Skip to contents

Changelog

Source: NEWS.md

CausalStress 0.2.1

Quarto Documentation and Bounded Maintenance

  • Documentation: Migrated the README, package articles, and all 12 registered DGP dossiers to governed Quarto sources; added canonical typed-scoring, estimator-extension, and native-DGP contribution guides.
  • Website: Added a pinned, validated pkgdown build with one registry-keyed report per DGP and a manual, tag-only GitHub Pages publication workflow.
  • Maintenance: Removed obsolete fingerprint fixtures, a thread setter, an unused estimator-result adapter, duplicated tau forwarding, and one-use vocabulary getters while preserving public behavior and governed identities.
  • Regression safety: Added characterization coverage for the duplicated per-seed runner paths without changing their current behavior or public contracts.
  • Release tooling: Pinned Quarto CLI 1.9.38, the quarto R package 1.5.1, and pkgdown 2.2.1 across local and remote documentation gates, including full-vignette Windows, Ubuntu release/devel, and macOS package checks.

CausalStress 0.2.0

Wave 1 Typed Estimand Registry

  • Governance: Ratified Constitution v2.0.1, preserving the v2.0 estimand registry while correcting the synthetic covariate contract to uppercase, consecutive, one-based X1...Xk and adopting real-data DGP terminology; activated the Wave 1 packet for typed scoring, schema-4 identity, and staged CATE support.
  • Scoring: Added canonical ATT, ATE, QST, and CATE target descriptors, typed estimator outputs, a legacy list(att, qst, meta) shim, and the three-way requested/produced/truth scoring join. ATT, ATE, and QST are executable; CATE is registered but deliberately staged as not implemented.
  • ATE: Added finite-sample synthetic ATE scoring from runner-owned structural truth, with ordinary estimator airlock protections and a typed oracle benchmark producer.
  • Persistence: Introduced schema-4 fit, score-record, and score-row fingerprints so one fit can produce multiple score records and one QST score record can contain multiple coordinate rows, while schema 1-3 resume attempts fail closed.
  • Runtime persistence: Replaced the archived qs dependency with atomic base-R RDS staging and oracle caches, added strict current-artifact identity checks, and made legacy .qs staging fail closed without reading or modifying it.
  • Surfaces: Added the canonical long-form typed score collection surface and preserved ATT/QST helpers as compatibility projections.
  • Staging: Registered CATE as a governed target while Wave 1 deterministically hard-rejects CATE-only tasks and emits target_not_implemented rows for mixed requests.
  • Gatekeeping: Preserved ATT/QST policy and added only an uncalibrated, non-consequential ATE gate slot.
  • Release hygiene: Fixed GenGC wrapper formulas for the v0.2.0 additive-formula parser, corrected README/roxygen claims and current registry inventories, and routed CI golden-value tolerance guidance into the horizon.

CausalStress 0.1.10

Governed Repair Cycle and Release Gate

  • Governance: Added the active design authority index, contract ledger, RFC cycle, release gate, and v0.1.10 spec packet with machine-readable tickets.
  • Correctness: Fixed ATT extraction for TMLE, batch error capture/reconciliation, dependency declarations, and version-aware DGP validation.
  • Reproducibility: Introduced schema-3 fingerprints, versioned DGP pins, RNG isolation for validation/planning/oracle paths, and substrate-scoped bitwise reproducibility claims.
  • Oracle safety: Replaced blanket oracle access with column-scoped airlock grants and recorded oracle-column provenance in result metadata.
  • Statistical gates: Hardened CI-less gatekeeper handling, QST run-failure arithmetic, low-bootstrap CI semantics, and registry/sidecar validation.
  • Release hygiene: Removed stale duplicate aliases, documented exported wrappers and runner arguments, cleaned package-check notes, and recorded release-gate evidence.

