What can a shell reveal?

A different observer changes the kernel.

The regular 600-cell boundary is a closed 3-sphere with chain dimensions (120, 720, 1200, 600), total 2640. A selected vertex link is an icosahedral 2-sphere with 62 simplex coordinates. It is the boundary of a vertex star, not the manifold boundary of the whole 3-sphere.

Fix the real, unit-metric, unweighted skew Hodge–Dirac generator. Ideal history uses every derivative of the selected signed simplex amplitudes. Its unobservable space is the intersection of the kernels of R A^k; a snapshot kernel alone does not give the dynamical answer.

The second configuration removes 605 of the original 747 hidden dimensions. The third exposes the remaining 142. These are different observers: the original 62-channel system still cannot see dynamics confined to its invariant 747-dimensional kernel. Bounds refer to normalized finite sample maps in model time, not calibrated physical instruments.

Sector ranks do not add across this readout.

Each complex C₁₀ character sector has dimension 264. The single-link readout has isolated-sector snapshot rank 51. Its history ranks are:

The chosen readout does not intertwine the C₁₀ action. Summing these isolated-sector ranks therefore does not give the full rank 1893.

Coefficient fields matter. The characteristic-13 calculation has observation rank 1877 and a 763-dimensional kernel, separately from the real result. The irreducible cubic x³−2 and quartic Φ₁₀ define a degree-12 compositum; this exact field arithmetic does not transfer real singular-value conditioning or set a Pachner recurrence period.

Finite accuracy needs a declared error contract.

The finite six-ratio protocol uses three known preparations, 12 scalar readings at h=1/32 and 20 projected inversion iterations. Under its stated bounded error assumptions, the tracking bound is below 6.189×10⁻¹³, hence below 2⁻⁴⁰ and the requested 2⁻³⁹ gate. This is a conditional mathematical certificate, not empirical calibration.

The linked packages separately address gain/rate/clock ambiguity, grade metrics, fitted-coordinate transfer, held-out discrepancy, local fault detection and the precision versus measurement-count frontier. The optimized restricted diagnostic protocol uses 493,290 readings at response error 2⁻⁵⁸; a stricter variant uses 394,632 at 2⁻⁶⁴. Their hypotheses and intermediate error targets differ.

The proposed random-matrix self-prediction script and Euler stepping are not held-out model validation. The 2⁻⁴⁹ residual gate is unsupported and rejected. The 2⁻³⁸ shape gate retains its typed calibration meaning; it is not a Borel-shear allowance.

Extract the spectrum without the 287-GiB array.

A single freely evolving trajectory has at most 105 temporal directions in this model. Adding Hankel columns alone cannot excite 1893 independent observable directions. The explicit 2640×37 preparation matrix supplies the required modal excitation; each preparation has squared norm 165/169.

The frozen primary record uses M=4096, 8191 temporal vectors per preparation and 62 history channels: 18,790,154 scalar readings. Its normalized Hankel matrix has shape 253952×151552; binary64 dense storage would use 286.75 GiB. Exact Hankel actions, energy accounting, verified 1893×1893 factor Grams and projection residual bounds avoid that allocation.

The nominal benchmark uses known modal factors. The later record-only proposer learns 105 temporal components from the record and contract, while the verifier checks all readings against the proposed model. The timing and worker-memory figures describe those frozen runs, not a universal performance guarantee. The 62 history channels remain distinct from the 189 diagnostic outputs and additional calibration ports.

Bounded noisy examples pass a narrower test.

All four certify 1893 singular values above the unchanged threshold. The temporal sketch theorem separates 105 components throughout the declared per-vector 2⁻²⁰ additive error ball. It does not establish universal frequency-fitting or optimizer convergence, or full calibration-family robustness. The single-vector fixture meets the vector ceiling but exceeds the sufficient scalar ceiling; no full acquisition-contract membership follows.

Threshold count is not exact algebraic rank. A certified weak impulse record has rank at least 2203 while retaining only 1893 singular values above threshold. The codec recovers every byte of the finite recorded history by retaining residuals. It does not reconstruct hidden bulk state or select a unique continuous-time generator through sampling aliases.

The parallel-work corrections remain visible.

• A coframe or a chosen right inverse selects a lift; it cannot recover an arbitrary discarded kernel without additional information.
• Boundary nilpotency does not impose coframe torsion flatness. A Pachner move needs typed incidence/coframe compatibility, not an automatic flatness-violation claim for an undefined edge value.
• The supplied Q₈ matrices over F₁₃ can be verified exactly. The claimed embedding 2I into SL(2,F₁₃) cannot hold: 120 does not divide 2184. No unsupported C₁₃ or dihedral branching rule is promoted.
• Twenty incident tetrahedra do not define a self-similar fractal or a Hausdorff dimension above its embedding dimension.
• Finite fields, geometric frames, norm conservation, clocks and physical gauge groups require separate definitions and correspondence evidence.

The source audits preserve valid finite arithmetic and counterexamples while separating them from unsupported gauge, curvature and leakage interpretations.

Review is a source-bound return.

A corrected return must identify the exact archive or commit, provide matching hashes, declare reviewer identity and independence, distinguish static inspection from fresh execution, and state scoped pass/fail/open findings. Constructor/verifier software separation alone does not establish independent human review.

All work remains in the development lane. Neither a task marked Done, a merged repository PR nor this public update enacts scientific or physical promotion.

Inspect every part of the session.

The earlier 17 archives were recovered after the transient checkout was lost. ZIP checks and all 307 original manifest entries passed during repository publication. Those scientific receipts are inherited; they were not all replayed during recovery. The nine later captures retain their own exact/numerical verification scopes.

Public reports are copied byte-for-byte and hash-bound. Code and datasets remain in the linked immutable repository evidence. The unavailable Genesis PNG and earlier excluded private source text are not republished. This is an accessible artifact record, not an export of hidden model state or a complete chat transcript.