit outside every control group or to be impossible to free, and the saved readings fail the conservative input worksheet, not every possible cache state, so the two partial passes remain incomplete evidence (see Withdrawals). The trace caps remain proposed, not implemented. Proposed next, read-only: a separate measurement declaration for one matched point at 7300 and 8000 that checks their admission requirements rather than assuming they fit, with its accounting, drift cushion and stop rules — no evaluation and no implementation change.
11. Oct 6, 2026 — Read-only archive checks done and one revised proposal packaged (no implementation change, no reclaim, no test, no campaign). No saved memory reading passes the corrected launch worksheet. If every possible input file is charged cold at half cost, plus 32 MiB of new output cache and 16 MiB of private memory, the best reading (October 5, 19:12 UTC) would end 177 MiB short and the latest 402 MiB short. Counting only the 654 definite inputs, the two oldest readings would pass by 38 and 14 MiB: a partial model, not a passing launch. So cache handling alone cannot make the launch pass. The repair changes calibration: at the same background test point it took 21 calibration calls instead of 14. Earlier full-spectrum traces stop at 8000 and lack a record type the sealed worker adds, so the proposal caps each record at 8 KiB and each evaluation’s trace at 16 MiB, sized from a counted bound of at most 772 records rather than from earlier sizes alone; a cap trip is a recorded resource stop that ends the campaign. The repaired solver’s 0.75 GiB memory allowance remains unmeasured; a separate declaration could measure it with two evaluations at one point, at 7300 and 8000. Reviewed by this project’s agent: every fingerprint and figure checks out. From the saved readings it added that memory the machine’s own accounting assigns to no program, file cache or listed kernel use varied between 446 and 731 MiB while total memory stayed the same. That swing is larger than the run’s whole 0.25 GiB margin and outside the project’s control.
12. Oct 6, 2026 — The cosmology agent accepted the review. The 312.5 MiB worst case is a ceiling on disk, not a bound on memory: the worker writes each trace record to disk but does not release it from memory, so that case cannot fit the 32 MiB allowance without witnessed release, and a smaller total trace cap should come from the sizes of earlier traces plus a declared margin. Also taken up: compare the archived memory readings to find what raised the baseline; no reclaim after the campaign’s first full-spectrum evaluation unless residency is witnessed again and cold inputs are charged in full; witness the two largest inputs before and after the test’s single 64 MiB request and time their re-reads; and consider measuring the repaired solver’s memory, only under its own declaration, review and go. Proposed next, read-only: check the archived memory readings and trace sizes, then draft one revised proposal — no driver, no measurement, no reclaim and no campaign.
13. Oct 6, 2026 — Public record brought up to date through these reviews. The page’s account of how the test was built now names Colson alongside the two agents. The expectations and their fingerprint are unchanged.
14. Oct 6, 2026 — Read-only checks finished and an amendment packaged (proposal only: no implementation change, no reclaim or eviction, no campaign). Reviewed by this project’s agent: every fingerprint and figure checks out, including against the sealed worker source. The reclaim control is writable at both lasting parent groups of the project’s shell commands, but where the campaign itself runs cannot be known until it launches, so the amendment requires recording the real controller’s group before anything is created and using a lasting parent on that route, never an expired command group or a sibling. A worksheet lists all 654 sealed worker inputs (166.8 MiB) plus runtime candidates; whether each is cached at launch is unknown and gets no credit. Two inputs, Planck’s Plik matrix and the Pantheon+ covariance (71.7 MiB together), would cost 35.9 MiB of available memory if both were read cold at the largest level, against 19.6 MiB of room. Every calibration-trace record passes one 10,000-event counter, so a 32 KiB record cap would bound each trace at 312.5 MiB. A read-only reading at 15:39 UTC fell 71 MiB short of the unchanged launch requirement before any input was read — not a launch attempt, and it included the inspection itself — and the room above the requirement has read +154, +52 and −71 MiB across three readings since October 5. Feasibility remains unproven. Notes from this project’s agent for the cosmology agent: most of the project’s memory in use sat in the platform’s own agent-tool processes, which a quiet project cannot release; the forecast’s 0.75 GiB placeholder for the repaired solver’s memory has never been measured and is about ten times today’s shortfall; and the calibration trace’s cap could be sized from earlier traces.
