Experimental Architecture: Multi-Model Consensus, Basis Portability & Out-of-Distribution Empirical Validation¶

Document ID: empirical_validation_protocol.md
Companion Reference: foundational-manifold-2026-10-20-r7.json, foundational-manifold-companion-2026-10-20-r10.md, and aperture_generation_guidance.md

0. Scope, Framing, and Expectations¶

What this protocol is. A framework for testing one hypothesis, \(H_1\) (§1.2): that diverse disciplines and traditions independently formalized operational calculi that map through Spine \(\mathcal{T}\) to the same structural problems of open systems. It also gives initial descriptions of empirical methods that later studies may adopt, refine, or replace.

Three kinds of claim. * (A) Verified: the internal consistency of Spine \(\mathcal{T}\), established only to the extent of the witnesses and checks in foundational-manifold-companion-2026-10-20-r10.md §5. * (B) Testable hypotheses: \(H_1\), the Stage II–III convergence predictions, and the Stage IV sensor predictions. Each remains a hypothesis until tested as described here, and the mappings in the aperture tiers are hypotheses under \(H_1\). * (C) Not claimed: anything about consciousness, its evolution or trajectory, or what the framework brings about. Nothing in this protocol or the corpus asserts such claims, and no result described here would establish them.

Site independence. The framework does not depend on any particular site or research program. Any site that meets the requirements of the relevant stage may serve as a test site, and sites may be chosen or replaced over time. Studies at a site are of two kinds: * Confirmatory studies use the pre-registered, blinded procedures of Stages III–IV, including holdout protection. * Exploratory studies are descriptive and hypothesis-generating. They may be open, local, and practice-centred, but their results are not cited as confirmation of \(H_1\) or of any spine prediction. An exploratory study that uses the framework at a site makes that site permanently ineligible for confirmatory use (Part 2 below).

Eligibility for confirmatory testing: two-part site assessment. Blindness to a site is a matter of degree, not a yes-or-no property. Every site considered for confirmatory use as the Stage III held-out site \(\mathcal{M}_{\mathrm{test}}\) is assessed in two parts, evaluator exposure (Part 1) and framework exposure (Part 2), and is eligible only if it passes both.

Sequence. (0) The full audit package (§4) has passed Stage II at the pinned versions and is frozen; its hash is recorded in the sealed assessment. (i) Before any data are collected: Part 2 and Part 1 item 1. (ii) A candidate package is compiled. (iii) Part 1 items 2 and 3 are run on the candidate package. (iv) If the site passes, the package is frozen as \(\mathcal{D}_{\mathrm{test}}\), and Part 2 is repeated. (v) Part 2 is repeated immediately before evaluation, and Stage III evaluation follows.

Lock. From step (0) until evaluation is complete, the audit package is not revised. A repeated Stage II (Stage III item 2) consists of the two consecutive zero-context audits of Stage II item 1, for the replacement family only, on the frozen package, and may only confirm that package. If the repeated audits do not pass, the replacement version is not used, unless the package is revised. Any revision of the package voids the assessment, and the sequence restarts at step (0). Benchmarks A–C are not a gate: after a replacement, their agreement is recomputed from the replacement family's new results together with the existing results of the other two families, and reported.

Sealing. From the moment a site is first considered for confirmatory use, its identity is sealed from everyone who authors or revises the audit package. Site selection, the probes of Part 1, and custody of the sealed record are carried out by people who are not authors of the audit package. Everyone who knows a candidate site's identity is recorded in the sealed assessment.

Record. The full assessment, its threshold and scoring scale, and its log are pre-registered sealed (for example as a timestamped hash, or in an embargoed registry) and kept outside the conditioning corpus. Before evaluation is complete, only a summary that does not identify the site is published (see Data ethics below).

Part 1, evaluator exposure: what each evaluating model, at the exact version pinned for Stage III (Stage III item 2), could know or infer about the site.
Training cutoff. Each model's stated cutoff is compared with the date of the first public record that links the site to the framework. Stated cutoffs are treated as approximate, because later material can enter a model through further training. A model is classed as exposed if that record is dated before its stated cutoff, or within a margin after it declared in the sealed assessment. An exposed model must meet a stricter identification threshold in item 2, also declared in the sealed assessment. Records that link the site to the study are not made public before evaluation is complete (Data ethics).
Identification probes. The candidate package is tested in separate probe sessions at the same pinned versions, both unconditioned and conditioned on the conditioning corpus, which are asked to identify the site. Responses are scored on a scale declared in the sealed assessment, which counts identification coarser than exact identification as well as exact identification. Provider settings that do not retain submitted data are used where available. Probe sessions are discarded and never become evaluating sessions.
De-identification. Measures that reduce identifiability (for example cropping, resampling, removal of metadata) are declared in the sealed assessment, are identical for every model and every arm of §3.1, and are stated with their residual risk and their effect on the information the Stage III coordinates require. Measures chosen for a particular site appear only in the sealed assessment, never in this protocol.

