LRA engine integration

bijux-pollenomics-lra owns numerical REVEALS, analytical and optimized LOVE, coupled reconstruction, exact-coordinate series and forward pollen deposition. The canonical host owns source admission, chronology, taxa, experiment splits, commands and product publication. The extension imports neither the host nor its development toolkit. Its implementation is Python, NumPy and SciPy.

The supported Python facade is bijux_pollenomics_lra.api. Installed metadata registers group bijux_pollenomics.extensions, name lra, target bijux_pollenomics_lra.api:extension_descriptor. Loading checks installed identity and compatibility; missing dependencies and incompatible implementations are observable refusals. There is no standalone engine command or automatic fallback.

Engine domain	Responsibility
models	Configuration, results, validated dimensions and errors
dispersal	Deposition factors and source-compatible LOESS curves
reconstruction	Regional/local equations, draws and forward observation model
optimization	Bounded solver, replay and convergence diagnostics
workflows	Stage composition and coordinate pairing
provenance	Input, backend and scientific content identity
outputs	Numerical serialization and byte checksums
resources	Versioned models, schemas and examples

Tests mirror those responsibilities. Exact-source R tooling lives under tests/reference and does not ship in the runtime wheel. The numerical core does not write products. Result serialization is explicit; the host decides whether an output is scientifically eligible for release.

The adopted calculations retain the explicitly named DISQOVER 0.9.14 source profile. Source geometry conventions, analytical estimates outside physical fractions, uncertainty omissions and optimizer failure are observable. A numeric pairing coordinate is not a canonical source ID or an automatically inferred age. Publication-specific multi-site aggregation requires its own verified profile.

Pollen and vegetation validation have different targets and scores. Holding every target pollen sample out requires a spatial predictor and a forward observation model; LOVE itself needs local pollen. Grouped folds and dependency closure guard the training boundary. Synthetic recovery and R parity establish numerical behavior; they do not establish ecological accuracy or qualified review.

prepare_forward_inputs constructs target-specific kernels and regional deposition through the engine facade, without target pollen or a target count total. Its corrected-matrix preparation profile is recorded separately from the forward observation profile. The host retains the parameter, geometry, time and training-source lineage; it never substitutes a training basin for target geometry.

Hatch/VCS, the workspace lock, Make/tox, test discovery and CI govern the package. Its GPL license and source attribution are package-owned. Repository license synchronization must not replace them with the host's Apache license. The release matrix builds the optional engine alongside the host. Publication checks bind staged distributions to the release revision and require a same-revision actual-R certificate. Matrix inclusion and local build evidence do not establish scientific acceptance or authorize publication.

A bounded C++ objective experiment agreed numerically with Python and R but did not meet its predefined performance thresholds against prepared NumPy. That kernel remains Python. This finding does not establish that every possible native optimization is unsuitable; further native adoption requires measured benefit, binding safety and platform evidence through the same Python facade.

Host boundary

The host facade is bijux_pollenomics.analysis.lra.api; its inputs, engine, execution and outputs packages have distinct owners and mirrored unit tests. The CLI delegates to the same service. validate-lra-input verifies a declared request and installed configuration; run-lra executes it and publishes a new local exploratory bundle. Neither command grants source or scientific approval.

The request records sample/site identities separately from exact numeric pairing coordinates. All source age intervals remain typed BP support. An explicit declared_bin_stationarity decision identifies common time bins containing the full support of every paired sample; matching a numeric key alone cannot establish chronological compatibility. This is a declared modeling assumption, not a chronology review verdict. Included/excluded sample identities must be disjoint and unique.

Geometry records use metres and square metres, with explicit source references. The optimized profile consumes terrestrial-to-ring area ratios; it needs regional rings beyond the separately declared local max_ring. The analytical profile does not apply partial terrestrial masks and refuses them. Arithmetic checks do not independently verify source geometry. The optimized source objective uses nominal PPE; its nested method manifest retains that source limitation even when other stages draw PPE uncertainty.

Products retain source-file hashes, explicit omissions, numerical units, actual installed engine identity, code/resource inventories, diagnostic files and separate operational/scientific hashes. Source and code changes during a run refuse publication. Before/after file inventories are mutation checks; they do not attest Python objects already loaded in a long-lived process. Certification uses fresh installed processes and pinned dependencies.

Product reopening and cache reuse belong to the host's output and execution domains. read_lra_product checks the complete file inventory, request and scientific identity graph, rejecting incomplete bundles, symlinks, unexpected files and changed numerical bytes. An independently retained manifest digest can identify an expected bundle; internal consistency alone does not establish authorship or scientific approval.