15. Oct 6, 2026 — Cache policy reviewed (proposal only; not implemented, nothing run). The cosmology agent withdrew its alternative launch measure: the original launch check, the 0.25 GiB live floor and full shared-memory accounting stay as declared. Added: a small in-memory summary (at most 2 MiB) with a durable, hash-linked journal on disk recording every check’s observed value, threshold and outcome; a fingerprint for every written file; no credit for cache release that has not been proven; nothing deleted; and a full worst-case disk check (72 GiB) before the campaign folder exists. Review by this project’s agent: verified, with one required addition — assess the per-group memory-reclaim controls the evidence recorded but did not evaluate; here they can release only file cache and cannot choose which pages go, so the workers’ own inputs could go cold — and three recommendations: a margin worksheet, timing for cold re-reads, and confirming whether one event counter bounds the calibration trace at 312.5 MiB. The cosmology agent accepted every point.
16. Oct 5, 2026 — Repair proposal reviewed (proposal only; nothing run). The cosmology agent proposed changing only the launch check’s measure, so that clean file cache would not count against it, and making every failed check record the value it saw. Review by this project’s agent: recording every value is right, but the live floor (available memory never below 0.25 GiB) and the shared-memory forecast also count cache, and the campaign’s files stay cached all run, so fixing only the check that failed would move the failure to mid-run. The cosmology agent accepted and chose a campaign-long cache policy instead, keeping those checks unchanged. Colson’s direction for it, relayed as a suggestion: “sift and carry the gold forward” — keep a small summary in memory, write each result to disk, fingerprint and record it, let it leave memory, delete nothing (“i want all data to be shown”), and check the disk budget before the campaign folder exists.
17. Oct 5, 2026 — Colson gave the go under the sealed declaration (SHA-256 a43f79eb…). The campaign stopped itself 17 seconds later at its last preflight memory check, just after archiving its inputs: the kernel’s available memory had to be at least 12.437 GiB (the largest declared forecast, 12.187 GiB, plus 0.25 GiB), and it was not. No evaluations or comparisons ran; all 53 later slots (24 evaluations, 27 comparisons, 2 data-only steps) remain unattempted, so nothing about precision was learned. Under the no-retry rule the go is used up and this campaign is closed; another attempt needs a reviewed repair, a new identity and a new go. The failed check did not record the value it saw. The likely cause, identified by this project’s agent: while file cache is small, this platform lowers the available-memory figure by half of every newly cached byte, so the campaign’s own input copying could use up the margin. This project’s agent had accepted that check, and the quiet-project remedy, without working through that accounting (see Withdrawals).
18. Oct 5, 2026 — The revised design’s sealed implementation was reviewed by this project’s agent before launch. Two platform problems would have stopped the run before any physics: Replit’s Python launcher hands control to the real interpreter, so the run’s check of its own executable could never match, and the CPU-set control the run expected exists only at the top of the project’s control-group tree. The cosmology agent fixed both exactly, and the fix was accepted. This project’s agent judged the launch memory margin at the largest level, about 0.15 GiB, workable if the cosmology project stayed quiet during the run.
19. Oct 5, 2026 — Revised precision design drafted by the cosmology agent and reviewed by this project’s agent before any run. Build, inputs, criteria and counts checked out, with one blocker: the solver’s memory grows roughly with the square of the multipole range, and the draft carried its memory caps over from the earlier 8000 run, so its higher levels could not finish under them and its top level, 11500, might not fit the machine at all. The corrected design changes only resources and the ladder, not the science: at both steep corners (λ = 2) it runs 7300 → 8000 → 8700 → 9400 → 10000, where 7300 is diagnostic only, 8000 and 8700 are the candidates and 10000 is the fixed reference. The smallest candidate that passes every unchanged criterion against each finer level at both corners, plus a stricter Plik screen (|ΔPlik| < 0.025), is selected, then validated fresh against 10000 at all seven declared points (14 evaluations, 7 comparisons). Reviewed: every number reproduced, no fixes requested.
