Benchmark and research instrument guide¶
This directory contains performance measurements, conditional model probes,
reserved-response producers and read-only record audits. These roles are
not interchangeable. The theory catalog owns mathematical
claims; the execution plan
is the sole task queue. A retained instrument is not automatically an active
research campaign, and a filename containing emergent proves no emergence.
Instrument families and owners¶
Engine, response and evidence¶
| Question | Instruments | Owner and scope |
|---|---|---|
| Does a held source retain information absent from regional contrast/rate? | Source-relative response | Derived form; supplied affine law and zero-source ablation |
| How do supplied joint form/phase laws differ? | Constitutive probes, cotangent C8 | Relational exchange, cotangent scope; probes and trajectories are distinct |
| How does capacity change the response? | Capacity producer, record audit | Capacity response; the audit does not rerun trajectories |
| Can a prepared temporal readout identify chi with bounded error? | P2 acquisition, read-only audit | Temporal admission; known-source software control, prior-derived nonlinear bounds and native Euler defects, not blind physical calibration |
| Can bounded phase samples distinguish the two named capacity laws? | K3 sampling response | Frozen interval-Taylor method; supplied models, certified sample errors and shared observations, not unknown physical-law selection |
| Does a prepared pattern recover or transmit deformation? | Local recovery, regional interaction | Joint-law owner; supplied support and reference geometry |
| Which observations close, and what hidden state carries memory? | Local composition, memory coefficients, reserved response | Composition, memory; fixed rational probes and finite/continuous error have separate scopes |
| Does an intermediary transmit a capacity-dependent response? | Mediated response | Finite mediator protocol; shared tangent memory versus omitted or instantaneous memory, with a frozen-capacity control |
| Can prior visible observations support a later sine-law prediction? | Three-stage prior forecast | Joint admission and analytic separation; separate preparation, prediction and evaluation, synthetic derivative evidence, full uncertainty and a counterfactual frozen intermediary |
| Does a return path distinguish regional orientations? | Static return geometry | Equilibrium admission; exact cycle reconstruction, certified scalar root and native residuals, with no trajectory or event |
| Does a geometric interval constrain a separate dynamic response? | Geometry-to-response producer | Inverse geometry and response; separate preparation, prediction and evaluation with fixed scales, one known-source control and no positive-time trajectory |
| Can retained environmental exchange induce finite regional winding? | Conservative winding certificate | Whole-window acquisition and distributed source geometry; one shared analytic instrument, full-law bounds, equal-storage reflection control for acute entry, no trajectory search or indefinite-maintenance claim |
| Does one retained environment produce a finite acute regional episode, and where does its storage go? | Conservative regional organization | Frozen environmental control, collective-pulse control, channel accessibility and reversible preparation; one instrument and separately frozen declarations. Optional evaluation_kind=reversible_preparation checks a source-to-retention inclusion. Historical defaults and saved-tube channel readers retain their own schemas; invocation modes belong below |
| Can the rational saddle preparation connect absent winding to a finite acute identity? | Metric connection trial | Retained-metric proof, shared connection owner; two fresh full-state directions, frozen literal source and whole-time growth/identity checks. A same-orbit connection is distinct from an independent source ball |
| Does the native law have a causally shared oscillatory mode? | Static collective pulse | Pulse admission; ideal trace-sign certificate, native tangent poles/residues and no-communication control, without a sustained-pulse claim |
| Does a finite trajectory change sector or enter a sufficient basin? | Formation response, capture response, endpoint audit | Capture scope; endpoint evidence is not a whole-path error bound |
| Is an ideal continuous transit enclosed rigorously? | Transit proof | Validated transit; also zero-form and reversed-form controls, with their own frozen protocols |
| What does the fixed regular source/receiver preparation do initially? | Seeded short response | Eight-coordinate continuous proof; exact-pi IVP, retained regular tubes and loss, no capture claim |