Saved primary tables must agree with the request's taxa, numeric coordinates, ring support, method and declared result units. Numeric fields reject malformed or nonfinite values; source-defined empty SD/local-statistic cells remain missing. Binary Monte Carlo arrays are byte-verified, without claiming a fresh numerical recomputation. A final inventory check detects changes during reopening; returned paths are not a promise that another process cannot subsequently change the files.

run_lra_cached revalidates source evidence and numerical configuration on every call. The key includes the complete request, installed engine identity, host and engine code/resources, Python/platform and dependency versions, and declared numerical thread environment. Parameter, source-identity or implementation changes select a different entry. A cache hit verifies the saved product before returning it. Corrupt entries refuse without replacement; simultaneous writers receive lra_cache_busy without polling. Interrupted writers may leave a lock and staging directory for explicit recovery after checking that its owner has stopped. The cache does not automatically infer that a lock is stale. This is a local product cache; no distributed scheduling or measured batch throughput is implied by this interface.

Directory publication requires an atomic no-replace operation on the destination filesystem. The Linux and macOS paths use their native exclusive-rename primitives; Windows uses its nonreplacing rename behavior. Unsupported systems or filesystems refuse publication instead of falling back to a racy existence check. Failed staging directories remain available for diagnosis. Local macOS execution evidence does not certify the other operating systems.

The primitive contracts are documented by the Linux rename manual and Apple's rename flag definitions.

Source admission and experiment ownership

analysis.lra.materialization owns the source-to-request boundary. Its Neotoma admission audit retains exact source cells, count units, chronology alternatives, conflicts and per-sample refusals. Necessary numeric count eligibility does not authorize taxon grouping, PPE, chronology or basin geometry. Source audits remain separate products until those scientific requirements have actual evidence.

The SEAD source-admission audit is available through analysis.lra.materialization.sead.audit_sead_lra_source. It consumes caller-pinned full-evidence acquisition identities, the existing normalized bundle and its country-boundary authority. The source-owned normalizer is replayed against every logical document before audit rows are emitted. Observation records retain native values, units, source taxonomy, elements, dataset methods and dimensions; chronology claims remain unselected at their original analysis entity. Numeric integer compatibility, an explicit source absolute-element-count declaration and an observed Palynology method link are reported separately. None constitutes approved pollen grain counts. The output is a hash-bound audit with zero scientific admissions, without importing or running the numerical engine. AdmissionReport.available_samples and selected_samples count unique entity-linked physical samples in this all-source audit; the manifest separately accounts for analysis entities, observation-bearing entities, the complete raw physical-sample denominator and samples without analysis entities. All61source-table row counts remain explicit, including empty tables.

The private analysis.lra.materialization.requests facade projects explicitly declared synthetic Neotoma fixtures into the existing lra-request.v1 contract. It requires synthetic origin and country inputs bound in the raw receipts, replays the existing source normalizer across all normalized surfaces and reconciliation, and binds every selected count cell and chronology claim to its source records. It preserves zero, refuses missing or incompatible values and incomplete denominators, retains the original evidence, and publishes through the shared no- replace filesystem primitive. Request projection does not import the optional numerical engine. Real or unspecified source provenance is refused; this facility supplies no qualified taxonomy, chronology, PPE, geometry or empirical reconstruction approval.

The shared analysis.classification.events.assess_classification_eligibility service checks reviewed source/taxon membership without requiring positive pollen detection. Zero, null and absent measurements remain distinct. This review projection does not admit counts, chronology or an LRA run. Its current pinned authority is Neotoma-specific and contains zero accepted mappings; the API cannot create scientific approval or extend that authority to another source.

analysis.lra_validation.experiments owns training-fold reconstruction and subsequent observation scoring. Its first profile is explicitly synthetic: prepare_experiment_predictions runs the existing batch/cache service, reopens training products, applies a named baseline and prepares target-specific forward geometry. The regional anchor retains its full training ancestry and is separately identified from local predictor sites. Both inversion and forward grids declare the same physical basin and selected outer support radius.

Prediction publication precedes score_experiment_predictions, which accepts a separate outcome inventory and records attempted outcome access. Every requested fold, baseline and target receives a score or explicit failure/exclusion result. Changing held-out pollen cannot affect that site's training fit or frozen prediction. Vegetation and pollen are scored separately; training mean, nearest site and inverse-distance-squared are named baselines. Their deterministic predictions have no predictive intervals, and overlapping-fold scores do not constitute independent fixed-model evaluation units. The synthetic profile does not admit final evaluation or empirical source observations.

The host CLI exposes the same two phases. Set the existing batch service's numerical thread controls, prepare the declared synthetic experiment, and retain its reported manifest_sha256 before scoring:

export OMP_NUM_THREADS=1 OPENBLAS_NUM_THREADS=1 MKL_NUM_THREADS=1
export NUMEXPR_NUM_THREADS=1 VECLIB_MAXIMUM_THREADS=1
bijux-pollenomics prepare-lra-experiment experiment.json \
  --input-root inputs --cache-root cache --output predictions --max-workers 1
bijux-pollenomics score-lra-experiment predictions \
  --expected-manifest-sha256 "$prediction_manifest_sha256" \
  --outcome-root outcomes --outcome-inventory outcomes/inventory.json \
  --output scores

Preparation dispatches the explicit request profile: synthetic withheld-site baselines, source-observation withheld-site baselines, fitted spatial prediction, or local-pollen assimilation. Local-pollen scoring accepts the separate vegetation-only outcome inventory; its pollen comparison remains an in-sample diagnostic. All profiles report the requested prediction count and retain their own target roles.