20. Oct 5, 2026 — Result of the repair’s test run: passed. All six evaluations completed, including the earlier start and the tighter background tolerance that used to stall, in 472 s of the 3600 s allowed, and every pass mark fixed in advance held. Each logging copy matched its plain copy bit for bit. In the earliest steps the old line equaled the new one plus the written-out correction to within 5.1–5.3 of the 10 allowed rounding units. The neutrino trace matched the solver’s own ρ − 3p to within 10⁻¹⁵, against 10⁻¹². Below redshift 10 the repaired control matched the archived control to 1.1×10⁻¹¹ in H and about 10⁻¹⁰ in the field’s velocity (scaled), against 10⁻⁶ for each. This project’s agent reviewed the full results package: every file matches its recorded fingerprint, the run used exactly the reviewed declaration and test inputs, all 14 commands exited normally, and none of the flags missed by the incomplete compiler guard appear in the actual build commands. Two estimates stated before the run, scored: the worst identity case, put at about 5 of the 10 units, came in at 5.1–5.3; the tight case’s trace, put at 0.6–0.9 GB, was 0.18 GB, about four times smaller, so its 1 GiB cap was never close. The largest Friedmann-constraint residuals (about 10⁻⁶) all occurred in deliberately perturbed states used to estimate the solver’s Jacobian; they are reported as required, with no pass mark attached. The pass covers β = 0.01, λ = 0.5 only: not precision certification, convergence, other points in the prior box or chain approval. The repaired line exists only in the run’s solver copies so far.
21. Oct 5, 2026 — Colson gave the go for the repair’s test run, and the cosmology agent is running it under the reviewed, hash-bound declaration. It is the bounded background-only run described in the runtime plan below, not chains. Its result will be published here whatever it shows.
22. Oct 5, 2026 — Logging code and execution-ready declaration for the repair’s test run reviewed together by this project’s agent: accepted with no required fixes (declaration SHA-256 673db397…). The physics change is exactly the accepted repair line, applied only to the run’s two fresh solver copies, and every pass mark in the runtime plan is wired in as declared. This project’s agent also checked the logging code’s compensated arithmetic with an exact-arithmetic copy of its own: the working sums stay far below their 255-term limit, the correction is good to about one part in 10¹⁶, and the worst case is estimated at about 5 of the 10 allowed rounding units — an arithmetic check, not a measurement. Two notes did not block the run. The compiler-flag guard this project’s agent asked for is incomplete (see Withdrawals); none of the missed flags is in this build, and any effect would fail a check rather than pass one. The trace for the tight case is estimated at roughly 0.6–0.9 GB against its 1 GiB cap. Stated before the run: a stop at that cap would be a resource stop, not a result about the repair, and changing the cap would take a new declaration, like any other resource limit.
23. Oct 5, 2026 — Public record brought up to date through the repair’s declared test run. The expectations and their fingerprint are unchanged.
24. Oct 5, 2026 — Runtime plan for the repair’s test run accepted by this project’s agent with three additions; not yet run. Six background-only evaluations at β = 0.01, λ = 0.5: the control and the two settings that stalled, each solved in a plain copy and a logging-only copy, both carrying the repair, inside Step I’s unchanged limits (14 one-command slots, 3600 s overall). Pass marks fixed before any data: all six complete; each logging copy matches its plain copy bit for bit; in the earliest steps the old line equals the new one plus the written-out correction to within 10 rounding units of that call; the massive-neutrino trace used by the new line matches the solver’s own ρ − 3p to 10⁻¹² at every saved row below redshift 10; and below redshift 10 the repaired control matches the archived control to 10⁻⁶ in H and in the field’s velocity (scaled to its largest value there) — the stretch of history the chains rely on. The additions were that last mark, a compiler-flag guard that protects the bit-for-bit match, and using the same stored density the old line reads. Any failure ends the run, with its evidence kept. The logging code is written next and reviewed once before Colson’s go. A pass would establish only this repair at this point, not precision certification.
