[CEFR] Add leakage-safe longitudinal language-profile and drift analysis after invariance gates

[seonghobae]

Temporal measurement gap

A single CEFR-aligned assessment result can support placement, but repeated results cannot be interpreted as language growth until form, task, rater, scoring, mode, population, and language-profile comparability have been evaluated. TEPP is the correct owner for temporal, multilevel, relational, and multiple-membership analysis of repeated language profiles.

This issue owns longitudinal and contextual language-profile analysis artifacts. It does not own assessment sessions, raw responses, task content, CEFR descriptor prose, cross-sectional scoring, cut scores, or LMS decisions.

Dependencies

Blocked by:

1. release of cwl_cefr_language_assessment/v1 from ContextualWisdomLab/learning-interoperability-contracts PR feat(temporal): add typed six-clock values and uncertain intervals #5;
2. immutable CEFR profile results from Psychometrics Commons issue feat(api): enumerate project histories via loopback collection GET #424;
3. Rust scoring/linking/DIF/classification evidence from fast-mlsirm issue #1484;
4. TEPP temporal/event foundation and persistence contracts being implemented on protected main.

Do not build this as a Python post-hoc regression over copied result JSON.

Analysis boundary

Consume only versioned, immutable result references and approved contextual membership references. Preserve at least:

result_reference
participant_reference
instrument_release_reference
assessment_blueprint_reference
scoring_profile_reference
cut_score_revision_reference
claim_status_code
target_language
domain_result_reference
observed_at
published_at
available_at
recorded_at
correction_or_supersession_reference

Raw responses, audio, transcripts, task content, rater payloads, and PII remain in their owning systems.

Multi-clock and leakage contract

For every analysis cutoff T, require:

available_at(result) <= T

Separate:

• assessment event/administration time;
• response observation time;
• result publication time;
• system ingestion/recorded time;
• correction/supersession time;
• analysis knowledge cutoff.

A later correction, rescoring run, cut-score revision, or RLD update must not leak into an earlier historical analysis.

Measurement structure

Evaluate, rather than assume:

• longitudinal configural/metric/scalar or appropriate approximate invariance;
• form, item/task, rubric, rater, AI-rater family/version, mode, and accommodation drift;
• cut-score and linking drift;
• population and language-profile drift;
• within-person versus between-person separation;
• irregular assessment intervals;
• cross-classified context: cohort, instructor, sponsor organization, course offering, content release, project, and learning environment;
• multiple membership when one learner/occasion belongs to multiple relevant contexts;
• missing-by-design domains and selective retesting;
• correction/supersession without historical mutation.

Do not treat ordinal CEFR labels as equally spaced numeric observations. Use the underlying versioned domain-score/posterior evidence or appropriate latent-state measurement model.

Rust arithmetic ownership

All result-affecting temporal and multilevel arithmetic is Rust-owned:

• longitudinal state transition;
• random intercept/slope and context effects;
• continuous-time or interval-aware transition where adopted;
• invariance/drift tests and uncertainty;
• change-point/onset detection where scientifically supported;
• trajectory estimation and prediction intervals;
• true-state simulation and recovery;
• CPU multithreading and GPU execution where benchmark evidence justifies it.

Python may validate contracts, orchestrate jobs, and render reports only.

First bounded vertical

Implement a research-only English A1–B2 repeated-placement analysis with:

• four domains: reading, listening, written production, spoken production;
• at least three occasions per simulated participant;
• irregular time gaps;
• two form releases;
• human and AI-rater-version drift in productive domains;
• course/cohort/instructor cross-classification;
• one multiple-membership sponsor/context dimension;
• delayed result availability and one later correction;
• explicit no-change verdict when invariance or linking evidence fails.

Outputs:

language_trajectory_artifact
language_invariance_artifact
language_drift_artifact
language_context_effect_artifact
language_change_interpretation_artifact

Every artifact pins exact source result identities, model/engine versions, data cutoff, membership revision, failure denominator, uncertainty, and limitations.

Recovery evidence

The simulator must inject known truth for:

• stable person differences and within-person change;
• domain correlations;
• form/linking shifts;
• item/task/rater drift;
• occasion effects;
• context random effects;
• multiple-membership weights;
• irregular lags;
• missingness and delayed availability;
• correction events.

Report bias/RMSE, interval coverage, state/transition recovery, drift-detection precision/recall, false change declarations, backend parity, worker-count determinism, failure denominator, and Monte Carlo uncertainty.

Acceptance

• No trajectory or growth claim is emitted when required invariance/linking evidence fails.
• No future-available result or correction enters an earlier cutoff.
• Ordinal CEFR labels are never averaged or treated as an interval scale.
• Within-person change is separated from stable between-person/context differences.
• Cross-classified and multiple-membership design is explicit; no primary_group shortcut.
• Historical results and trajectories are immutable; correction creates superseding artifacts.
• Every number is reproducible from exact input and model artifacts.
• Production statement/branch coverage and public API docstrings are 100%.
• True-state recovery, CPU/GPU parity, memory/throughput evidence, and no skipped required GPU/recovery tests.
• Visual outputs include exact-value accessible tables, uncertainty, provenance, data cutoff, and no causal language unless a separate causal design supports it.
• No Council of Europe endorsement, validation, certification, or logo-use claim.

Standards and research baseline

• CEFR Companion Volume (2020).
• Council of Europe examination-linking and test-development manuals.
• AERA, APA, & NCME Standards (2014).
• Longitudinal measurement-invariance, DSEM/continuous-time, multilevel, cross-classified, and multiple-membership primary literature documented in APA 7th form.

TEPP may measure change in a governed language-profile system; it must not turn repeated labels into growth by assumption.