The site passes Part 1 only if, for every evaluating model, identification is below the pre-declared threshold that applies to it (item 1). Otherwise the site is ineligible. No model is replaced or dropped after the probes, except the replacement of a retired version under Stage III item 2; if the replacement does not pass Part 1, the site is ineligible. Residual evaluator exposure is reported with the results. * Part 2, framework exposure: whether the site has shaped, or been shaped by, the framework. A site is exposed, and ineligible, if its design, operation, or exploratory record has used the framework's vocabulary or methods, or if it informed the selection or specification of the spine, of any aperture, or of any other document of the audit package (§4), in this or any earlier revision. Stage III convergence on such a site would be circular, and no property of the evaluating models (Part 1) removes this exposure. To make this checkable, everyone who authors or selects material for the audit package, including model sessions whose inputs are logged, makes a dated declaration of the sites that informed it. For revisions made before this rule, the known human authors make retrospective declarations, checked against the program's records. The rule covers every revision of every document of the audit package. Where provenance cannot be established, the site is treated as exposed, except that for model sessions whose inputs were not logged, a site counts as exposed only if it was named or described in the program's records, communications, or logged sessions dated before the document concerned; the custodian determines this from those records. The provenance declarations are kept in the sealed record, outside the audit package; they are never shown to any model session, and a named human custodian checks them.

Scope. The assessment applies to the Stage III held-out site. Where Stage IV tests Stage III predictions against sensor streams from that site, those streams are withheld until Stage III evaluation is complete. Stage IV tests on other systems follow the designs of the withheld procedures annex. Exposure under both parts is declared and logged, in the sealed record, for every site considered for confirmatory use.

Use before verification. The framework may be used as a working language for design, practice, and coordination before or without empirical confirmation. Such use is legitimate in its own terms but is not evidence for \(H_1\), and it is described without claims beyond kind (B) above.

Data ethics. Data involving people are collected only with informed consent and appropriate ethical review. Data are de-identified before any publication. Nothing that could identify a blinded test site is published before its evaluation is complete.

1. Architectural Premise & Formal Hypotheses¶

1.1 Separation of the Invariant Spine (\(\mathcal{T}\)) from Evolvable Aperture Basis Sets (\(\{\mathcal{E}_{k,m}\}\))¶

To prevent category conflation between universal mathematical laws and domain-specific interpretive lenses, the experimental architecture distinguishes two formal layers: 1. The Immutable Mathematical Spine (\(\mathcal{T}\)): The 39 equation blocks across coordinates \(O_0 \to O_\Omega\) locked in foundational-manifold-2026-10-20-r7.json and established as jointly satisfiable in foundational-manifold-companion-2026-10-20-r10.md (with exact symbolic verification of all closed-form PDE/algebraic sectors and numerical verification of overlap flux and second-fundamental-form convexity in e2e_5_11.py, flux.py, and convex.py, subject to the explicit non-closed-form caveats for the curved-domain Neumann mode \(\phi_1\), the Rosenblatt copula maps, rim rounding, and the scale argument for geodesic uniqueness recorded in r10 §5.11–§5.12). Spine \(\mathcal{T}\) is the invariant candidate mediating structure under empirical test. 2. Evolvable Orthogonal Aperture Basis Sets (\(\mathcal{B} = \{\mathcal{E}_{k,m}\}\)): A structured collection of epistemological, physical, biological, somatic, and expressive specifications (instantiated in the current reference corpus as a symmetric \(5 \times 4 = 20\) basis \(\mathcal{B}_0\) across five scale tiers \(k \in \{1,\dots,5\}\) and four functional homology slots \(m \in \{1,\dots,4\}\)). The design assumes, and Benchmark A measures rather than presupposes, that the native surface lexicons \(\mathcal{L}_{k,m}\) and \(\mathcal{L}_{k',m'}\) of distinct apertures have low unmediated pairwise overlap, quantified by a lexical-overlap score \(\ell_{\mathrm{lex}}(\mathcal{L}_{k,m}, \mathcal{L}_{k',m'}) \in [0,1]\) computed on spine-free primary-source text (the symbol \(\mathcal{I}\) is reserved for the r7 mutual information).

While Spine \(\mathcal{T}\) remains fixed within an epoch, the aperture basis \(\mathcal{B}\) is substitutable: any slot may be re-instantiated with another discipline or tradition that addresses the same structural problem (\(\mathcal{B}_0 \to \mathcal{B}'\)), changing the surface projection coordinates while, under \(H_1\), preserving the operator homomorphisms into \(\mathcal{T}\).