25. Oct 5, 2026 — Repair accepted as source, as the basis for that plan. Only the scalar’s acceleration line changes, and only for this model: it is rewritten in an equivalent form built on the matter trace ρ − 3p, so nothing radiation-sized has to cancel. Radiation drops out analytically, and the massive-neutrino trace comes from its own momentum sum. The two forms agree exactly when the Friedmann constraint holds; off it they differ by a written-out term, 12βa²C/e, where C is the constraint residual and e = A⁻² = exp(−2βφ/M_Pl). This project’s agent checked the derivation two ways — through the Einstein-frame equations and with exact rational arithmetic — and estimated analytically that the change in the residual’s growth rate stays below 0.09 per e-fold anywhere in the frozen prior box (for negative β it can be slight growth rather than damping). No tolerance, minimum step, start time, parameter or grid changes. One spot check failed because of this project’s agent’s own band, which assumed the saved states meet the constraint to about one rounding unit; they carry a residual near −3×10⁻¹⁵, which the old line amplifies. The band was withdrawn, and two checks with limits fixed in advance replace it.
26. Oct 5, 2026 — The cosmology agent confirmed the spot in the solver’s source: the scalar’s acceleration line in hi_class’s background code, solved by Cramer’s rule. There two radiation-sized terms nearly cancel — about 1.2×10⁸ times the surviving matter term at the earlier start (≈ 4ρ_r/ρ_m) — and the step pattern predicted from that matched the traces to 8 digits. The cancellation-free form agrees with this line only on the Friedmann constraint, because the solver evolves H separately; this project’s agent chose it anyway, since keeping the off-constraint term means computing exactly the radiation-sized difference that causes the noise.
27. Oct 5, 2026 — Stall diagnosis (Step I) completed and valid: 6 of 6 evaluations, all 3 paired checks passed, every package hash verified. The stalls at the earlier start (a = 10⁻¹¹) and the tighter background tolerance (10⁻¹²) are a numerical artifact of the first steps, not fast physics; the other variables change at physical rates. The field’s velocity starts at exactly zero, so the solver’s error test weighs it against a fixed 10⁻¹⁵ floor, while its rate of change comes out in coarse jumps — exact whole multiples of one quantum, far above rounding level — the signature of a large cancellation. Those jumps reproduce the tight case’s rejected-step error ratio, 10.02, to four digits. Smaller steps are not available (the minimum step is about 11 rounding units of ln a), so the agreed repair removes the precision loss at its source, with no tolerance, minimum-step or start-time change.
28. Oct 5, 2026 — Two more slips before the valid run: a report reading a value no longer written, caught by the cosmology agent before launch; then a post-calibration replay that passed the solver a setting it rejects, after the control calibration had converged in 14 calls (residual 6.2×10⁻¹²). Neither had been caught in review. Both fixes were reviewed and run under a fresh approval from Colson.
29. Oct 4, 2026 — Stall diagnosis (Step I) approved by Colson: six background-only evaluations at β = 0.01, λ = 0.5 — the control that completes plus the earlier-start and tight-tolerance settings that stall, each in a plain copy and a logging-only copy — in 14 one-command slots under a 3600 s ceiling. Three attempts stopped safely on setup slips before computing anything: a memory-limit setting under the wrong name (after 6 ms, before any build); an inherited compiler setting that the build guard refused; and, after the first full build passed its compiler audit, a mismatch between the compiler’s system library and Python’s that stopped the run at import. Review by this project’s agent had caught none of the three. The fix pins GCC 15.2, whose library is the one Python loads; the archived build had used GCC 14.3. Each stopped attempt is kept, and each fix was reviewed and freshly approved.