The scoring command verifies the expected prediction identity before opening the outcome inventory. Both commands emit JSON identifying the saved product and its accounting. Preparation exits 0 when every requested prediction was generated, 2 for a refusal or partial result, and 130 for an accounted-for interruption. Scoring exits 2 if any requested score is missing, refused or excluded. A zero exit reports operational success; scientific_acceptance remains not_evaluated. The original prediction and separate score products retain all requested denominators and target-specific results.

The exploratory source-observations-withheld-site-baselines-v1 profile requires request_template.data_kind: source_observations and raw observation files with the same declared data kind. It uses the same train-only grouped folds and source-byte checks as the synthetic baselines. The separate scoring inventory uses schema_version: lra-source-pollen-outcomes.v1 and data_kind: source_observations; each record reference must point to the exact raw file and SHA-256 recorded in the prediction catalogue. Scoring opens those counts after prediction authentication and emits only held-out pollen scores. Predicted vegetation is not measured vegetation accuracy. Source identity, chronology, basin geometry, taxa, parameters and permissions still need an eligible cohort; this software profile does not grant source or release approval.

prepare_local_pollen_predictions adds the distinct synthetic local-pollen-available profile. It runs actual paired regional REVEALS and target analytical LOVE, retaining the regional anchor and explicitly observed target pollen channel in their original outer fold. Independent vegetation outcomes are opened only during the separate scoring phase. Its predicted pollen is an in-sample diagnostic against the assimilated counts and is explicitly ineligible for held-out pollen acceptance. This method is not a renamed spatial baseline.

Baseline, local-pollen and fitted preparation reverify their consumed pollen, geometry and template evidence immediately before publication. A changed source refuses publication; unused pollen and outcome files remain outside that check. Interrupted batches retain explicit incomplete rows and support cache-based resume. The local profile currently produces point predictions, with uncertainty and empirical acceptance remaining separate requirements.

The separate analysis.lra_validation.spatial.fit_spatial_model facade fits clr-wgs84-rbf-ridge-v1: centred log-ratio vegetation fractions regressed on radial basis features centred only at declared training sites. It reuses the host's unrounded WGS84 distances. Lengthscale, ridge penalty and matched-support identity are required declarations. All training fractions must be positive; zeros refuse without an invented replacement rule. The numerical least-squares backend is loaded only when fitting, through the optional NumPy dependency.

The CLR definition and ridge objective define its transformation and fitted objective. This is an exploratory geographic model choice, not evidence that the model is ecologically accurate. It has an unpenalized intercept, one equally weighted composition per predictor group, and explicit solver diagnostics. The basis features do not assume a geographic covariance kernel. Every fitted coefficient, centre, parameter and denominator is bound to the model digest. Target coordinates are supplied after fitting; target observations are not accepted. The profile produces point predictions, with uncertainty, parameter selection and final evaluation left to separately governed experiment stages. The current baseline experiment runner does not silently substitute this fitted model for any named baseline.

TrainingAdmission records an explicit whole-group subpartition of an original outer fold. It binds each physical group's complete record membership, admitted and excluded training records, and the declared selection criterion. Excluded training observations cannot re-enter through external dependencies; they do not become test observations. A fitted model with this admission uses a distinct v2 manifest, while the default all-training model retains its existing identity.

Shared support separation uses unrounded WGS84 distance minus both footprints and coordinate uncertainties. The same arithmetic serves experiment buffering and final evaluation. The declaration alone does not prove the criterion was applied; the experiment must retain and validate its actual selection evidence.

analysis.lra_validation.temporal.admit_forward_time_training evaluates explicitly declared synthetic calendar supports on the BP1950 axis. Every source/target interval-component pair must establish positive order, satisfy the declared embargo and stay within the maximum lead. Admission compares exact binary64 endpoint ratios; rounded reported differences never decide a boundary.

The temporal receipt preserves the original fold, whole physical-site groups, missing support and required regional-source exclusions. If any alias fails, its entire group is excluded. Missing required regional support leaves otherwise eligible groups visible but supplies no usable training admission. Consumers must still validate their quantitative dependency closure with assert_training_inputs. This component reads no pollen or vegetation outcomes and performs no prediction, chronology conversion, review authentication or uncertainty propagation.

prepare_fitted_spatial_predictions connects training-only regional REVEALS and local LOVE products to the explicit CLR spatial model, followed by target-specific forward pollen deposition. Its synthetic singleton folds withhold both target observations. Prediction products retain training products, regional ancestry, geometry, fitted state and separate vegetation/pollen fractions.