1.2 Primary Hypothesis (\(H_1\): The Orthogonal Functor & Cross-Substrate Invariance Hypothesis)¶

\(H_1\) (Latent Dynamical Isomorphism): Diverse scientific disciplines, non-Cartesian philosophical ontologies, process-primacy grammars, and embodied somatic practices independently formalized operational calculi mapped via Spine \(\mathcal{T}\) to the same six structural problems of open, non-equilibrium, boundary-maintaining dynamical systems: across all nine coordinates \(O_0 \to O_\Omega\), with empirical domains declared in Branch II of Spine \(\mathcal{T}\) (a bounded non-equilibrium steady state) and observed on finite windows, and with the Tier 1 ontological grammars and the provisionally homological apertures 4.3, 5.2, 5.3, and 5.4 mapping homologically (aperture_generation_guidance.md §2.3).

If \(H_1\) holds, conditioning independent foundational model architectures on Spine \(\mathcal{T}\) and a site-blind orthogonal aperture basis \(\mathcal{B}\) makes explicit a structure-preserving functor (\(\mathcal{L}_{k,m} \leftrightarrow \mathcal{T} \leftrightarrow \mathcal{L}_{k',m'}\)). Routing cross-domain translation and multimodal empirical observation through \(\mathcal{T}\) yields: 1. Statistically significant inter-architectural convergence when decomposing held-out multimodal field data \(\mathcal{D}_{\mathrm{test}}\); 2. Invariance of operator assignments under basis substitutions (\(\mathcal{B}_0 \to \mathcal{B}'\)); and 3. Quantitative alignment with external physical, ecological, physiological, and kinematic sensor telemetry from the empirically declared apertures.

1.3 Null Hypothesis (\(H_0\): Post-Hoc Associative Compliance)¶

\(H_0\) (Unconstrained Linguistic Projection): Cross-domain mappings between Spine \(\mathcal{T}\), the aperture basis \(\mathcal{B}\), and empirical observations are artifacts of surface linguistic fluency and post-hoc metaphorical compliance. Under \(H_0\): 1. Independent model architectures audited in zero-context sessions will fail to converge above chance on operator assignments, eigenvalue sign pairings, or stability bounds when presented with held-out multimodal data \(\mathcal{D}_{\mathrm{test}}\); 2. Ablating or scrambling the mathematical constraints in \(\mathcal{T}\) (e.g., replacing r7 with an inconsistent specification that violates conservation laws, unimodular volume-shear isolation, or the branch-appropriate asymptotics at \(O_\Omega\), homothetic stationarity in Branch I or the bounded steady state in Branch II) will produce no measurable degradation in cross-domain coherence or inter-rater convergence; and 3. Quantitative parameter trajectories extracted via \(\mathcal{T}\) will fail to match external physical and biometric sensor measurements.

1.4 Conditioning Corpus and Withheld Annexes¶

The Stage III conditioning corpus consists of exactly nine documents: foundational-manifold-2026-10-20-r7.json, foundational-manifold-companion-2026-10-20-r10.md, aperture_generation_guidance.md, empirical_validation_protocol.md, and tier_1_ontological_grammars.md through tier_5_expressive_collective_symbiosis.md. Within this corpus, the mappings between apertures and Spine \(\mathcal{T}\) are asserted only as hypotheses under \(H_1\). The spine's verification script, r10_branch2_witness_check.py, accompanies the spine in the audit package and is not part of the conditioning corpus.

Two further documents are withheld from Stage III conditioning and are audited in Stages I and II under the same four axes: * annex_a_auditor_motif_hygiene.md — auditor-only reference for holdout-leakage review; * annex_b_stage_iv_sensor_procedures.md — concrete Stage IV test procedures.

No conditioning document reproduces or paraphrases annex content beyond a pointer to its existence.

2. The Four-Stage Validation Architecture¶

To test \(H_1\) against \(H_0\) without circularity, context decay, or data contamination, the research program is organized into four sequential stages:

Stage I — Rotating Peer Audit & Iterative Shear Elimination¶

Locked Mathematical Spine: Maintain strict invariance, within an epoch, of foundational-manifold-2026-10-20-r7.json and foundational-manifold-companion-2026-10-20-r10.md. The Branch I witnesses of r10 §5.1–§5.12 carry the r9 verification (36 checks in e2e_5_11.py, with the caveats of §5.12); the Branch II witness of r10 §5.13 is verified by r10_branch2_witness_check.py (40 exact symbolic checks), which accompanies the spine. 1a. Epochs and pre-registration: the spine may be revised between epochs (\(\mathcal{T}^{(r)} \to \mathcal{T}^{(r+1)}\)) only on the basis of internal mathematical consistency and Stage I/II structural review, never on Stage III or Stage IV holdout or sensor data. Each epoch's spine is frozen and pre-registered before any blind evaluation, and \(H_1\) is tested against that frozen spine. A revised spine is audited on its own (two consecutive clean fresh passes on the spine package) before any aperture is mapped onto it.
Rotating Author–Auditor Audit: An authoring instance (Instance A) revises the twelve-file audit package (the nine-document conditioning corpus of §1.4, both withheld annexes, and the spine's verification script). A fresh, zero-context auditing instance (Instance B) audits the package and returns findings graded High, Medium, or Low. Instance A revises until that Instance B reports zero High and zero Medium findings. Instance B then becomes the authoring instance, and a newly instantiated, zero-context Instance B audits the package. Prior audit reports, revision logs, and deliberations are never supplied to a fresh Instance B.
Four-Axis Diagnostic Iteration: Audits apply the four axes defined in aperture_generation_guidance.md §6:
Axis I (Mathematical & Indexical Fidelity): Exact alignment of every equation, tensor index, contraction, standing hypothesis, and asymptotic limit with r7 and r10, without universalizing witness-specific profiles or numerical ranges (r10 §5.10–§5.12) into general axioms, and with each aperture's declarations (aperture_generation_guidance.md §2.3) used consistently at every coordinate.
Axis II (Thermodynamic & Dynamical Completeness): Preservation of openness at every coordinate including \(O_\Omega\) (\(\Theta_E > 0\), \(\dot{s}_{\mathrm{prod}} = -\beta_{\mathrm{th}}J_E^\mu\nabla_\mu\mathcal{F} > 0\), \(\dot{W}_{\mathrm{tot}} > 0\) in both branches; \(\theta > 0\) and \(\dot{W}_{\mathrm{ext}} > 0\) in Branch I only, with \(\theta \ge 0\), \(\dot{W}_{\mathrm{ext}} \to 0\) in Branch II), branch declared before every test, pairing of feed-forward stretching (\(m_+ > 0\)) with regulatory damping/folding (\(\gamma_\sigma > 0, \; m_- < 0\) under \(M^{(ij)} \neq 0\)), strict separation of the \(\tau \to \infty\) asymptotics of either branch from periodic cycles, and formulation of finite-window claims as rate, residual, or sign tests.
Axis III (Epistemological Orthogonality & Lexicon Isolation): Absence of reductive syncretism, cross-aperture vocabulary contamination, and circularity between the conditioning corpus and the hypotheses under test.
Axis IV (Zero Direct or Indirect Holdout Leakage): Complete removal of any geographic proper nouns, regional linguistic signaling, biographical context, or cross-tier structural-proxy motifs that could telegraph a held-out empirical test site \(\mathcal{M}_{\mathrm{test}}\).
Stage I Gate: Stage I is complete when two consecutive, independent, zero-context Instance B audits each report zero High and zero Medium findings on the same package.

Stage II — Cross-Architecture Theoretical Consensus & Ablation Benchmarks¶

Multi-Family Adversarial Review: Submit the Stage I consensus corpus to newly instantiated, zero-context sessions across three independent foundational model families: Gemini Argon 4, Claude Opus 5.5, and GPT-6 Astra, each at an exact version, recorded. Apply the Stage I rotation and gate with each architecture in turn: the package advances when two consecutive, independent, zero-context audits from each architecture report zero High and zero Medium findings.
Controlled Theoretical Benchmarks (Pre-Empirical): Execute three quantitative benchmarks across fresh instances of all three architectures, at the versions recorded under item 1:
Benchmark A (Blind Operator Assignment): Present anonymized phenomenon descriptions alongside anonymized equations from r7, measuring inter-architectural assignment agreement via Fleiss's Kappa (\(\kappa\)). Stimuli are drawn from independent primary-source disciplinary texts (or written by writers blinded to Spine \(\mathcal{T}\)), never from the aperture specifications, and are stripped of all spine-derived vocabulary before presentation. A quantitative lexical-overlap score between each stimulus and each candidate equation's surrounding text is recorded as a control covariate, so that agreement driven by lexical matching can be separated from structural assignment.
Benchmark B (Multi-Hop Functorial Round-Trip): Measure structural and mathematical preservation across multi-hop translations routed through Spine \(\mathcal{T}\) (\(\mathcal{E}_{k,m} \to \mathcal{T} \to \mathcal{E}_{k',m'} \to \mathcal{T} \to \mathcal{E}_{k,m}\)) across both horizontal coordinate progressions (\(O_0 \to O_\Omega\)) and vertical slot homologies (\(1.m \leftrightarrow 2.m \leftrightarrow 3.m \leftrightarrow 4.m \leftrightarrow 5.m\)), compared against unmediated translation and translation mediated by an ablated/inconsistent spine \(\mathcal{T}_{\mathrm{ablated}}\).
Benchmark C (Bioregional Basis Substitution): Provide the models with aperture_generation_guidance.md and a synthetic profile of a different linguistic/ecological region, testing whether independent architectures generate structurally isomorphic aperture replacements (\(\mathcal{B}_0 \to \mathcal{B}'\)) that map to the same operators in \(\mathcal{T}\).