CausalStress 0.1.8

Patch-Only: Reproducibility & Audit Closure

  • Integrity: Oracle truth is now cached to disk (tools::R_user_dir), avoiding redundant re-simulation in parallel runs.
  • UX: Runner enforces ci_method = "none" when bootstrap = FALSE to prevent accidental compute waste (“Silent Bootstraps”).
  • Governance: Registry now fails closed (aborts) if a DGP sidecar is missing.
  • Audit: Runner captures estimator_reported_version alongside registry versions for provenance.
  • Experimental parallel gating: parallel=TRUE now requires experimental_parallel=TRUE, emits a causalstress_experimental_parallel warning once per call, and records provenance flags.
  • Fingerprint schema v2: config_fingerprint_schema=2 added; max_runtime is now part of the config fingerprint; legacy v0.1.7 pins remain resumable (with max_runtime=Inf only).
  • Wide & Shallow in experimental parallel: forces num_threads=1 and applies thread-cap env vars (OMP/MKL/OpenBLAS/Veclib) with provenance capture.
  • DGP sidecar metadata corrected + validated against runner-side executable meta mapping (no DGP generator changes).
  • Added accessors: cs_science_payload(), cs_provenance(), cs_meta_flatten().

CausalStress 0.1.7

Scale Readiness & Protocol Hardening

  • Enforced “no parallel persistence without staging” protocol (parallel=TRUE + board requires staging_dir).
  • Deterministic campaign execution order via optional campaign_seed (no ambient sample() usage).
  • Pin metadata now includes ATT/CI summary fields for pin-meta-first resume checks.
  • Stage-and-gather persistence leaves staged files in place on write failures for retry.

CausalStress 0.1.6

Major Improvements (Hardening Phase)

  • RNG Locking: Implemented strict worker-side RNG enforcement (Constitution Article II).
  • Robust Bootstrap: Added tryCatch protection and a 90% success quality gate for CIs.
  • New Estimators: Added bart_att (BART) and tmle_att (TMLE) with soft dependency checks.
  • Validation: Added cs_validate_registry() tool; verified 12 baseline DGPs.

Oracle & Truth Generation:

  • Refactor cs_get_oracle_qst: Removed “Shadow Library” of duplicate simulators.
  • Implemented “Registry-First” truth generation: Oracle now calls cs_get_dgp().
  • Added recursion guard (cache sentinel) to prevent infinite loops (DGP->Oracle->DGP).
  • Enforced ORACLE_SEED <- 99999L to isolate truth generation from user seeds.

Registry & Metadata:

  • Created YAML sidecars (inst/dgp_meta/*.yml) for ALL DGPs (including placebos).
  • Refactored cs_dgp_registry to read tags dynamically from these YAMLs.
  • Refactored Gatekeeper to filter placebos via metadata tags, not regex.
  • Added yaml to Imports.

Runner & Code Health:

  • Extracted cs_extract_estimator_result helper to DRY up cs_run_single.
  • Unified result parsing logic for main execution and bootstrap loops.
  • Added test-integrity-capabilities.R to prevent Registry/Estimator drift.

Operational Hygiene:

  • Exported cs_gather_results for explicit parallel support.
  • Renamed dgp-synth-placebo.R -> dgp-synth-placebo-tau0.R (Constitutional fix).
  • Updated .Rbuildignore/.gitignore to exclude sidecar cache artifacts.

CausalStress 0.1.5

Scientific Safety & Oracle Refactor

  • Oracle Truth: Completely refactored cs_get_oracle_qst to strictly adhere to the “Single Source of Truth” principle. It now calls the official DGP generators from the Registry (using a reserved seed 99999L) rather than using a hardcoded “shadow library” of simulators. Implemented a recursion guard to safely handle synthetic DGPs calling the Oracle.
  • Gatekeeper Robustness: Eliminated brittle “magic string” matching (regexing “placebo” in filenames). The Gatekeeper now identifies placebo DGPs by querying strict tags (“placebo”) from the Registry.
  • Registry-Sidecar Integration: The DGP Registry now dynamically reads metadata (tags, stress profiles) from immutable YAML sidecars (inst/dgp_meta/*.yml) instead of hardcoded lists.

