Changes in version 0.12.0.9000 Development - Development resumes after the cross-platform CI-certified stable v0.12.0 release. - Stable release provenance remains anchored at commit 4d272f025e70207f31fb2baf10cba31ee8bdb996; the v0.12.0 tag and published release assets are unchanged. CRAN publication status — 2026-09-28 - CRAN release: eyeprocess 0.11.1 was published on CRAN on 28 September 2026. - Canonical CRAN package page: https://CRAN.R-project.org/package=eyeprocess. - CRAN package DOI: https://doi.org/10.32614/CRAN.package.eyeprocess. - Release channels remain explicit: CRAN is currently 0.11.1; the newer stable GitHub release is v0.12.0; master is the 0.12.0.9000 development line. Changes in version 0.12.0 Release validation status - Release status: v0.12.0 is the stable GitHub release. It was promoted from pre-release after the exact tagged commit passed the complete cross-platform GitHub R-CMD-check matrix. - Local Windows validation — 2026-09-19: installed-package targeted tests PASS; strict measurement-intelligence lane PASS; full local test suite PASS; clean survival-vignette rebuild PASS; pkgdown local build PASS; final package installation PASS. - R CMD check: Windows R 4.6.1, --as-cran, 0 errors / 0 warnings / 0 notes. - Cross-platform certification — 2026-09-19: macOS release, Windows release, Ubuntu release, Ubuntu devel, and Ubuntu oldrel-1 all passed on commit 4d272f025e70207f31fb2baf10cba31ee8bdb996. The pkgdown push workflow, pkgdown release workflow, and GitHub Pages deployment also passed. - Release provenance: tag v0.12.0 remains permanently anchored to commit 4d272f025e70207f31fb2baf10cba31ee8bdb996; the validated release source artifact and SHA-256 remain unchanged. Event detector multiverse and inference robustness - Adds explicit detector specifications, deterministic detector branches, temporal event matching/agreement, detector-to-AOI/feature propagation, identical-model inference, stability summaries, diagnostic plots, and reproducible reporting. - Uses the full declared detector multiverse as the convergence denominator; failed, missing-term, unavailable-backend, and non-converged branches remain in the scientific record. Duplicate coefficient terms from custom callbacks are rejected. - Adds per-detector input_audit accounting for propagated rows, target-AOI rows, quality exclusions, non-finite outcomes, model rows used, and model status without recoding missing outcomes to zero. - Adds the main methodology vignette, failure-clinic vignette and executable example, reporting/interpretation guidance, API documentation, pkgdown discovery, and cross-language contract coverage. - Adds a runnable pkgdown visual-reporting vignette using built-in detector branches only, with event-count, temporal-agreement, AOI-feature, coefficient-stability, model-input-accountability, publication-sequence, interpretation-boundary, and API-handoff sections. Censored gaze-latency survival examples and site guidance - Adds a standalone evidence-verification example for time to first source/evidence AOI entry, including right-censoring construction, clustered Cox, Weibull/log-normal AFT sensitivity models, diagnostics, interpretation, and manuscript reporting guidance. - Restores the survival methodology articles and the complete survival API group to pkgdown navigation so later site additions do not hide the survival workflow. - Adds a troubleshooting clinic covering incomplete windows, time-zero events, sparse events, flagged proportional-hazards diagnostics, low-quality gaze, and competing-event boundaries without changing the public statistical API. - Adds a survival reproducibility/preregistration checklist and a site-contract regression test that fails if survival articles, API references, or required worked-example files disappear from the source tree. - Expands the plot-rich survival website guidance with Kaplan-Meier, single-event 1-KM, and empirical hazard/risk-set visualizations plus interpretation boundaries. Standardized spatial data quality - Adds a dedicated Data Quality plot gallery and reporting clinic covering accuracy, RMS-S2S, BCEA, sampling intervals, the four-panel dashboard, review-rule sensitivity, limitations, and manuscript-ready reporting. - Repairs public export/governance for simulate_gaze_survival_inputs() and restores missing multilevel-mediation Rd coverage so package/vignette and pkgdown builds can execute the documented workflows. - Adds vendor-neutral target-referenced accuracy, RMS sample-to-sample and population-SD precision, BCEA with explicit probability, realized sampling behavior, and reason-aware data loss. - Adds a canonical gaze_quality_report with metric-specific units, provenance, study-defined review thresholds, insufficiency/timestamp flags, and