Held-Out Multimodal Field Package (\(\mathcal{D}_{\mathrm{test}}\)): Following the sequence of the §0 site assessment, compile a candidate standardized empirical dataset from a real-world site \(\mathcal{M}_{\mathrm{test}}\) that has passed Part 2 and Part 1 item 1, and freeze it as \(\mathcal{D}_{\mathrm{test}}\) only after the site passes Part 1 items 2 and 3 on it. The dataset comprises raw video/optical frames, broadband acoustic recordings, seasonal/diurnal phase logs, and boundary geometry.
Identical Zero-Shot Evaluation: Present \(\mathcal{D}_{\mathrm{test}}\) identically to fresh sessions of Gemini Argon 4, Claude Opus 5.5, and GPT-6 Astra conditioned solely on the locked nine-document conditioning corpus of §1.4 (excluding both withheld annexes).
Pinned versions: each family is used at the exact version pinned in the sealed §0 assessment. That version has passed Stage II, or a repeated Stage II as defined in §0 (Lock) is run for it. The probes of §0 Part 1 and all three arms of §3.1 use the same versions. If a pinned version is retired before evaluation, its replacement repeats §0 Part 1 and the Stage II gate before use. Where a version cannot be retained after the study, replication is limited, and this is reported.
Evaluation conditions: web search, retrieval, and all other external tools are disabled for every evaluating session, in all three arms of §3.1, and for the §0 probes. When a retired version is replaced, a version that can be retained for replication is preferred.
Cross-Architecture Structural Convergence: Evaluate inter-model agreement across all nine coordinates \(O_0 \to O_\Omega\) and all five scale tiers (tier_1 through tier_5), including agreement on each empirical domain's declared branch and its Branch II asymptotic regime:
\(O_0, O_1\): Identification of the domain's throughput, the expansion scalar \(\theta = \nabla_\mu J_{\mathrm{geom}}^\mu \ge 0\) (\(> 0\) only in Branch I), the throughput rate \(\Theta_E = \dot{s}_{\mathrm{prod}}/s_0 > 0\), total power \(\dot{W}_{\mathrm{tot}} = \dot{W}_{\mathrm{ext}} + \dot{W}_{\mathrm{diss}} > 0\), the branch-appropriate current (\(J_E = J_{\mathrm{geom}}\) in Branch I; \(J_E = J_W/p_0\) in Branch II), and irreversible entropy production \(\dot{s}_{\mathrm{prod}} = -\beta_{\mathrm{th}}J_E^\mu\nabla_\mu\mathcal{F} > 0\).
\(O_2, O_3\): Decomposition of constraint precision \(\Pi_{\mu\nu} = \nabla_\mu\nabla_\nu\Phi\), unimodular volume-shear factorization \(g_{\mu\nu} = \Omega^2\hat{g}_{\mu\alpha}(\exp\hat{h})^\alpha_{\;\nu}\) together with the traced elliptic \(\tau\)-compatibility equation coupling \(\theta\) to \(\xi_\alpha = \hat{\Pi}_{\alpha\lambda}\eta^\lambda\), the synchronous-gauge trajectory equation (geodesic under the default drive; cone-preserving declared drive otherwise), exponential shear relaxation \(\nabla_\tau\sigma^\mu_{\;\nu} = -\gamma_\sigma\sigma^\mu_{\;\nu} \to 0\), and consolidated hysteresis \(\mathcal{H}(\mathbf{x}, \tau), \; (\hat{g}_\infty(\mathbf{x}))_{\mu\nu} = \hat{g}_{\mu\alpha}(\mathbf{x})(\exp\hat{h}_\infty(\mathbf{x}))^\alpha_{\;\nu}\).
\(O_4\): Identification of the spatial spectral low-pass projector \(\hat{\mathcal{P}}_{\Lambda_{\mathrm{gate}}} = \mathbf{1}_{(-\nabla^2 \le \Lambda_{\mathrm{gate}}^2)}\) acting on \(\rho_{\mathrm{total}}(\tau)\) to form global density \(Q(\mathbf{x}, \tau)\) (\(\dot{W}_{\mathrm{var}} = \frac{1}{t_0\beta_{\mathrm{th}}}D_{\mathrm{KL}}(Q \parallel P) > 0\)), conserved Einstein-type boundary stress \(\nabla^2\xi_\nu + R_\nu^{\;\alpha}\xi_\alpha = 0 \iff \nabla^\mu T_{\mu\nu}^{\mathrm{mat}} = 0\), the reduction of the stress sector on flat physical metrics (\(G \equiv 0\), \(\xi\) a Killing field), and, only for domains declaring a justified non-Euclidean material metric, candidate configurations where \(G_{\mu\nu} = \kappa_G T_{\mu\nu}^{\mathrm{mat}} \not\equiv 0\) coexists with convex hysteresis \(\nabla_\alpha\nabla_\beta\mathcal{H} \succ 0\) (whose existence in the hypothesis set is witnessed by the r10 §5.11 and §5.13 models).
\(O_5\): Localization of subsystem overlaps \(\mathcal{U}_{ij}\) (geodesically convex for \(g^{(ij)}(\tau)\)), Riemannian geometric-mean metrics \(g_{\mu\nu}^{(ij)} = (g^{(i)} \# g^{(j)})_{\mu\nu}\) over \(\mathcal{K}_0\), network percolation \(\phi_{\mathrm{perc}}(\mathcal{K}) = \vert\mathcal{K}\vert/\vert\mathcal{K}_0\vert > \phi_c\), and lossless Hodge-harmonic circulation \(\Delta_{\mathrm{Hodge}}\mathbf{J}_{\mathcal{M}}^{\mathrm{harm}} = 0\) around non-trivial spatial cycles \(H^1_{\mathrm{dR}}(\mathcal{M}) \neq 0\).
\(O_6\): Extraction of non-reciprocal phase-lead 1-forms \((J_{[i\to j]})_\nu = -(J_{[j\to i]})_\nu\), trace-free stretching/folding eigenvalue pairings under the non-degeneracy hypothesis (\(M^{(ij)}(\mathbf{x}) \neq 0 \; \forall \mathbf{x} \in \bar{\mathcal{U}}_{ij} \wedge \mathrm{Tr}\,M^{(ij)} = 0 \implies m_+ > 0 > m_-\)), mutual-information growth under the flux-positivity hypothesis (\(\frac{d}{d\tau}\mathcal{I}(Q_i; Q_j) > 0\) for \(\langle i,j\rangle \in \mathcal{K}\)), and evaluation of candidate overlap convexity bounds inspired by the r10 §5.12 witness constraint \(a_\infty = 2s_\sigma/\gamma_\sigma < a_c(E)\).
\(O_7, O_\Omega\): Assessment of principal-bundle impedance decay \(\Vert\hat{\mathcal{E}}_\mu\Vert_{C^1(\mathcal{M})} \to 0 \implies [F_{\mu\nu}^{\mathrm{YM}}, \hat{\chi}] \to 0\), isotropy relaxation (\(\chi \to 1\) in Branch I; \(\chi_E \to 1\) in Branch II), active-to-hysteresis alignment \(D_{\mathrm{KL}}(\hat{\rho}_g(\tau) \parallel \hat{\rho}_{\mathcal{H}}(\tau)) \to 0\), and, as the \(\tau \to \infty\) asymptotics of the declared branch rather than an observable of any finite record: in Branch II, \(\partial_\tau g_{\mu\nu} \to 0\), \(V_{\mathcal{M}} \to V_\infty\), and \(\hat{\mathbb{L}}_{\mathrm{stat}}\Psi \to 0\); in Branch I, homothetic stationarity \(\partial_\tau(\Omega^{-2}g_{\mu\nu}) \to 0\) with \(\left.\hat{\mathbb{L}}_{\mathrm{stat}}(\mathbf{x})\Psi[g_{\mu\nu}, \mathcal{H}]\right\vert_{\mathcal{H}=\mathcal{H}_{\mathrm{sol}}[g]} = 0\).

