Longitudinal evaluation from receipts

Use plb trend to compare a fixed model's results over time without collecting prior raw datasets. Inputs are verified, already-shareable evaluation-receipt/v1 artifacts. The result is a versioned longitudinal-series/v1 artifact with source receipt digests, per-period scores, available uncertainty, practical changes, slice trajectories, and quality decisions.

plb trend jan.shareable.json apr.shareable.json aug.shareable.json --time 2026-01-01 --time 2026-04-01 --time 2026-08-01 --series local-assay-study@v1 --model published-model@weights-v2 --metric rmse --practical-threshold 0.02 --critical-slice site=lab-a --out reports/assay-trend.json

Identity and time

Declare one collection date or timezone-aware timestamp for every receipt, in chronological order. Dates mean midnight UTC; timestamps are normalized to UTC. Duplicate times or receipts fail. Source publication/file dates are not inferred as collection dates: the release boundary deliberately removes those local metadata fields.

Model ID/revision, series ID/revision, and collection times are declarations. Available signer model statements and external-validation model identities must agree; immutable model artifact digests cannot conflict. Pin trusted issuers with --trust-store and --require-trusted-signer when identity matters. A self-contained hash alone does not prove honest execution or that runs belong to the claimed study.

Every source must have a successful release decision and compatible protocol digest, task, benchmark identity, input schema, and metric basis. Changing the model or scientific protocol starts a separately identified series. --metric defaults to the protocol primary metric; unsupported metrics fail. The default reference is the first period; --reference-index selects another zero-based reference explicitly.

Changes and uncertainty

Each period records the selected score and available interval, raw and direction-adjusted changes from the reference and previous period, and relative degradation. Positive degradation always means worse performance. Relative change is unavailable when the reference is zero. Undefined scores and missing/incompatible intervals stay explicit.

A point change above --practical-threshold is marked degradation. Its separate uncertainty_assessment reports whether available marginal interval bounds are compatible with no material change. A small worsening point estimate is not automatically evidence of drift.

Receipts do not contain the joint bootstrap samples needed for a new paired or independent-cohort interval. The trend artifact instead reports the conservative difference of available marginal bounds. Two nominal 95% marginal intervals imply a union-bound coverage floor of 90% for that difference conditional on their marginal coverage; this does not require independent periods, and it is not a 95% paired interval. Bootstrap marginal coverage is approximate. Source-versus-itself comparisons are exactly zero. Perturbed metrics do not inherit clean-metric intervals.

Period/slice scans remain exploratory and unadjusted. The artifact does not claim a simultaneous confidence band, significance after scanning many periods, causal attribution, or proof of distribution drift. Use the raw-data cohort diagnostic workflow when you have both cohorts and need its explicit resampling assumptions or feature-distribution summaries.

Critical slices and source diagnostics

The artifact retains already-approved slice metrics for each period and compares slice values/prevalence with the reference. It identifies new/missing groups, ranks the largest observed degradations, and marks explicitly named critical slices. --prevalence-threshold defaults to an absolute proportion change of 0.1. Missing groups have no invented performance estimate.

This supports recurring site, batch, assay, and other configured subgroup diagnostics using aggregate receipts. When a source receipt includes a degradation-summary/v1 cohort analysis, the period records that the analysis is available in its source receipt; the source digest identifies that evidence. Cross-period feature-distribution tests or joint resampling results are not fabricated from marginal summaries.

Recurring quality decisions

Pass --quality-policy quality.yaml to apply the existing quality gate to every source receipt. Without it, an included source gate's status is retained when available. Requirements needing private evidence that was not released return investigate.

The CLI exit status follows the latest period's quality result: 1 for fail, 2 for investigate, and 0 for pass or successful construction without a declared quality gate. Earlier failures remain in the artifact and do not permanently fail a recovered model's current check. Read the reported quality status; exit zero without a policy does not establish model acceptability.

Public release

The local artifact is written by default. To export the series and its Markdown summary, approve exact series/model identities, all collection times, and any complete slice partition:

schema_version: release-policy/v1
id: reviewed-assay-series
enabled: true
require_privacy_audit: false
include_protocol: true
include_quality_gate: true
include_source_receipts: true
public_series:
  id: local-assay-study
  revision: v1
public_models:
  - id: published-model
    revision: weights-v2
public_times:
  - '2026-01-01'
  - '2026-04-01'
  - '2026-08-01'
slices:
  site: [lab-a, lab-b, lab-c]

plb trend jan.shareable.json apr.shareable.json aug.shareable.json --time 2026-01-01 --time 2026-04-01 --time 2026-08-01 --series local-assay-study@v1 --model published-model@weights-v2 --release-policy series-release.yaml --out reports/assay-trend.json
plb verify reports/assay-trend.shareable.json

The example assumes every source already passed its own reviewed release policy. Set the series policy's audit requirements to match your study; the new artifact does not reconstruct withheld audit evidence. A series policy requiring fresh privacy/design evidence remains blocked when that evidence is unavailable.

Every period must satisfy the new sample floor and any known independent-unit floor. Approved slices must cover a complete disjoint partition in each period; their observed labels can vary within the approved union. Small groups or incomplete partitions block release. Omit slices to withhold all slice trajectories and critical labels. Collection times, study membership, and source hashes can themselves be sensitive; include_source_receipts explicitly approves publishing hashes of already-shareable source receipts, never raw dataset hashes.

The exporter recomputes changes, prevalence, rankings, and interval bounds from approved numeric evidence; caller-supplied change fields and unknown metadata are discarded. Paths, receipt IDs, local policy/context, and unapproved labels are excluded. Failed reruns remove stale public JSON and Markdown outputs. The workflow stays local and requires no account, telemetry, or hosted monitoring service.