The current grouped-outer producer has a versioned paired payload. Before fitting the spatial model it freezes a distinct training-mean model over the same exact original-fold reconstructions, then projects both methods through the governed target forward model. Reconstruction failure applies to both methods, while a spatial-fit refusal can coexist with a successful baseline. Every fold/method is predicted, refused or part of the deterministic interrupted not_run suffix. The strict reader recomputes the baseline mean and complete input closure and keeps its admitted_reconstruction_training_mean identity separate from the fitted model. Historical fitted-only products retain their legacy reader branch; the paired schema never accepts an arbitrary method subset.

prepare_spatial_block_predictions uses a separate declared spatial-block profile. It preserves the original physical-site groups, excludes training groups using unrounded footprint separation, selects an admitted regional anchor and fits one model for each held-out block. Each target then receives its own spatial prediction and forward pollen calculation. The same admitted reconstruction set also freezes an independent training-mean vegetation baseline and projects it through the governed target-specific forward model. All target pollen remains withheld.

Saved block products retain the complete experiment catalogue, admission decisions, the fitted state when available, and a separately identified baseline model. A fitted-model refusal can leave the already frozen baseline usable; a baseline target refusal does not discard its fitted counterpart. The reader regenerates the declared folds and verifies source, support, model and target identities before scoring. Every fold/record/method key is exactly predicted, refused or not run, with separate counters for the latter two states. Score pairs include the target record identity, so sibling targets in one block cannot reuse one another's training-mean score. Successfully consumed inputs remain subject to final hash checks even when later reconstruction fails; this lifecycle rule also applies to singleton fitted and local-pollen experiments.

The retained reader recomputes baseline vegetation arithmetic and authenticates its admitted reconstruction lineage. Forward pollen bytes remain protected by the caller-pinned product manifest rather than replaying a potentially different optional engine after publication. This boundary does not supply uncertainty, empirical accuracy or scientific acceptance.

adapt_authenticated_spatial_block_local_vegetation_rows joins the strict spatial-block local-pollen reader to that paired training-mean product before any vegetation outcome is accepted. It requires the same complete fold universe, block admissions, target group/bin/taxon denominator, geometry and per-target training admission. The analytical LOVE row receives its own target-specific local_pollen_assimilation source identity and explicit observed-pollen channel; it is never relabelled as a fitted spatial model. Its training-mean comparator is authenticated from the separate block product and cannot carry target pollen or a fitted-model alias. Predicted rows are frozen with the generic v2 envelope, while parent refusals/not-run states remain failures without invented values.

This adapter supplies only the outcome-free spatial-block local_pollen_available vegetation pair. No governed predictive-observation interval for local vegetation exists, so the cell remains ineligible for final coverage acceptance. The adapter does not bind a release receipt, open vegetation truth, establish first-access timing, approve a threshold or grant scientific review. The grouped-local vegetation adapter and the factory-only six-cell release product remain separate boundaries.

adapt_authenticated_spatial_block_withheld_vegetation_rows performs the paired point-row projection for the spatial-block all-target-withheld vegetation cell. It reopens the caller-pinned product, requires the exact frozen fold and complete fitted/training-mean terminal universe, checks target group, bin and taxa, and preserves refused/not-run parents as value-free failures. Successful rows carry the producer-authenticated fitted or admitted-training-mean identity and remain free of target pollen and vegetation outcomes. No governed vegetation predictive interval exists, so this adapter also cannot establish empirical coverage or release acceptance.

adapt_authenticated_spatial_block_withheld_pollen_rows builds the matching complete pollen point-row universe and sends it through one batch authentication of the caller-pinned conditional-observation product. Both the fitted method and training-mean baseline receive producer-authenticated v2 source identities and predictive-observation intervals. Parent prediction refusals remain value-free failures. A successfully predicted parent without a generated interval refuses the whole ledger rather than silently shrinking its uncertainty denominator. These conditional intervals propagate declared counting uncertainty only; empirical coverage, first-access timing and scientific acceptance remain false.

These spatial experiments remain synthetic and exploratory. Joint model uncertainty requires separate profiles and evidence.

read_withheld_predictions authenticates a caller-pinned saved baseline, legacy fitted-only or versioned paired grouped-outer product before outcome access. It rechecks original folds, the exact method/terminal universe, training lineage, forward dimensions and exploratory status, and returns independently owned data. Reading and scoring retained products work without installing the numerical engine. These consistency checks establish neither external authorship nor scientific acceptance; numerical reproduction is separate evidence.

adapt_authenticated_grouped_withheld_vegetation_rows joins the paired grouped-outer product to the exact frozen evaluation scope. It authenticates the two-method terminal denominator, singleton target ownership, support, taxa and the complete training/external dependency closure. Successful fitted and training-mean rows receive distinct source identities; refused and interrupted parents remain value-free. This is an outcome-free point ledger and supplies no vegetation uncertainty or accuracy claim.