an explicit no-automatic-exclusion contract. - Adds deterministic six-profile 9-point validation data, frozen R/Python numerical fixtures, sensitivity-analysis guidance, manuscript-reporting guidance, and dedicated pkgdown navigation. - RMS-S2S never bridges missing samples; mixed units fail; malformed thresholds fail; missing grouping identifiers are retained and audited. - Long observed intervals are distinguished from the estimated number of nominal samples represented by those gaps. AOI perturbation and uncertainty analysis - Adds vendor-neutral AOI geometry validation, reproducible dilation/erosion/translation/jitter and anisotropic perturbation specifications, and explicit pixel/degree-of-visual-angle conversion. - Adds sample/fixation reassignment matrices, participant/trial/AOI stability summaries, feature recomputation, empirical assignment frequencies, and optional fixed model-callback propagation with retained failure and convergence evidence. - Adds rectangle and convex-polygon support, explicit overlap ambiguity, screen-boundary policies, synthetic cross-language parity fixtures, base-R robustness plots, and a runnable advertising/interface vignette. - Robustness frequencies remain descriptive sensitivity summaries for the declared perturbation set and are not probabilities that a scientific conclusion is true. CRAN portability hotfix - Make default Arrow/Parquet compression portable across Arrow builds: prefer zstd, fall back to snappy, then uncompressed when the default codec is unavailable. - Preserve explicit compression requests: an explicitly requested unavailable Arrow codec fails with a clear error instead of being silently substituted. - Replace repository-relative README links with absolute GitHub URLs. - Add a package spelling allowlist for legitimate eye-tracking, biometric, and storage terminology flagged by CRAN incoming checks. Changes in version 0.11.0 M4 trait-conditioned latent response-process states - Added M4 as a sequential extension of the frozen M3 response + RT + gaze + pupil measurement backbone, with marginalized latent response-process states over explicitly ordered sequences. - Added K=1 as a formal null, K=2 as the conservative reference, RT-anchored state-label identification, posterior state probabilities/entropy, trait-conditioned initial/transition probabilities, and state-dependent RT/gaze/pupil deviations while keeping the scored-response Rasch equation state-independent. - Added structural/posterior identifiability auditing, state diagnostics, PPC, focused M3-to-M4 ablation and response-target process-information evidence, six negative controls, K/modelling sensitivity, deterministic state-aligned recovery, and multi-domain validation statuses. - Added consolidated S3 plots, twelve versioned examples, six M4 articles, and an explicit computational boundary: recovery/sensitivity/controls are design-only unless analysts opt into backend fitting. - M4 state labels remain neutral statistical response-process states. State detection is not construct validation and does not automatically identify strategy, attention, engagement, effort, cognitive load, emotion, guessing, or misconduct. Changes in version 0.10.0.9000 M3 response + RT + gaze + pupil measurement evidence - Added the M3 four-channel response + RT + gaze + pupil reference likelihood with correlated person/item structures and explicit pupil measurement nuisance terms for baseline, luminance, gaze position, quality, blink status, interpolation status, and time-on-task. - Added retained-truth simulation with informative, weak, null, redundant, and confound-only pupil scenarios plus MCAR/quality/gaze/ability/device pupil-dropout stress mechanisms. - Added the full eight-model response-anchored channel-ablation lattice, response-target PSIS-LOO/ability-uncertainty process-information evidence, non-additivity/redundancy contrasts, channel-conflict diagnostics, and sensor value-of-information screens. - Added pupil-specific negative controls, an explicit functional-pupil score bridge to the existing trajectory/deconvolution machinery, recovery/coverage tooling, M3 PPC, consolidated S3 plots, nine examples, and six articles. - Pupil remains a neutral measurement channel: M3 does not equate pupil dilation with cognitive load, effort, attention, or arousal; device equivalence and nonignorable missingness remain validation questions rather than defaults. M2 three-way response + RT + gaze evidence - Added a likelihood-faithful M2 reference implementation for Rasch responses, lognormal response times, and negative-binomial gaze-fixation counts with correlated person-side and item-side parameter structures. - Added compatible M0/M1 Stan reference models for response-target channel ablation. - Added M2 simulation