| What holds for directed transport? | Dynamics, U2 metrics, transients, time, capacity boundary | Directed transport; conditional matrix/metric comparisons |
| What do support changes and capacity actually do? | Birth/transport, pressure feedback, forced support, capacity, selection | THOL, capacity, support balance; declared operators and laws |
| Can prior THOL or C6 evidence answer the question without another run? | thol_*audit.py, regional THOL instruments, c6_winding_*.py |
THOL feedback, C6 audit; retained chains of producers, exact controls and audits, not independent new campaigns |
Observation, auxiliary models and performance¶
| Role | Instruments | Boundary |
|---|---|---|
| Static and temporal observation | Static interface, temporal, multichannel, external phase gate | Interface guide owns measurement preparation, label availability, estimator limits and fair comparisons |
| Terrestrial transport comparisons | TCLab, Volts comparison, Volts loader | TCLab protocol, passive transport; exploratory maps are not admitted physical variables |
| Field diagnostics | Method battery, curvature example | Tetrad; declared diagnostic scope, not autonomous dynamics or universal safety thresholds |
| Arithmetic and spectral comparisons | arithmetic_*.py, composition, commutant, Paley, residue/prime/word controls |
Arithmetic; supplied encodings, exact algebra or finite comparisons |
| REMESH/Riemann projections | remesh_infinity_riemann_*.py |
Riemann scope; parked spectral controls, not RH or complete-runtime stability proofs |
| Auxiliary geometry | emergent_*.py, pulse and symmetry probes |
Ontology boundaries; supplied graph/wave/simplex models, not identified particles or physical dimensions |
| Actual performance | Node lookup, offset batching | Source, seed, workload, timings and trajectory hashes; no universal speedup claim |
The complete source inventory can be read without importing or running scripts:
rg --files benchmarks -g '*.py'
Paths remain stable where tests, imports and frozen source manifests consume them. Grouping the instruments here avoids moving a scientific producer merely to make a directory look tidier. Inspect the source and its owner before using an unlisted family member; do not batch-execute this inventory.
Reuse the mechanism appropriate to the question¶
Use the SDK for operator studies and its regional/relational guide for shared observations, explicit joint evolution and capture admission. Benchmarks add preparations, comparisons and evidence, not a second production solver.
Prefer a read-only record audit when the needed trajectory already exists.
An audit separates record consistency, the original decision and compatibility
with current source/runtime. Some THOL/C6 audits require local records under
artifacts/research/ that are not versioned or supplied by cloning the repo;
missing records make those checks unavailable. Sparse checkpoints cannot authenticate omitted
steps or reconstruct a continuous integral. A content digest does not establish
physical acquisition, trustworthy chronology or a complete transitive dependency
archive unless that archive is explicitly included.
Use derived memory and the scale bridge for closure questions. The auxiliary gasket instruments reuse the pure graph preparation, including its ordered outer corners. That supplied construction is not a THOL event or a selected physical dimension. The Bell instrument now uses those corners at every level; earlier level-two and higher fixtures used the first inserted nodes instead. The CHSH bound is unchanged, while those graph fixtures are corrected; historical outputs were not recomputed.
A rank calculation for rational surrogate coefficients is not a nonlinear closure proof for actual irrational coefficients. An observation of winding, a standing mode or a supplied potential does not identify a physical particle.
Running and reporting¶
Install the repository in editable mode, then run the selected instrument from the repository root. Dependencies and command-line support differ by module:
python -m pip install -e ".[dev-minimal]"
python -m benchmarks.nodal_lookup_scaling --sizes 10 --repeats 2
python -m benchmarks.stable_node_offset_scaling --sizes 6 --repeats 2 --scoped
These small examples exercise timing/report plumbing; meaningful performance
comparisons require identical input, source, hardware, backend, warm-up and
measurement scope. Retain absolute times and trajectory agreement, not just a
speed ratio. --wheel selects a separate baseline without installing it.
pytest benchmarks --benchmark-only is not a supported directory-wide suite;
scientific regression tests live under tests/. The optional test-performance
extra supplies a third-party plugin only when explicitly requested.