adapt_authenticated_grouped_withheld_pollen_rows builds the corresponding pollen point ledger and authenticates the complete conditional-observation denominator once. Every successful parent must have its generated interval; authenticated parent failures remain value-free and cannot shrink the ledger. The intervals propagate only the declared pollen counting process and do not establish empirical coverage, preregistration timing or scientific acceptance.

read_local_pollen_predictions, exported from both experiment facades, authenticates grouped-local request bytes, raw batch operations, numerical products and exact fold terminals. Operational failures and unobserved items are published as scientific refusals unless a genuine interruption starts the deterministic not_run suffix. Raw status and error evidence remain bound to each attempted failure; a post-completion projection refusal is recorded separately from the completed numerical batch item.

adapt_authenticated_grouped_local_vegetation_rows joins that strict local product with the independently retained grouped training-mean comparator. It requires singleton target ownership, the exact fold/bin/taxon/geometry/support universe, target-specific observed-pollen bytes and the complete causal input closure. The comparator must declare the exact pollen-free training and external lineage admitted by the frozen scope. Either parent may retain an independent refused or interrupted terminal without inventing a prediction or source. The successful pair uses distinct authenticated local-assimilation and training-mean source identities and the generic v2 prediction envelope.

This completes the six outcome-free protocol/cell adapter paths, not a release evaluation product. Callers still cannot treat an assembled row tuple as release evidence; a governed factory must bind the six adapter outputs and their reopened product identities before any separately authenticated observation join. The grouped-local cell has no vegetation predictive-observation interval, outcome, threshold, historical first-access proof or scientific approval.

build_six_cell_prediction_product is that engineering factory boundary. It accepts only the two typed release protocols, their factory-minted release admission and fixed-schema path/digest locators. It dispatches the six adapters itself; callers cannot supply rows, manifest mappings, observations or adapter callbacks. Only after every governed reader has reopened and authenticated the retained local, withheld and conditional-uncertainty products does the factory record those product identities. It then requires the exact protocol, cell, fold, record, method and terminal denominator and reopens every successful v2 prediction envelope.

The resulting private-constructor SixCellPredictionProduct stores a canonical receipt and immutable outcome-free rows. Reopening the typed product regenerates release admission and repeats the row, source, taxa, interval and claim-boundary checks. This does not score outcomes: independent observation attachment remains a separate governed producer, and caller-built row tuples remain outside this release-product boundary. The receipt explicitly records that empirical evaluation and scientific acceptance were not performed.

build_observation_access_plan derives the next, still value-free boundary from that complete prediction product. It removes the model dimension so each protocol/cell/fold/record outcome will later be shared identically by the model and its baseline, binds separate typed locators for independent measured vegetation and held-out pollen counts, and records only mismatch guards and reader/reference identities. Building the plan never opens, stats or hashes an outcome path.

start_outcome_access reopens the plan, revalidates both policies at its internally captured UTC instant and requires every successful prediction to have frozen strictly earlier. It then exclusively publishes and fsyncs a canonical start receipt before returning. That receipt contains no outcome reads or values, refuses overwrite, and explicitly says it authenticates this process ordering—not historical human non-access. HYDE, ESSD and reconstruction output remain ineligible as independent measured vegetation; actual source products, observation joins, scoring and qualified review are separate downstream gates.

read_observation_source_products implements the next engineering gate. It accepts only that durable start receipt, reads the fixed measured-vegetation and held-out-pollen locators in order, verifies their exact bytes and governed reader identity, and publishes either a canonical completion receipt or a canonical failure receipt. Its two typed products retain complete available, missing and refused denominators; source snapshot/member, rights, reviewed-mapping and normalization identities; exact ownership and support; vegetation fractions or nulls; and nonnegative integer pollen counts, effort and all-zero status. Reopen replays those fields from embedded canonical source bytes without reopening an outcome path. These declarations are byte-bound, not independent authentication of a provider, rights holder or scientific reviewer.

build_authenticated_observation_product accepts only a successful typed source completion. It rederives the model-free access assignments and retains the exact source-unit envelopes. build_observed_six_cell_product then expands each assignment over the already frozen model/baseline pair. Prediction failures stay prediction failures; missing or refused observations remain a separate source terminal and do not rewrite prediction bytes. Available observations attach the same values, support and provenance identity to both comparison methods.

evaluate_observed_six_cell_product is the release-oriented scoring entry point. It rejects a raw tuple, mapping or lazy iterable before row access, reopens the joined product, and evaluates grouped-outer and spatial-block protocols under their own policy and scope. The result binds every prediction, access, source, observation and protocol parent while retaining missing and refused source statuses separately. evaluate_final remains the generic compatibility scorer; its result alone is not release-product evidence. Neither entry point verifies reviewer authenticity, grants scientific approval or supplies missing reviewed thresholds and real admitted observations.

bijux_pollenomics_lra.sample_pollen_observations supplies a separate conditional observation model. It accepts predicted pollen fractions, declared counting effort, seed, nominal coverage and the host's prediction/effort evidence hashes. The host must freeze those declarations before observation access; the numerical function cannot authenticate their origin. It accepts no target count vector.