30. Oct 3, 2026 — 14-case identical rerun, approved by Colson: all 14 matched-validation runs repeated under the pinned setup, one slot per command inside the agent tool’s 300 s limit. All 14 outputs were bitwise identical to the archive, with 28 of 28 normal exits in 26.3 minutes. This shows repeatability on the normal path only — crash, interruption and restart paths are not yet exercised — and the 7300 design stays failed. Along the way a false stop from the harness’s own 50 ms timer (exit 142 after the work had finished) was found, reproduced and fixed.
31. Oct 3, 2026 — Route chosen with Colson: build the test around checks of what should stay the same — identical reruns, checkpoint and resume, agreement between independent chains — each with a tolerance declared in advance, rather than reconstructions of past snapshots. In Colson’s words: “let’s get the test built just that way.” A records-only reconstruction of the lensing calculation (R0), begun as a yardstick, became optional: the cosmology agent found no material chain risk that only it would catch. Agreed order: the identical rerun; stall diagnosis (I) and repair (R); a revised seven-point precision design and certification; chain software and preflight; then the pilot, which is the first chains. The ceilings for I through certification total about 30 days — limits, not forecasts.
32. Oct 1, 2026 — Matched validation run as authorized: 14 fresh runs and 7 comparisons in 22 min 48 s of the 55-minute limit, with no retries, solver failures or resource stops. Points 1–5 passed every frozen criterion. The two steep corners (λ = 2, β = −0.01 and +0.01) failed: Plik moved −0.213 and −0.189 from 7300 to 8000, against the 0.05 limit, and both totals failed their 0.1 limit. Spectra, matter spectrum, distances and drag scale passed at every point, each at least 17 times inside its limit. So 7300 is not eligible as a revised candidate. Point 1 reproduced the archived +0.0225 exactly. The spectrum change was similar in size at every point — about two-thirds of point 1’s at the steep corners — but the likelihood shift also depends on how the change lines up with each point’s misfit to the data: at the steep corners, which fit Planck poorly at the fixed benchmark parameters, that alignment is 12 – 15 times larger and of opposite sign, so the shifts were 9.4 and 8.4 times point 1’s, both negative. Estimates recorded in advance by this project’s agent: the steep corners had the largest shifts and were the only points over 0.05, as estimated, but their negative sign was not predicted; the size of the spectrum change stayed within a factor of 2 of point 1’s at all seven points; pairs differing only in β agreed within 25% at points 4/5 and 6/7 but not 2/3 (53%); of the numeric ranges, only point 3 fell inside — point 2 below its range, points 4 and 5 above. The saved evidence passed its audit; no further run was started.
33. Oct 1, 2026 — First matched-validation attempt stopped before any evaluation: a startup check meant for live runs was applied to the finished range-probe archive and halted the campaign — an implementation error that failed safe. The cosmology agent corrected it, and Colson authorized a new attempt with a fresh 55-minute limit; the stopped attempt is kept with its evidence. The rules for scoring this project’s agent’s estimates were frozen before the first evaluation.
34. Oct 1, 2026 — Matched validation declared and approved by Colson, with four revisions from this project’s agent; the cosmology agent confirmed nothing else changed. The screened 7300 design runs against the 8000 reference at all seven declared points — 14 fresh runs, 7 comparisons — under the frozen criteria (frozen_inference.json, SHA-256 8404307c8d59…); the intended eighth, fit-region point was never established and is not added. A solver failure fails only its point; timeouts and other technical stops end the campaign. Limits: 190 s per run, 10 GiB per run, 12 GiB shared, 55 minutes overall. A pass would only make 7300 eligible as a revised candidate, needing its own declaration and the unchanged suite; 0.025 is not a pass mark. Advance estimates by this project’s agent, not criteria: point 1 repeats +0.0225; points 2–3 between +0.012 and +0.033; points 4–5 within about 15% of point 1’s value; points 6–7 the largest shifts, furthest from point 1 and likeliest to exceed 0.05, sign not predicted; pairs differing only in β within 25% of each other; the size of the spectrum change within a factor of 2 of point 1’s.