with retained truth, structural/data identifiability audit, posterior predictive W/L/M checks, response-target PSIS-LOO and ability-variance process-information summaries, negative controls, and repeated estimator recovery. - Added consolidated M2 plot methods, six examples, four development articles, and explicit no-fallback CmdStan gating. - M2 remains experimental/evidence-gated: synthetic recovery and model diagnostics do not establish construct validity, empirical generalizability, or causal interpretation of process channels. - Starts the Unified Multimodal Latent Process-IRT development programme with a canonical person-item-trial measurement contract, deterministic four-channel simulation, posterior process-information metrics, channel ablation, explicit backend gating, validation objects, native plots, development Stan programmes, examples, tests, and articles. - M2/M3 CmdStan estimators remain experimental/gated until recovery, SBC, misspecification, identifiability, and empirical-reproduction evidence is frozen. Availability is not treated as construct validity. Changes in version 0.9.0.9000 Executed validation evidence and governed IRT — Milestone #2 - Adds deterministic validation-plan, acceptance-rule, MCSE, stress, reliability, negative-control, evidence-freeze, and evidence-atlas infrastructure. - Adds transparent IRT probability, information, scoring, residual-fit, local-dependence, targeting, classification-precision, missing-by-design, and prior-sensitivity utilities. - Adds multidimensional/testlet/latent-regression design support, scale linking/equating diagnostics, DIF/DTF effect summaries, CDM/Q-matrix utilities, adaptive-testing governance, and response-time/process-aware IRT contracts. - Adds exact optional-engine adapters for mirt, TAM, GDINA, LNIRT, eRm, equateIRT, catR, and mirtCAT; unavailable engines are explicitly gated and never replaced by simpler estimators. - Adds 11 articles, deterministic synthetic fixtures, validation-execution scripts, paper-ready evidence tables, and 12 standalone examples. - All Milestone #2 validation outputs are software-validation evidence. They do not establish construct validity of gaze, pupil, response-time, sequence, or psychometric measures. - Closes the 0.9 public-API lifecycle registry across all 1,139 current exports using a curated module policy and explicit experimental/gated overrides; future unregistered exports remain unreviewed by design. Empirical validation and recovery - Adds explicit process-validation designs, known-truth simulation fixtures, recovery/bias/RMSE/coverage/convergence summaries, Monte Carlo error diagnostics, frozen validation references, and validation evidence matrices. - Adds simulation-based calibration rank diagnostics and interval-coverage calibration curves for Bayesian/process estimators where posterior draws are available. Governed workflows - Adds explicit analysis specifications and dependency-aware pipelines with manifests, resumable execution, audits, reports, Graphviz/Mermaid views, and targets interoperability templates. - Substantive preprocessing/model choices remain user-declared; the pipeline layer does not silently select methods. API consolidation - Adds API inventory, lifecycle registry, canonical-interface mapping, replacement audits, surface summaries, and migration-oriented lifecycle diffs. - Existing interfaces are not aggressively removed in 0.9; unclassified APIs remain explicitly unreviewed. Measurement quality and reliability - Adds a guarded process-measure registry, split-half/test-retest reliability, ICC(A,1), Bland-Altman repeatability, temporal stability, and bootstrap reliability. - Adds empirical calibration-error models, RMS successive-sample gaze precision, effective sampling frequency, sampling-irregularity audits, data-loss profiles, calibration drift, uncertainty propagation, probabilistic AOIs, and analysis-resolution compatibility checks. - Adds pupil preprocessing-order and baseline-window sensitivity audits. Sensitivity and negative controls - Adds defensible multiverse grids, specification curves, decision stability, decision leverage, fragility/rank diagnostics, and method-comparison wrappers. - Adds temporal feature provenance, leakage sentinels, outcome-blind feature construction audits, placebo-window summaries, permutation/shift negative controls, and null benchmarks. Reproducibility and evidence - Adds analysis/session manifests, reproducibility fingerprints, file hash manifests, lightweight PROV-oriented lineage graphs, RO-Crate 1.3 metadata export, and provenance DOT export. - Adds software-paper claim matrices, evidence bundles, descriptive readiness/gap audits, evidence freezing, and manuscript reproducibility manifests. Computational validation - Adds benchmark designs, runtime/memory summaries, synthetic corruption plans, robustness stress tests, and stress-tolerance frontiers. Scientific governance - Process features remain measurement channels rather than automatic mental-state or diagnostic labels. - Data-quality thresholds, resolution guards, and stability thresholds are explicit workflow rules rather than universal validity cutoffs. - Negative-control/leakage flags identify analysis properties and are not misconduct labels. - Exact gated frontier methods remain gated; 0.9 does not substitute easier estimators under existing frontier function names. Changes in version 0.8.0 Process measurement and deployment governance - Added biometric/process pre-flight governance and explicit exclusion manifests. - Added deployment-drift auditing across device, site, vendor, and stimulus version. - Added temporal process windows, AOI trajectory features, growth models, and window-sensitivity audits. - Added advanced pupil frequency, activity, event-deconvolution, confound, fatigue, and filtering workflows. - Added visual-context/testlet IRT and multiblock process representations. - Added process-profile mixtures, external-validity tools, streaming scoring, and validation bundles. - Added item-parameter seeding and candidate-bank auditing. - Added nonclinical presentation/accessibility sensitivity workflows. - Added process-decision proxies and structured/unstructured process-feature comparisons. - Added advanced Rasch, mixture, imputation, Bayesian, and gaze-aware 3PL diagnostic infrastructure. - Added explicit gated frontier estimators where an exact validated implementation is not available. - Added 11 new pkgdown articles, focused 0.8 tests, reference documentation, and validation registries. Validation - Focused 0.8 tests: PASS. - Complete eyeprocess test suite: PASS. - pkgdown::check_pkgdown(): PASS. - R CMD check: 0 errors, 0 warnings, 0 notes. - Installation with vignettes and installed-package smoke validation: PASS. - GitHub PR #9: all 8 CI checks passed before merge. Scientific governance - Process evidence is not automatically interpreted as a mental-state, clinical, ability, or misconduct diagnosis. - Frontier estimators remain explicitly gated rather than silently substituting simpler estimators. - Experimental and custom model specifications retain conservative maturity labels. Changes in version 0.7.0.9000 Process-IRT validation framework - Added the process-IRT validation framework and associated validation contracts. - Expanded advanced process-model validation, diagnostics, and evidence infrastructure. - Preserved explicit experimental and validation-gated maturity boundaries for advanced estimators. - Added focused validation coverage and pkgdown documentation for the 0.7 programme. Changes in version 0.6.0 Measurement uncertainty and AOI evidence - Added probabilistic AOI assignment, fuzzy transition summaries, compositional AOI analysis, and explicit measurement-uncertainty propagation and budgets. - Added calibration-drift detection, offline recalibration, process reliability and Generalizability Theory, and cross-device linking/equivalence audits. Temporal, spatial, and pupil-process science - Added pupil phase-amplitude registration, informative-missingness and MNAR sensitivity, recurrence and cross-recurrence analysis, experimental fixation point-process models, representative scanpaths, and cognitive-episode segmentation. Decision intelligence and provenance - Added multi-objective item-bank optimization, dynamic process-DIF and fairness drift, conditional process centiles, and auditable evidence/decision provenance graphs. - Added a shared S3 plotting framework, 18 targeted test modules, eight articles, CI, manual documentation, and a standalone validation programme. Changes in version 0.5.0.9000 Research-scale validation and reproducible orchestration - Added deterministic Monte Carlo job plans, atomic checkpoints, resumption, parallel execution, manifests, recovery/calibration/failure summaries, plots, reports, and model-promotion audits. - Added manual research-validation workflows and chunk-based execution scripts; full simulations remain opt-in and are not run during ordinary package checks. Advanced model hardening - Hardened dynamic IRTree models with transition designs, structural zeros, irregular time, uncertain states, multinomial and optional Stan hidden-state engines, decoding, residuals, comparison, and recovery programmes. - Added functional pupil–IRT, theory-constrained strategy-mixture, and Wiener gaze-diffusion engines with explicit scientific-evidence gates. Vendor evidence, contracts, storage, and reproducibility - Added independent multi-vendor corpus registration, fingerprinting, redaction, semantic