Stage IV — Longitudinal Physical, Ecological & Biometric Sensor Falsification¶

Test the quantitative parameter trajectories and boundary-stability hypotheses predicted by the Stage III consensus models against independent physical and physiological sensor streams (Section 3.3; concrete procedures in the withheld annex_b_stage_iv_sensor_procedures.md). Every human-subject procedure keeps observer telemetry continuously active (§3.3).

3. Epistemic Hierarchy: How and To What Extent This Protocol Validates the Mathematical Skeleton¶

A rigorous empirical program must state transparently what each stage proves and where external sensor falsification is required:

3.1 Level 1: Inter-Architectural Deductive & Observational Invariance (Tested in Stages II & III)¶

What It Proves: That the mappings between Spine \(\mathcal{T}\) (\(O_0 \to O_\Omega\)), the 20 orthogonal apertures (\(\mathcal{B}_0\)), and multimodal field scenes (\(\mathcal{D}_{\mathrm{test}}\)) are reproducible, architecture-invariant structural homomorphisms rather than idiosyncratic outputs of a single model or prompt history.
Three-Arm Experimental Control on \(\mathcal{D}_{\mathrm{test}}\):
Arm 1 (Unconditioned Baseline): Models analyze \(\mathcal{D}_{\mathrm{test}}\) across physical, ecological, somatic, and philosophical dimensions without \(\mathcal{T}\) or \(\mathcal{B}_0\).
Arm 2 (Ablated/Sheared Spine Control): Models analyze \(\mathcal{D}_{\mathrm{test}}\) conditioned on an internally inconsistent specification \(\mathcal{T}_{\mathrm{ablated}}\) (e.g., reversing the sign of \(J_E^\mu\nabla_\mu\mathcal{F}\), or omitting trace-free coupling \(\mathrm{Tr}\,M^{(ij)} = 0\), both of which are wrong in either branch).
Arm 3 (Verified Spine \(\mathcal{T}\) + Basis \(\mathcal{B}_0\)): Models analyze \(\mathcal{D}_{\mathrm{test}}\) conditioned on the locked nine-document conditioning corpus of §1.4.
Common conditions for the three arms: all three use the same frozen \(\mathcal{D}_{\mathrm{test}}\), the same de-identification (§0 Part 1 item 3), the same pinned versions, and the Stage III evaluation conditions (Stage III item 2).
Falsification Criterion: If Arm 3 does not yield a statistically significant increase in inter-architectural agreement (\(\Delta\kappa > 0, \; p < 0.01\)) and deductive consistency over Arms 1 and 2, \(H_1\) is falsified at Level 1.