Freeze before evaluating a reserved response¶
The following producers accept --prepare --output PATH; evaluation uses the
same command and path without --prepare:
Producer module under benchmarks |
Role |
|---|---|
source_relative_form_response |
Held-source prediction and ablation |
phase_form_exchange_comparison |
Prepared constitutive probes, not a trajectory |
relational_capacity_response |
Capacity intervention |
relational_coefficient_acquisition |
One known-source P2 acquisition; full retained state/error chain and source archive |
relational_capacity_sampling_response |
Four fixed K3 candidate/arm cases; single-origin Taylor enclosures and frozen sample/contrast error gates |
relational_seeded_response |
Fixed regular source/receiver short response; exact-pi interval initial state, whole-time proof and storage ledger |
relational_receiver_barrier |
One fixed full eleven-node sine source; whole-prefix receiver potential and dissipative full-storage tail, with failed or unresolved evidence retained |
relational_local_recovery |
Prepared local recovery |
relational_region_interaction |
Transmitted regional response |
relational_formation_response, relational_capture_response |
Finite sector/endpoint responses |
relational_memory_response |
Matched-grid memory prediction and full-engine comparison |
relational_mediation_response |
Matched-grid mediator-capacity intervention; retained memory versus instantaneous and omitted-memory controls |
relational_transit_proof |
Continuous proof audit; --zero-form and --reverse-form are mutually exclusive controls |
conservative_regional_organization |
Whole-time regional observation or reversible preparation, selected by the declaration; --declaration PATH requires an explicit new --output PATH and the same declaration in both preparation/evaluation invocations |
For example, use a fresh path for a current-source regression:
python -m benchmarks.relational_capacity_response --prepare --output artifacts/relational-capacity/result.json
python -m benchmarks.relational_capacity_response --output artifacts/relational-capacity/result.json
The inverse-geometry instrument has a third, independently retained prediction stage. For a separately declared current-source check, use one fresh path for all three invocations:
python -m benchmarks.return_path_geometry_response prepare --output artifacts/research/return_path_geometry_response/current-check.json
python -m benchmarks.return_path_geometry_response predict --output artifacts/research/return_path_geometry_response/current-check.json
python -m benchmarks.return_path_geometry_response evaluate --output artifacts/research/return_path_geometry_response/current-check.json
This fixed known-source protocol checks a bounded initial derivative. It does
not run a positive-time trajectory or supply an independent physical measurement.
The analytic conservative_regional_winding instrument instead writes one
certificate and source archive; it has no --prepare stage and executes no
solver. Its --declaration chooses the preparation, not a parameter search.
conservative_regional_organization --analyze-channels SOURCE_RESPONSE
--output NEW_PATH is a third kind of operation: a retrospective analysis of
saved whole-time tubes. It does not take --declaration, replay a solver or
rewrite the original response. Keep the separate analysis source archive,
original file hashes and original verdict. A successful file audit or a new
channel bound does not convert a negative acquisition result into a success.
sine_metric_connection uses a separate declaration and mandatory output:
python -m benchmarks.sine_metric_connection --prepare --declaration PATH_TO_DECLARATION --output artifacts/research/sine_metric_connection/new-response.json
python -m benchmarks.sine_metric_connection --declaration PATH_TO_DECLARATION --output artifacts/research/sine_metric_connection/new-response.json
Preparation verifies the literal rational saddle recipe without evolving it. Evaluation executes both declared directions once through the shared engine, retaining whole-time matrix and identity evidence, including partial failures. Its study-level pass requires the same-orbit connection and both complete horizons. An image of the prepared family remains distinct from an independent winding-zero source box. The retained declaration is not permission to replay an already evaluated response or scan alternative budgets.
The retained evidence bundle
contains the original response, protocol and evaluated source archive. Its
manifest
records their unchanged byte hashes. Read the
result owner
for the distinction between certified connection prefixes and the incomplete
declared horizon; the producer's original passed=false is retained.