comparison, round-trip loss audits, and declared/fixture/empirical compatibility matrices. - Added stable API/object contracts, atomic partitioned Arrow/Parquet/CSV/RDS storage, schema migration, corruption detection, optional-engine adapters, and a fully synthetic multimodal public benchmark. Changes in version 0.4.0.9000 - Added Eye-Tracking-BIDS import/export and optional RDS/Parquet/Arrow storage. - Added interoperability contracts for eye-tracking, sequence, GDINA, diffIRT, and OpenMx ecosystems. - Added multi-vendor empirical-validation audits and an executable validation-release programme. - Added parameter-recovery, interval-coverage, simulation-based calibration, advanced-model evidence gates, engine-equivalence, empirical-reproduction, grouped-validation, leakage, multiverse, benchmarking, reporting, and public-benchmark infrastructure. - Added dynamic gaze-state IRTree, functional pupil-IRT, theory-defined strategy-IRT, and gaze-informed diffusion model families with explicit experimental safeguards. - Added tests, plots, templates, scripts, and four research vignettes. Changes in version 0.3.0.9001 Changes in version 0.3.0.9000 - Added run_gazepoint_workflow() as the complete, reproducible downstream workflow for real Gazepoint folders. - Added media-run reconstruction into explicit person-by-item-by-trial intervals with optional study-specific item and condition maps. - Added integrated QC evidence for file pairing, canonical validation, timebases, coordinate spaces, sampling rates, gaze and pupil validity, trial coverage, episodes, and multimodal clock overlap. - Added pupil interpolation/filtering and blink derivation with baseline correction disabled by default unless explicitly declared. - Added valid-only biometric analysis values and trial-level EDA, heart-rate, IBI, dial, pupil, gaze, fixation, and AOI summaries while preserving native data. - Added one-row-per-person-item-trial process tables, AOI/fixation/pupil/biometric tables, feature dictionaries, response templates, and conditional response/response-time matrices. - Added a complete plot suite, canonical export, source fingerprint, workflow specification, rerun script, session information, Markdown/HTML report, and workflow-result object. - Added workflow validation, documentation, a dedicated vignette, Windows runner script, and regression tests with and without observed responses. - The workflow never fabricates responses, scores, IRT estimates, or psychological interpretations. Changes in version 0.2.0.9002 - Renamed the packaged Gazepoint Analysis 7.2.0 regression fixtures to portable, space-free filenames. - Updated all fixture references while preserving filename-based inference of the canonical participant identity User 3. - Added regression coverage requiring portable packaged fixture paths. - No importer, canonical schema, biometric mapping, or empirical-validation behavior changed. Changes in version 0.2.0.9001 - Preserved the successful real Gazepoint Analysis 7.2.0 corpus validation while aligning legacy synthetic tests with current channel semantics. - Treats the legacy GSR fixture field as gsr_raw; processed conductance remains represented by GSR_US/EDA columns. - Added an explicit recording_id argument to folder-level Gazepoint import and forwards it safely for single-recording folders. - Rejects a single recording_id override for multi-recording folders rather than creating duplicate canonical identifiers. - Added regression coverage for direct and generic folder imports with recording-ID overrides. Changes in version 0.2.0.9000 - Added empirical support for Gazepoint Analysis 7.2.0 User *_all_gaze.csv, User *_fixations.csv, and multi-section Data_Summary_export_*.csv files. - Added filename-based participant and recording identity inference for paired Gazepoint exports whose USER field is blank. - Added TIMETICK(f=...) normalization to zero-based recording seconds while retaining native ticks, media-relative time, source media identifiers, AOI labels, saccade measures, and video-frame values. - Namespaced vendor fixation identifiers by media because Gazepoint restarts FPOGID for each media item. - Added explicit Gazepoint biometric channels and validity handling for raw GSR, conductance in microsiemens, tonic/phasic components, heart rate, IBI, and engagement dial values. - Added read_gazepoint_summary() and conversion of Gazepoint Data Summary reports into AOI definitions and participant-AOI feature records. - Added paired-folder import, real-structure fixtures, regression tests, and a private six-recording empirical validation corpus. Changes in version 0.1.0.9003 - Updated the validation-corpus regression test to match the intentional, idempotent behavior of init_validation_corpus(). Repeated initialization