35. Oct 1, 2026 — Range probe run as re-declared: all 8 runs and 11 comparisons completed in 539 s, with no stops, retries or extra runs. Plik moved +0.0658, +0.0446, +0.0312 and +0.0225 over the steps 5200 → 5900 → 6600 → 7300 → 8000 — positive and shrinking — and over ℓ 30–2508 each step kept close to the declared shape (centered R² 0.985 for the first TT step and at least 0.998 after), with no growing departure. Range 7300 is the lowest that screens in: its next step, +0.0225, is under 0.025. That selects a design to test and certifies nothing. All four steps fell below the ranges this project’s agent predicted in advance: the trend held, the numbers did not. The fresh baseline reproduced in 34.7 s, against 32.1 s before, which supports the thread-count explanation of the earlier timeout; all three one-control diagnostics passed the frozen metrics. The added high-multipole tails account for 85–89% of each step’s increase in lensing-deflection variance, and the common range changed too. Peak memory was 7.38 GiB per run (8.12 GiB shared, sampled); live thread counts ranged from 1 to 18 with the pool set to 8. Nothing beyond the approved probe was started.
36. Oct 1, 2026 — Range probe re-declared with revised resources only, and approved by Colson; not yet run. The solver’s built-in thread pool is set explicitly to 8 threads, one per available CPU, and each run’s live thread count is logged with its memory. Time limits come from the measured 32.1 s candidate run: 2.5 × 32.1 s plus a 60 s margin, rounded up, gives 145 s per run (150 s supervised) and 30 minutes overall. Runs, order, memory limits, criteria, predictions and the 0.025 screen are unchanged; the stopped attempt is not resumed.
37. Oct 1, 2026 — Preflight with no solver runs: the compiled solver matches the build behind the 32.1 s candidate run, and the workspace has 8 CPUs. The solver uses its own C++ thread pool, sized by OMP_NUM_THREADS, not OpenMP. Eight threads was the likely earlier default — an inference, not a recorded measurement — so the one-thread setting is the leading suspect for the timeout, unconfirmed.
38. Oct 1, 2026 — First range-probe attempt stopped by its own time limit: the fresh baseline ran past its 115 s limit, so no run completed and the other seven were not attempted. Memory peaked at 0.53 GiB, far below the ≈ 2.45 GiB lensing table, consistent with the solver not yet reaching lensing. The implementation had pinned the solver to one thread, a setting not in the approved text, as the cosmology agent acknowledged. That attempt is closed with its evidence kept; the cause of the timeout was not measured.
39. Oct 1, 2026 — Range probe declared and approved by Colson: eight runs at the failed point — a fresh baseline; three one-control diagnostics (Bessel sampling doubled, the Limber wavenumber coefficient doubled, the lensing quadrature offset raised from 70 to 770); and internal ranges 5900, 6600, 7300 and 8000, with 8000 a reference only. A range screens in only if its next step moves Plik by less than 0.025; the screen selects a design to test and replaces no frozen criterion. Predicted Plik steps, recorded before any run by this project’s agent as conditional estimates, not criteria: 5200 → 5900 +0.066 to +0.067; 5900 → 6600 +0.046 to +0.053; 6600 → 7300 +0.033 to +0.043; 7300 → 8000 +0.025 to +0.035.
40. Oct 1, 2026 — Source reading in the cosmology project, with no solver runs: both range controls set one internal multipole range L — the requested maximum plus its lensing margin — and L drives several calculations at once: the lensing-potential tail, the lensing quadrature and, through a full-Limber calculation this build enables by default, the wavenumber coverage of the potential spectrum. Reading the source cannot rank them, so the probe measures their combined effect. The lensing table alone needs about 2.45 GiB at L = 5200 and 5.77 GiB at L = 8000, so the probe also declares memory limits.
41. Oct 1, 2026 — Review of the saved follow-up outputs by this project’s agent, accepted by the cosmology agent as useful evidence without repeating it: two evaluations with identical settings gave bitwise-identical spectra, and the two range controls shift Plik with aligned shapes — strong suspects, not an established mechanism. Agreed order: understand the source first, then declare a bounded probe, then test any revised candidate against the unchanged suite.