The same-source constitutive audit consumes that bundle without rerunning the producer. Its fixed one-percent cubic-current comparison uses the unchanged source's full conserved storage and a proved sector barrier; see the contract and retained result. It is a read-only model comparison, not another trajectory campaign.
Preparation and evaluation are separate invocations. Keep sibling prediction
or protocol files and any source archives together. For a separately identified
numerical correction of relational_seeded_response, use a fresh output path
and supply --correction-of PRIOR_RECORD to both invocations. This retains the
prior bytes' digest and labels the calculation as a correction, not an
independent blind replication.
For the fixed source/receiver target-budget study, pass --study target-budget
and an explicit fresh --output to both invocations. This declares total time
one from the original exact initial state; it does not substitute the previous
response's rounded endpoint. A separately frozen continuation can pass the
same --horizon 9/8 to both invocations, still starting from the original IVP.
The shared owner admits a positive exact horizon of at most 256 steps; no
automatic extension is performed. A resolved negative budget and a resolved positive
budget both pass the numerical discrimination gate. Read the separate target
verdict: positive storage is not successful formation. The default short
study and its retained declarations remain unchanged.
Admission binds declared inputs and recorded source/runtime fingerprints; changing source bytes, line endings or dependency versions can invalidate a replay. Prepare a new pair rather than rewrite the first experiment. Use the CLI rather than feeding raw JSON into producer functions that expect native exact-rational mappings.
docs/assets/ contains immutable first predictions, responses and source
archives linked by each theorem owner. New output belongs in artifacts/.
The commands above describe instrument use, not an instruction to replay every
retained experiment. Consult the execution plan before a new scientific run;
documentation, interface and admission checks normally reuse saved evidence
or independent bounded fixtures.
Failed decisions must remain visible; an inconclusive certificate is not a
pass. The later continuous transit certificate does not replace the earlier
finite-executor verdict. Consensus (target_sector=0) and unavailable evidence
are different results; do not interpret the sector as a Boolean.
For every new numerical claim, retain source/configuration, input construction, seed, backend/precision, law or operator sequence, raw artifact and provenance. Separate calibration from evaluation and disclose supplied labels or factors. Finite measurements establish their measured scope, not universal physical laws.
Read-only temporal-acquisition audit¶
Inspect the existing P2 response without regenerating it:
python -m tnfr.research.relational_acquisition artifacts/research/relational_coefficient_acquisition/result.json
This command reads the sibling protocol and archive, then reconstructs saved
contrast/error evidence through the shared auditor.
It has no prepare/evaluate mode and writes no artifacts. Exit 0 means record
consistency, not that the experiment passed; recorded_passed and
reconstructed_passed retain the original verdict separately. Exit 1 denotes
inconsistent evidence and 2 missing/unavailable files. Local artifacts may be
absent after cloning; their absence is not permission to replay the producer.
The contract
owns admission, work limits and the authentication boundary.
Multichannel observation records¶
The multichannel instrument computes one window series for both ranking and
block summaries. Malformed ARFF rows reject the record rather than silently
removing samples; binary labels must remain aligned with signal samples.
The byte limit covers cached payloads and expanded ZIP members as well as the
download. Unavailable class means serialize as null; reports use the shared
strict JSON and atomic writer.
Check auc_available and auc_unavailable_reason before interpreting scores.
When windows or a label class are absent, retained legacy AUC values of 0.5
are compatibility placeholders, not measured chance performance. Available
orientation-free rankings remain same-sample descriptions, not reserved
forecasts; the interface guide
owns the methodological boundary.
Retired instruments¶
The retirement record and recovery manifest identify unsupported timing comparisons, obsolete threshold/U2 investigations and orphan utilities removed in this audit. Useful mathematical results, negative boundaries, maintained tests and frozen evidence remain with their owners; retirement does not assert that every alternative model is impossible.