Joint multinomial draws retain their draw IDs, taxon dependence and total effort. The result binds nominal and effective probabilities, the explicit binary64 closure operation, numerical backend and dependency versions. Independent returned dictionaries cannot mutate its canonical result. The sampler uses NumPy's Generator.multinomial. Internal sequential subtraction remains subject to floating-point rounding.

Binomial marginal predictive bounds condition on fixed probabilities and effort. For effort at most 256, exact rational CDF checks resolve discrete probability ties in the provider's returned quantiles. Larger supports use SciPy's binomial CDF and quantiles with an explicitly recorded numerical boundary check; ambiguous endpoints refuse. Discrete marginal coverage may exceed nominal coverage. Simultaneous coverage, empirical calibration and R-source parity are not claimed. These bounds omit spatial fitting, PPE, chronology, geometry, reconstruction and overdispersion uncertainty, and do not provide vegetation prediction intervals. The experiment runner still requires a separately governed integration of those uncertainties.

analysis.lra_validation.observation_uncertainty attaches this conditional sampler to an authenticated synthetic singleton-fold baseline or fitted product. An immutable ObservationUncertaintyPlan pins the parent and declares effort, seed, taxa and counting-design evidence for every requested fold/method unit, including failed predictions. Generation accepts no outcome inventory. Joint draw cells are bounded globally, and source, parent and implementation identities are rechecked before publication.

score_pollen_observation_coverage separately consumes the pinned prediction, uncertainty and existing score products. Observed counts must have the same taxa, bin and exact predeclared total; a mismatch is an explicit coverage refusal. Every requested unit remains counted. Results are per-taxon conditional marginal coverage, with neither simultaneous coverage nor independent-site sampling claimed. A supplied timing declaration cannot authenticate historical preregistration. The numerical engine is required to validate retained draws; no new simulation is run during coverage scoring. Local-pollen diagnostics, latent vegetation intervals and scientific final-policy admission remain outside this profile.

The CLI exposes conditional sampling and coverage as separate saved phases. Serialize the independently declared plan with ObservationUncertaintyPlan.as_dict(); its parent prediction and counting-design evidence hashes are required inputs.

bijux-pollenomics prepare-lra-observation-uncertainty predictions \
  --plan counting-plan.json --evidence-root counting-designs --output uncertainty
bijux-pollenomics score-lra-observation-coverage predictions \
  --uncertainty uncertainty --scores scores \
  --expected-uncertainty-sha256 "$uncertainty_manifest_sha256" \
  --expected-scores-sha256 "$scores_manifest_sha256" --output coverage

Generation reports every requested unit and its status, returning 130 for an accounted-for interruption and 2 for refusals or partial results. Coverage returns 2 when any requested unit is refused, including counting-effort mismatch. Both commands emit exploratory JSON; successful execution grants no scientific acceptance. Base installations support help and input refusals, while actual sampling and retained-draw validation require the numerical extra.

analysis.lra_validation.evaluation.api evaluates a complete, separately frozen observation ledger against an explicit v2 policy. The policy supplies reviewed reference identities, per-target thresholds, independent site/block minimums, permitted exclusions and predictive-coverage requirements. This API supplies no thresholds or scientific approval. Its receipt distinguishes passing engineering criteria from authenticating the declared review.

Every baseline/model pair must share its observation, taxon denominator and geographic/temporal support. Errors are averaged equally within independent site groups and then across groups; failures and exclusions retain their requested denominators. Unrounded WGS84 separation accounts for footprint radii and coordinate uncertainty, including regional source support. Exact support matching includes coordinates and uncertainty, even when footprint hashes agree.

The shared channel guard preserves the original outer fold. Local pollen requires explicitly pinned target pollen channels; held-out vegetation, other target pollen and actual dependencies on held-out record owners remain forbidden. Ownership metadata is distinct from a quantitative dependency edge. The host must verify the declared channel bytes; an identity catalogue alone is not source authentication.

Conditional site/block bootstrap summaries require one fixed fitted model and a complete external evaluation cohort. They preserve paired comparisons and equal group weighting with unequal block sizes. They cannot treat overlapping cross-validation fits as independent models or replace predictive observation intervals. Synthetic policy examples demonstrate these controls without claiming reviewed vegetation references or empirical reconstruction accuracy.

The prepare-lra-experiment command also dispatches the synthetic-spatial-block-fitted-v1 profile through the same governed lifecycle. Its operational exit status preserves complete, partial and interrupted accounting. score-lra-experiment verifies the saved prediction identity before opening outcome labels and reports vegetation and pollen as separate score rows. These synthetic experiments do not establish empirical accuracy.