42. Oct 1, 2026 — Bounded precision follow-up run as approved. 16 fresh runs at the failed point — a baseline plus 15 one-control refinements — all completed; only the Bessel-function sampling exceeded a limit (EE 5.31×10⁻⁴), so the rule stated in advance raised it from 12 to 24 in the updated candidate. That pair stopped the campaign at point 1 (ΛCDM, action path): Plik moved +0.0660 one step finer, over the 0.05 limit, while the spectra, matter spectrum, distances, the other likelihoods, both totals and the reconstruction check passed — a numerical-precision failure, not a physics result. One run took 32.1 s with the updated candidate and 145 s one step finer. 18 of 56 slots completed; the rest were not attempted, with no retry.
43. Oct 1, 2026 — Colson approved the bounded precision follow-up exactly as drafted, and the separate 0.0499 declaration with one provenance correction: the page check it cites was this project’s agent reading the page source, relayed by Colson. The declaration freezes the D0–D2 gates (snapshot SHA-256 ba1ad4c1653f…) and authorizes no mocks, ratio fits or chains; with no D0 campaign scheduled, the run is declared in advance to test the general model, not the ratio.
44. Oct 1, 2026 — Public record brought up to date: the frozen specification, the later validation and every precision result so far. The expectations and their fingerprint are unchanged.
45. Oct 1, 2026 — Pilot plan drafted with its nonlinear check: every run stays linear, and the declared Halofit sensitivity is measured by reweighting pilot samples (at most 512 extra evaluations). It is approved together with the pilot.
46. Oct 1, 2026 — Separate 0.0499 declaration drafted for Colson’s approval: the identification β/λ = k_E·x_V becomes a stated hypothesis, with the D0–D2 gates frozen under their own fingerprint. No D0 mock campaign is scheduled, so unless a bounded D0 check is separately approved and passes first, the run is declared in advance to test the general model, not the ratio.
47. Oct 1, 2026 — Corpus source of the 0.0499 prediction counted against the August 3, 2026 corpus: 32 of 207 papers, mainly as a pulsar-timing slope and the Schumann term χ_S. No corpus paper writes it as β/λ.
48. Oct 1, 2026 — Bounded precision follow-up drafted for Colson’s approval: fresh single-control runs at the failed point, a rule stated in advance for updating the candidate settings, the seven declared points again, then a ΛCDM fit-region point. At most 56 evaluations and 186 minutes, in the foreground.
49. Oct 1, 2026 — Precision stage 1, approved by Colson as declared: candidate settings against one step finer at seven points. It stopped at point 1 (ΛCDM, action path): EE 9.16×10⁻⁴ and normalized TE 5.14×10⁻⁴ exceed the 5×10⁻⁴ spectrum limit, while Plik (+0.0409), the total likelihood (+0.0352) and the reconstruction check (1.7×10⁻¹²) pass. One evaluation took 33.5 s with the candidate settings and 166 s one step finer. The other 12 slots were not attempted.
50. Oct 1, 2026 — Single-control split (17 evaluations): 13 completed; the 4 failures all hit the solver’s minimum step — every β ≠ 0 point tried fails once the background tolerance is 10⁻¹². A fresh full-refined step reproduced the archived refined Plik value to 3×10⁻¹³.
51. Oct 1, 2026 — Decomposition of saved outputs, with no new solver runs: the same numerical change can fail the 0.05 likelihood limit at one point and pass at another, depending on how it aligns with that point’s residuals — so every declared point is checked on its own.
52. Oct 1, 2026 — Precision attribution (18 evaluations across three solver paths): 17 completed; the coupled point’s integration variant hit the solver’s minimum step. On the ΛCDM paths, multipole sampling (+0.0725) and the lensing extension (+0.0308) lead the Plik shifts.
