observer_dynamics

Observer dynamics: executable evidence interface

These tools verify bounded observer-like self-reading patches: local port loads, repair attempts, boundary readback, accumulated records and public receipt bundles. The immutable producers, specifications and compact numerical inputs live in ../../evidence/observer_dynamics_20260925/.

From the RER root:

python3 code/observer_dynamics/verify.py
python3 code/observer_dynamics/verify.py --replay-refinement --write-receipt
python3 evidence/exact_federation_tower_20260924/verify_tower.py
python3 evidence/source_net_causal_poset/verify_sampled_manifold.py
python3 evidence/observer_dynamics_20260925/lean/verify.py

The first command checks immutable file hashes, recomputes finite-chain and finite-window receipts, checks the 48 terminal-state digest crosspins, verifies the complete scale/angular population, estimator identities, null normalization and joint schedule-moment consistency, and runs independent and mutation tests. The second also reconstructs the fresh-noise streams for all 61 retained refinement chains and writes a separate canonical verification receipt. Neither command modifies historical evidence. The tower verifier streams load reconstruction in 8 MiB chunks. The manifold command uses the original independently written verifier and tests, recovered with hashes matching the sampled-run receipt.

archive_adapter.py answers only the historical producers' explicit git show requests for the frozen simulator revision from the archived source blobs. No sibling simulator checkout, network service or simulator Git history is needed. It also serializes relative provenance paths with / on every operating system; the frozen producer bytes and numerical comparisons remain unchanged. Install the repository-root requirements.txt for the complete default controls, including the pinned CAMB interpolation checks. Skipped controls fail verification. The causal controls use a separate process because they vendor another source revision. The tests can also be run through this entrypoint without introducing those modules into an unrelated Python process.

The large terminal arrays and geometry caches remain in their original archive. Digest crosspins and retained statistical identities do not constitute a fresh microscopic replay of those arrays. Large transport spectra and boosted-clock measurements are preserved with their original producers and checks; the bounded verification reruns their small controls. The sampled manifold record lacks the original large event samples and joint estimator covariance, so its large-run uncertainty cannot be independently rebuilt here.

The conditional CMB comparison uses the retained standard-transfer calculation and supplied source amplitude, tilt, cosmological background and temperature. It is a reproduction of a conditional observer distribution and measurement comparison, not a source-dynamics derivation of these inputs. Plot images are not required by this numerical package. Exact finite-chain identities, numerical receipts, retained large-run observations and compiled Lean witnesses remain distinct evidence types.

Retained scale and angular consistency

The scale receipt contains 51 scale rows, each with 16 schedules, and 240 angular rows. The checker binds their census to the archived specification; binds carrier count, port count, conserved mass and geometry digest to each original settlement receipt; and checks their derived statistics independently of the archived scale producer. It does not import that producer to calculate the expected answers. The new controls run through verify.py in CI.

Two previously accepted corruptions explain why the extra checks are needed. Doubling the L8 ancestor_1 equilibrium denominator while halving all 16 normalized variances changed their mean from 1.32427 to 0.66214. Every old ratio identity still passed, reversing enhancement versus suppression. Setting schedule noise to zero and common power to total power also passed, although the retained unequal responses alone require noise at least 7.02851e-6. The unchanged archived data pass the independent checks; these are defects in the scientific cross-check, not discovered errors in the frozen observations. The outer archive manifest already rejects changed frozen bytes.

For P ports, R raised ports and disjoint groups with c_i cells, let a_i be their normalized estimator weights. Exchangeable fixed-total binary loads and iid initial loads give, respectively,

C = sum_i (a_i-a_i^2)/(12 c_i),
V_exchangeable = R(P-R)/(P(P-1)) C,    V_iid = (35/12) C.

For G equally weighted groups, C=(G-1)/(12 G h) with harmonic cell mean h. For cell-weighted groups, replace h with arithmetic mean m. These identities hold even when the groups cover only part of the ports. Ancestral group sizes follow from the specification and parent population. Cube checks use the retained cell-count summaries and compare the weighted/unweighted pair. They also enforce necessary integer-count bounds: the arithmetic mean must admit an integer total with the declared minimum and maximum attained; convexity of 1/c then bounds the reciprocal sum and hence the harmonic mean. Angular pixel extrema must likewise fit the total carrier population. These checks do not reconstruct the absent geometry arrays or prove that every interior summary corresponds to an integer histogram.

The schedule check uses a further connection: response, power and noise must belong to one vector population. For the shared initial field x, set a_i=<x,y_i>/||x||^2, p_i=||y_i||^2/||x||^2 and let n be the unbiased schedule noise in those same units. Writing y_i/||x||=a_i e_0+v_i gives

r_i = ||v_i||^2 = p_i-a_i^2,
R_sum = ||sum_i v_i||^2
      = S sum_i p_i - S(S-1)n - (sum_i a_i)^2,
max(0, 2 max_i sqrt(r_i) - sum_i sqrt(r_i))^2
    <= R_sum <= (sum_i sqrt(r_i))^2.

The polygon bounds are necessary and sufficient for some real vector population when the residual space has dimension at least two: vectors of specified lengths in a plane attain every resultant length between those bounds. Centered scalar groups have residual dimension G-2, and real degree-ell angular coefficients have residual dimension 2 ell; all retained rows meet this condition. This does not prove realizability by bounded integer loads, reproduce an unseen covariance matrix, or authenticate the raw arrays. It does reject impossible moments that separate row identities cannot detect.

The executable check is a numerical consistency diagnostic. It assembles the supplied binary scalars as exact fractions and uses relative tolerance 1e-9 on the relevant moment scale, without an absolute small-power floor. It propagates residual subtraction uncertainty before taking square roots, with outward rational root bounds. A tiny negative computed noise within roundoff is allowed; reported statistics are never rewritten. This preserves valid nearly collinear and identical schedules. For the response spread, the exact sample variance V and variance about a rounded mean m_hat differ by S/(S-1) (m_hat-mu)^2. The checker bounds the mean's binary64 summation and division error, then permits that quadratic centering increment. It retains a relative lower bound on V and binds the standard error to the reported standard deviation; an erased or inflated resolvable spread still fails. For example, 16 original vectors with responses [1+ulp(1),1,...,1] have a NumPy variance about 6.7% above the exactly centered variance, although their absolute discrepancy is only rounding of the common center. Controls exercise both consumers at small, ordinary and large scales. Negative unbiased common power and noise fractions greater than one are also possible and remain accepted.

The repair changes verifier acceptance and its live verification summary. Frozen inputs, producer bytes, numerical observations, claim registrations, papers and physical classifications are unchanged. It supports the standing evidence audit #1033 without supplying a microscopic law, physical calibration or CMB prediction.