Saved baseline and fitted experiments are reopened against the complete declared site catalogue, original folds, template, and retained batch request bodies. Source hashes and PPE declarations must agree with consumed training requests and the target forward calculation. Changed metadata cannot acquire scientific authority merely by receiving a new file hash. Valid historical products retain their recorded fields; omitted historical declarations are not manufactured. The separate analysis.lra_validation.temporal.persistence API implements the synthetic forward-time persistence profile. It reconstructs admitted older training sites, fits their spatial model, and explicitly transports the predicted vegetation unchanged to one younger held-out target before forward pollen calculation. No target pollen or younger observations enter generation. This is a declared persistence assumption, not a fitted model of vegetation change.

The existing prepare-lra-experiment command dispatches the temporal profile, and score-lra-experiment authenticates its retained prediction semantics before opening the outcome inventory. Exit codes preserve complete, refused and interrupted accounting; temporal scores keep both target denominators.

Temporal products retain full chronology/geometry admission, older fitted state, transport assumptions and original batch requests. Reading joins request bodies, source references, ordered items, batch summaries and fitted identities. Missing pre-attempt or interrupted receipts retain their actual failure state. A later post-batch interruption remains distinct from an interrupted numerical batch.

Separate scoring checks the younger observation support and reports vegetation and pollen errors independently, including failed-target denominators. Constant synthetic vegetation is reproduced in the demonstrated profile; changed younger vegetation exposes the persistence assumption's error. These tests establish workflow behavior, not LANDCLIM/Neotoma accuracy, qualified scientific review, final acceptance thresholds or joint model/PPE/chronology uncertainty.

Retained score review

analysis.lra_validation.evaluation.api exports read_scored_experiment_evidence and audit_scored_evaluation_readiness for reopening saved synthetic baseline, fitted, local-pollen, spatial-block and forward-time experiments. Supply the independently retained manifest digests of both the prediction and score products:

from pathlib import Path

from bijux_pollenomics.analysis.lra_validation.evaluation.api import (
    audit_scored_evaluation_readiness,
    read_scored_experiment_evidence,
)

evidence = read_scored_experiment_evidence(
    Path("predictions"),
    expected_prediction_manifest_sha256=prediction_manifest_sha256,
    scores=Path("scores"),
    expected_scores_manifest_sha256=scores_manifest_sha256,
)
audit = audit_scored_evaluation_readiness((evidence,))
print(audit.canonical_audit_json)

This base-host path reads retained products without loading the numerical engine, rerunning a model or opening original outcome files. It reconciles descriptive metrics through the existing scoring functions and preserves every requested fold, method and target, including failed predictions and refused observations. Vegetation and pollen have separate accounting; local pollen remains an in-sample diagnostic. A retained source refusal without an observation body cannot be re-executed here and is labelled accordingly.

admit_final_evaluation_release is the prospective engineering boundary for a final benchmark. It requires the local-pollen-available vegetation cell and both all-target-pollen-withheld vegetation and pollen cells under each of grouped outer folds and predefined spatial blocks. Each protocol binds the complete canonical split manifest, its split kind and digest. Both protocols must use the same frozen site and alias universe, and a spatial fold cannot place one block in both training and test. The returned receipt records that no empirical evaluation was performed and that scientific acceptance was not evaluated.

For one fitted or training-mean all-target-withheld pollen prediction, freeze_authenticated_pollen_interval attaches a governed conditional observation interval before outcomes are available. The batch freeze_authenticated_pollen_rows boundary instead requires the complete grouped method/fold ledger. Both reopen caller-pinned prediction and uncertainty products, check planned uncertainty units and effort declarations, and require observation-process-only propagation with exact taxa, model and training-input identities. The frozen prediction binds the uncertainty-product manifest and governed draw. This verifies lineage and conditional interval semantics; it does not authenticate historical preregistration timing, establish empirical coverage or grant scientific acceptance.

The immutable evidence binds its projections to retained product identities. Caller-pinned product bytes establish retained consistency; they do not independently authenticate external raw observations. The readiness result always states final_evaluation_performed=false and scientific_acceptance=not_evaluated. Missing historical access timestamps, the remaining prospective product/release joins, reviewed reference support, thresholds and complete interval-bearing benchmark products remain explicit gaps. A release-matrix receipt alone cannot retroactively freeze an experiment or authenticate its source observations. Record-specific local-pollen sources and overlapping cross-validation fits do not become an independent fixed-model bootstrap cohort.

Saved-product reporting

bijux_pollenomics.reporting.lra_products.api owns the bridge from verified saved products to exploratory reports. It uses the host's existing static atlas writer, map renderer and shared exclusive-output publisher. Reporting never invokes the numerical engine. Display geometry is a separate, hash-pinned input bound to the product and site identities; reconstruction ring coordinates do not become map locations or geographic footprints.