now verifies preservation of the manifest, README, and existing case files rather than expecting an obsolete non-empty-directory error. - No package implementation or analysis behavior changed in this patch. Changes in version 0.1.0.9002 - Removed repeated write.csv() qmethod warnings from canonical-folder serialization on Windows. - Added explicit canonical units for imported biometric channels, including heart rate, interbeat intervals, electrodermal activity, and engagement-dial streams; unknown channels retain the explicit unit vendor_units. - Made init_validation_corpus() idempotent and non-destructive. Re-running it now preserves existing manifests and case files unless overwrite = TRUE. - Added regression tests for warning-free canonical serialization, populated biometric units, and safe repeated corpus initialization. Changes in version 0.1.0.9001 - Corrected binocular eye-sample validation to use the composite key recording_id + sample_id + eye, eliminating false duplicate-key errors. - Added versioned canonical-folder serialization with an explicit missing-value token, preserving the distinction between NA and genuine empty text while retaining compatibility with folders written by earlier development builds. - Prevented Pupil Labs and EyeLink detectors from probing proprietary binary files as delimited text and suppressed detector-level parsing warnings. - Made validate_real_exports.R exit cleanly with instructions when an initialized corpus has no cases, rather than raising an error. - Added regression tests for binocular keys, missing-versus-empty round trips, and warning-free expected rejection of SMI IDF input. - Added reproducible empirical export validation for individual files, folders, and initialized, versioned multi-vendor corpora. - Added declared-versus-fixture-versus-empirical compatibility profiles. - Added source manifests with file hashes, field inventories, delimiter and format-detection evidence. - Added canonical schema-coverage and source-preservation audits. - Added deterministic dataset fingerprints, canonical round-trip comparison, and table-level compatibility evidence. - Added linked-identifier anonymization and shareable validation bundles that exclude raw vendor data, source paths, original filenames, and file hashes by default; human disclosure review remains mandatory. - Added dedicated source validators for Tobii, Pupil Labs, EyeLink, SMI, and generic mapped exports, so all built-in adapters expose a validation hook. - Added unique case identifiers, manifest-relative path resolution, exact format-family evidence assignment, a real-export validation vignette, manifest template, plots, reports, and regression tests. Changes in version 0.0.0.9004 - Added eyeprocess-package to a dedicated Package overview section in the pkgdown reference index, resolving the final pkgdown::check_pkgdown() configuration error. - No analytical, import, schema, plotting, or modelling behaviour changed. Changes in version 0.0.0.9003 - Fixed the Windows installer version check by removing read.dcf() field-name attributes and comparing normalized character scalars. This eliminates a false mismatch when the installed and source versions are identical. - Added the canonical pkgdown site URL to DESCRIPTION while retaining the GitHub repository URL, satisfying pkgdown::check_pkgdown(). - Retained all 0.0.0.9002 package-check and runtime fixes. Changes in version 0.0.0.9002 - Removed stale hello() scaffold artefacts during local installation and validation. - Replaced the unexported brms::gaussian() reference with stats::gaussian(). - Removed non-standard evaluation from stream construction and Gazepoint biometric matching. - Excluded the repository-level CITATION.cff from built source packages while retaining inst/CITATION. - Hardened the Windows installer against reinstalling a loaded namespace and against false version-success messages. Changes in version 0.0.0.9001 - Fixed type-preserving row binding when canonical tables contain zero rows. - Added stable identifiers to the interpretive-warning registry. - Removed all-missing extrema warnings from fixation and saccade preprocessing. Changes in version 0.0.0.9000 - Introduces the canonical eye_dataset relational object. - Adds generic CSV/TSV mapping and an extensible adapter registry. - Adds first-class Gazepoint Analysis and Gazepoint Biometrics importers. - Adds dedicated Tobii Pro Lab, Pupil Labs Neon/Core, EyeLink ASC/report, and SMI BeGaze text adapters. - Adds explicit clock and coordinate-space management. - Adds trial reconstruction, AOI registration and assignment, ocular-event derivation, pupil preprocessing, quality audits, feature extraction, plots, simulations, reporting, and optional psychometric modelling interfaces.