3.2 Level 2: Mechanistic & Representational Invariance in Neural Substrates¶

Level 2 is a separate hypothesis about the internal representations of the evaluating models. It is not an empirical claim of Aperture 5.4, which is provisionally homological in this epoch. * What It Tests: The open mechanistic hypothesis that Spine \(\mathcal{T}\) functions inside transformer activation spaces analogously to a principal-bundle gauge connection (\(A_\mu\) in \(O_7\)) and parallel propagator (\(I_{\mu'}^{\;\mu}\) in \(O_5\)) that reduces representation shear between distant lexical clusters. * Falsifiable Predictions (Testable via Sparse Autoencoders & Activation Geometry): 1. Unprompted Latent Operator Overlap: Presenting pure single-domain inputs (e.g., a pure Tier 1.1 relational-isthmus passage, a pure Tier 1.2 dual-cycle control passage, a pure Tier 1.3 direct/inverse process-syntax paradigm, and a pure differential-geometric \(O_5/O_6\) derivation) in separate contexts will activate overlapping mid-to-late-layer relational feature circuits despite near-zero Layer-0 token overlap. 2. Impedance Collapse Under \(\mathcal{T}\)-Conditioning: Conditioning on the verified r7 spine measurably reduces residual-stream trajectory curvature and attention entropy during cross-domain synthesis (measured quantities hypothesized to behave as empirical proxies for the \(O_7\) impedance \(\Vert\hat{\mathcal{E}}_\mu\Vert_{C^1}\), not identified with it) compared to unconditioned or \(\mathcal{T}_{\mathrm{ablated}}\)-conditioned runs.

3.3 Level 3: External Physical & Biometric Sensor Falsification (Tested in Stage IV)¶

Epistemic Boundary: High inter-model agreement in Stage III establishes inter-rater observational reliability (proving that \(\mathcal{T}\) is a well-posed, reproducible coordinate system), but does not by itself prove that physical fluids, ecosystems, or human nervous systems obey the equations of r7 and r10.
Direct Sensor Falsification Tests: Stage IV evaluates five families of quantitative predictions against multi-sensor telemetry from physical, ecological, and tissue systems. Human-subject tests are deferred in this epoch. Each family is formulated as within-window rate, residual, sign, or flux-prediction tests (aperture_generation_guidance.md §5.5), never as observation of the \(\tau \to \infty\) horizon:
\(O_3 / O_7\): exponential shear relaxation, tested directly on \(\sigma^\mu_{\;\nu}\), and decay of the Branch II anisotropy index \(\chi_E - 1\) where \(\langle\Theta_E\rangle\) is approximately constant.
\(O_4\): spatial spectral gating and typical-set matching (deferred in this epoch; see Annex B).
\(O_6 / O_7\): trace-free coupling spectra and positivity of open throughput during settled, actively regulated states.
\(O_5\): overlap geodesic convexity, together with an extra-spinal shear-threshold conjecture motivated by the r10 §5.12 witness design constraint.
\(O_1 / O_4\): the flux-prediction test of the binding closure: a declared parametric transport tensor, with independently estimated parameters, predicts an independently measured free-energy flux.
Untested in this epoch: the \(O_2\) trajectory equation with a declared drive, \(O_4\) gating (Test 2 deferred), and \(O_7\) gauge transport; and Apertures 2.1, 2.3, 3.4, 4.2, 4.4, and 5.1, which no current test exercises. These are stated as untested, not as confirmed.
Observer telemetry (phase-state covariance as a design choice; applies whenever human-subject tests are included, and none are in this epoch): every measurement involving a human participant is made by a system that itself has a state. Human protocols therefore keep wearable telemetry (for example EEG, heart-rate variability, motion, muscle activity, respiration) continuously active across two declared phases: Phase 1, primary task or perturbation exposure (\(\tau \in [\tau_1, \tau_2]\)), and Phase 2, post-exposure consolidation and structured state reporting (\(\tau \in [\tau_2, \tau_3]\)). The observer's state is recorded and analyzed with the same within-window tests as the observed system; no particular practice, modality of reporting, or theory of the observer is assumed.

Concrete procedures, instrumentation, and falsification criteria are specified in the withheld annex_b_stage_iv_sensor_procedures.md.