The default mapped profile admits synthetic paired reconstructions with one nonnegative, typed BP bin and explicit point anchors. Unsupported geography, profiles or temporal support remain in the tables and refusal accounting. Regional and local model anchors carry textual R/L cues. These are model context, not measured vegetation, propagation edges or accepted scientific reconstructions.

For a synthetic paired request whose samples explicitly declare calendar_bp, set temporal_profile="calendar-bp1950-signed.v1" in render_lra_report, or use the corresponding report command option:

bijux-pollenomics render-lra-report product \
  --expected-manifest-sha256 "$product_manifest_sha256" \
  --geometry display-geometry.json \
  --expected-geometry-sha256 "$geometry_sha256" \
  --temporal-profile calendar-bp1950-signed.v1 \
  --output signed-report

This opt-in preserves negative and fractional calendar BP values, with 1950 as zero. It requires one sample per role, one shared bin and one contained source support component per sample. The closed source-age support and closed declared stationarity envelope remain separate in the map and downloads. Their endpoints are finite numbers within [-(2**53-1), 2**53-1]; Boolean values, numeric strings, partial missing intervals and unsupported chronology refuse. The total display span must also fit within 2**53-1 years. BP 0 is an instant, not a substitute for a missing age. The underlying numerical product is retained unchanged. If both sites lack admissible display geometry, the report retains their rows and refusal accounting and produces no map.

Retained joint-realization products use a separate lossless presentation facade, render_joint_lra_report, and installed command:

bijux-pollenomics render-joint-lra-report joint-product \
  --expected-manifest-sha256 "$joint_manifest_sha256" \
  --output joint-report

This path performs no numerical recomputation and accepts no display geometry. It copies the externally pinned product, preserves exact rational law mass and all draw/bin outcomes, and emits complete-bin means plus the full ordered covariance. Incomplete and interrupted bins retain null moments and refused/not-run mass. The result is exploratory, has no calibrated-interval or empirical-coverage claim, and is not publication authorization or scientific acceptance.

Signed reports use explicit younger/older endpoint controls. Fractional and point selections retain the same endpoints through filtering, URL restoration, download and native browser capture. This profile does not interpret HYDE calendar labels, convert native CE/BCE or radiocarbon ages, or admit calibrated chronology. Legacy source propagation and chronology-media profiles retain their existing scope; signed native captures cannot be used as legacy media evidence.

One native selection controls roles, sites, taxa and bins across the table, anchors, URL and selected download. Shared time and layer controls also constrain mapped rows. Rows refused geographic display remain inspectable under the native selection, with that temporal-refusal policy declared in the export. Original CSV decimal text and missing values retain their source locators and units.

The native view has an independent digest, avoiding a cycle with the enclosing report manifest. URLs bind that digest and a validated selection; stale or malformed selections visibly refuse export until corrected. Selected downloads declare their JSON serialization and hash the exact emitted UTF-8 bytes. Neither a matching digest nor an interactive report establishes ecological accuracy.

Supplied joint realizations

analysis.lra.api exports JointLRAPlan, run_joint_lra_ensemble and read_joint_lra_ensemble. The synthetic-explicit-joint-analytical-v1 profile computes a finite law supplied by the caller. It does not estimate a prior, posterior or chronology distribution from source intervals.

bijux-pollenomics run-joint-lra joint-plan.json \
  --input-root inputs --output joint-ensemble --cache-root child-cache

The lra-joint-realizations.v1 plan pins an admitted synthetic LRA/series request and a declaration of the joint law. Every atom supplies a positive rational mass, one shared PPE vector, integer counts for every original sample with its original total, and a complete calendar BP age realization. Atom masses must sum exactly to one. Component identities bind their values; the workflow never independently redraws components or creates their Cartesian product.

Explicit nonoverlapping bins allocate modeled points using younger_bp <= age_bp < older_bp. Each point must also lie in its original closed source support, including any disjoint components. Original evidence, counts and age supports remain separate from the modeled transformations. Every draw/bin requires exactly one regional and one local sample; missing or ambiguous pairs remain refused. The first profile admits synthetic calendar_bp only.

Each eligible pair uses the existing deterministic analytical workflow with the same realized PPE in both stages and a coupled regional result. Source rounding and signed analytical LOVE estimates remain visible. Complete bins expose finite mass-weighted population summaries and regional/local covariance. These are not Monte Carlo standard errors, posterior intervals or measured ecological accuracy. Any failed mass prevents a complete summary; failed atoms are never removed and renormalized. Predictive pollen observations require their own forward/count layer.

The bundle retains every requested draw/bin and all child products. Reopening requires the retained manifest digest and checks the parent-to-child request and byte inventories without running the engine. CLI exit codes are zero for complete computation, two for incomplete/refused work and 130 for interruption. None grants scientific acceptance. Geometry, fall speed, taxonomy, parameter-domain selection, spatial fitting and source/calibration authenticity remain outside this profile.