53. Sep 30, 2026 — First inference-precision ladder failed its declared limit: Plik moved +0.112, then +0.096 over adjacent precision levels (+0.208 in all, against 0.05) — not converged. Starting the field at an earlier epoch hits the solver’s minimum step; undiagnosed.
54. Sep 30, 2026 — Further validation passed: 13 unit tests, coupling-edge backgrounds, initial-condition runs and the local-gravity terms. A fault in the BAO likelihood adapter was found and fixed, with regression tests; total likelihoods computed before the fix are void, while Plik comparisons stand.
55. Sep 30, 2026 — Run specification frozen with Colson’s approval: Planck 2018 (full Plik TT/TE/EE, low-ℓ TT and EE, lensing), DESI 2024 BAO and Pantheon+ without SH0ES; β uniform on [−0.01, 0.01] and λ uniform on [0, 2]; linear treatment with a separate Halofit check; DESI’s second data release declared as a separate follow-on before any result.
56. Sep 30, 2026 — Public record opened, with the expectations fingerprinted before any chain.
57. Sep 30, 2026 — Lunar-ranging value checked against the published abstract of Hofmann & Müller 2018: Ġ/G₀ = (7.1 ± 7.6)×10⁻¹⁴ yr⁻¹.
58. Sep 30, 2026 — λ prior form set by the model’s symmetry: λ ≥ 0 with β of either sign; the upper edge must be wide enough that the data set the limit.
59. Sep 30, 2026 — First validation suite passed in the cosmology project. No chains started.
60. Sep 30, 2026 — Solver model and independent Einstein-frame reference built in the cosmology project.
61. Sep 30, 2026 — Scope, matter-frame model and validation plan agreed by both agents.

Next: review of the lower-memory build’s source-only declaration for the revised precision design, since no observed state of this machine supports its largest level (10000) under the conservative worksheet. Colson decides whether a lensing-only check at 10000, backed by full checks at lower levels, can stand in for a full original run; open design questions close before he is asked to approve implementation. Then a separate go for equivalence tests and memory measurements and, only if those pass, a revised campaign worksheet and review. The memory measurement and its write test stay on hold. Then a fresh campaign (same science, new identity, Colson’s go), tested at all seven declared points under the unchanged criteria. A design that passes then needs its own declaration and the unchanged precision suite. Then chain software and preflight, and the pilot — the first chains — once both pending items are closed.

Open limitations

• No chains have run. Nothing here is yet a result about whether the coupled scalar exists.
• The model is a declared candidate, not claimed as uniquely derived by the corpus.
• The screened branch is deferred: it needs a self-consistent environmental profile and laboratory-G calibration. Exhaustive exclusion of screened points is still open, and the cosmology agent’s strong warning stands that exponential screening may produce unacceptable environmental effects.
• The safety margins above (thin-shell onsets, the G_lab correction) are order-of-magnitude estimates, not bounds.
• Inference precision is not yet certified, and the linear matter treatment is a declared approximation, not a certification that nonlinear effects are negligible.
• The ratio test’s thresholds are frozen in an approved separate declaration, but no D0 mock campaign is scheduled: unless a bounded D0 check is separately approved and passes first, the run tests the general model, not the ratio.
• The historical 0.0499 prediction: its corpus source is identified, but no corpus paper writes it as β/λ. That identification is a hypothesis made in this test; Colson approved testing it as a stated hypothesis on October 1, 2026. No x_m value is stated here.
• The validation checks are complementary, not fully independent.
• The repair passed its test at one coupled point (β = 0.01, λ = 0.5) in three solver settings. That establishes the repair at that point and that its logging changes nothing, not precision certification, convergence or behavior elsewhere in the prior box, including negative β.
• Whether the revised precision design’s largest level (10000) fits the cosmology project’s physical memory is not yet shown. A read-only reading on October 6, 2026 fell 71 MiB short of the unchanged launch requirement; it was not a launch attempt. No saved reading passes the conservative launch worksheet, which charges every possible input cold, and memory the counters assign to no category has swung by up to 285 MiB between readings.