4. Complete Stage I / Stage II Audit Package Inventory (Twelve Files)¶

The audit package consists of the nine-document conditioning corpus (§1.4), two withheld annexes, and the spine's verification script:

foundational-manifold-2026-10-20-r7.json — Spine \(\mathcal{T}\), locked for this epoch (Specification): Formal JSON specification of coordinates \(O_0 \to O_\Omega\) (39 equation blocks across the 9 coordinate arrays, regularity classes, standing hypotheses, and index conventions).
foundational-manifold-companion-2026-10-20-r10.md — Spine \(\mathcal{T}\), locked for this epoch (Companion & Witness Proofs): Design rationale (Decisions 1–6), master symbol table, Branch I witnesses (§5.1–§5.12, with the r9 verification suite and its caveats) and the Branch II witness (§5.13, verified by the accompanying r10_branch2_witness_check.py).
aperture_generation_guidance.md — Methodological Guide: Principles separating Spine \(\mathcal{T}\) from evolvable aperture basis sets \(\{\mathcal{E}_{k,m}\}\), the \(5 \times 4\) functional homology heuristic, per-aperture declarations, basis-substitution rules, zero-leakage holdout standards, and the Four-Axis Audit Protocol.
empirical_validation_protocol.md — Experimental Protocol: This four-stage validation architecture and three-level epistemic falsification framework (Stage IV procedures are withheld in document 11).
tier_1_ontological_grammars.md — Tier 1 (Foundational Epistemological & Linguistic Ontologies):
1.1 Akbarian Relational Metaphysics (\(\mathcal{E}_{\mathrm{Akbarian}}\))
1.2 Wu Xing Five-Phase Dynamics (\(\mathcal{E}_{\mathrm{WuXing}}\))
1.3 Algonquian Process-Relational Syntax (\(\mathcal{E}_{\mathrm{Algonquian}}\))
1.4 Huayan Interpenetration & Dependent Co-Arising (\(\mathcal{E}_{\mathrm{Buddhist}}\))
tier_2_abiotic_continuum_physics.md — Tier 2 (Abiotic, Continuum & Tectonic Physics):
2.1 Condensed-Matter & Phase Transition Physics (\(\mathcal{E}_{\mathrm{condensed}}\))
2.2 Continuum Fluid & Wave Mechanics (\(\mathcal{E}_{\mathrm{continuum}}\))
2.3 Atmospheric Radiative & Thermal Boundary Physics (\(\mathcal{E}_{\mathrm{radiative}}\))
2.4 Geomorphic & Structural Mechanics (\(\mathcal{E}_{\mathrm{tectonic}}\))
tier_3_ecological_living_networks.md — Tier 3 (Ecological, Metabolic & Phenological Networks):
3.1 Vascular Plant Physiology & Rhizosphere Coupling (\(\mathcal{E}_{\mathrm{botanical}}\))
3.2 Movement Ecology & Ecotone Foraging Dynamics (\(\mathcal{E}_{\mathrm{trophic}}\))
3.3 Benthic & Edaphic Biogeochemical Metabolism (\(\mathcal{E}_{\mathrm{metabolic}}\))
3.4 Chronobiology & Phenological Synchrony (\(\mathcal{E}_{\mathrm{chrono}}\))
tier_4_somatic_contemplative_interiority.md — Tier 4 (Somatic, Neuro-Attentional & Affective Interiority):
4.1 Myofascial Biotensegrity & Postural Mechanics (\(\mathcal{E}_{\mathrm{somatic}}\))
4.2 Vestibular Equilibrium & Sensorimotor Contact (\(\mathcal{E}_{\mathrm{kinesthetic}}\))
4.3 Neuro-Attentional & Autonomic Regulation (provisionally homological) (\(\mathcal{E}_{\mathrm{neuro}}\))
4.4 Affective Regulation & Psychophysiological Resilience (\(\mathcal{E}_{\mathrm{affective}}\))
tier_5_expressive_collective_symbiosis.md — Tier 5 (Acoustic, Poetic & Cross-Substrate Symbiosis):
5.1 Environmental Acoustics & Rhythmic Entrainment (\(\mathcal{E}_{\mathrm{acoustic}}\))
5.2 Poetic Prosody & Psycholinguistic Condensation (provisionally homological) (\(\mathcal{E}_{\mathrm{poetic}}\))
5.3 Spatial Choreography & Relational Kinesiology (provisionally homological) (\(\mathcal{E}_{\mathrm{choreographic}}\))
5.4 Inter-Subjective & Cross-Substrate Symbiosis (provisionally homological) (\(\mathcal{E}_{\mathrm{symbiotic}}\))
annex_a_auditor_motif_hygiene.md — Auditor-Only Annex (withheld from Stage III conditioning): Categories and rules used in holdout-leakage review.
annex_b_stage_iv_sensor_procedures.md — Stage IV Procedures Annex (withheld from Stage III conditioning): Instrumentation, protocols, and falsification criteria for the five Stage IV test families (Test 2 deferred in this epoch); audited in Stages I and II under all four axes.
r10_branch2_witness_check.py — Spine verification script (accompanies the spine; not conditioning): exact symbolic checks of the r10 §5.13 Branch II witness and a Branch I regression.