Operational Hygiene & API

  • Parallel Staging: Exported cs_gather_results() to the public API to officially support “Stage & Gather” parallel patterns.
  • Build Artifacts: configured .Rbuildignore and .gitignore to exclude machine-specific RMarkdown cache directories (*_cache, *_files) from the package source.
  • Constitutional Compliance: Renamed dgp-synth-placebo.R to dgp-synth-placebo-tau0.R (and associated tests) to strictly match the DGP ID.

Internal Code Health

  • DRY Runner: Refactored cs_run_single to use a unified helper cs_extract_estimator_result for parsing estimator outputs, eliminating logic duplication between the main run and the bootstrap loop.
  • Integrity Tests: Added test-integrity-capabilities.R to enforce that estimator outputs (oracle, QST support) match their Registry declarations.

CausalStress 0.1.5

  • Parallel execution: Added parallel=TRUE support to cs_run_seeds and cs_run_campaign with future/furrr, plus optional staging_dir for crash-safe runs.
  • Load balancing: Shuffled task ordering and chunked scheduling to reduce idle workers in wide campaigns.
  • Staging & resume: Stage-and-gather workflow for pins boards; improved fingerprint hashing now includes estimator config and tau to avoid stale caches.
  • Progress: Wired progressr into parallel runs for clearer feedback during campaigns.
  • Robustness: Timeout/safety checks in runners; fingerprint mismatch now surfaces clear errors with force/skip overrides.

CausalStress 0.1.4

  • GenGC Integration: Added est_gengc adapter with soft dependency.
  • Distributional Intelligence: Runner now bootstraps QST curves, computes pointwise CIs, and enriches outputs.
  • Governance: Added cs_summarise_qst and implemented the Gatekeeper 10/10 rule for QST (fails if >10% of quantiles exclude zero).
  • Visualization: Updated cs_plot_qst to support confidence bands and truth overlays (“Blue Band”). # CausalStress 0.1.3

Safety & Suites

  • Suites Layer: Introduced cs_run_suite() and the internal Suite Registry. Users can now run curated batches (e.g., suite_id = "placebo") with a single command.
  • Gatekeeper Protocol: Added cs_summarise_gatekeeper() to automate Article IV validation. It applies the 90% coverage rule to detect “unsafe” estimators.
  • Visualization: Added cs_plot_att_error() and cs_plot_placebo() for standardized ggplot2 reporting.
  • Instrumentation: Run results now capture granular timing metadata (run_time_dgp, run_time_est, run_time_total) to identify bottlenecks (e.g., bootstrap overhead).

Governance & Infrastructure

  • Article VII Compliance: Refactored the DGP registry to support immutable, versioned helpers. Added status flags (stable, experimental) and rationale metadata.
  • Registry Validation: Added cs_validate_dgp_registry() and cs_validate_tau_grid() to enforce constitutional invariants.
  • Run Fingerprinting: Implemented cs_build_config_fingerprint() to prevent unsafe resumes when simulation settings (e.g., bootstrap, n) change.
  • Dependencies: Added ggplot2 and tidyr to Imports.

CausalStress 0.1.2

Core Features

  • Atomic Persistence: Enabled crash recovery via pins. Results are saved to disk immediately after each seed is generated.
  • The Airlock: Introduced cs_airlock() to centrally strip forbidden columns (y0, y1, p, structural_te) before data reaches an estimator.
  • Provenance: Implemented cs_audit() to track git hashes, timestamps, and session info for every run.
  • Tidy Output: Added cs_tidy() and collectors (cs_collect_att, cs_collect_qst) to flatten rich result objects into analysis-ready tibbles.

Safety & UX

  • Resume Guards: Runners now raise informative errors when skip_existing = TRUE encounters a pin computed with mismatched settings (e.g., missing bootstrap CIs).
  • Documentation: Added the “From Run to History” vignette and a comprehensive README.
  • Aliases: Added cs_run_campaign() as an alias for cs_run_seeds() for consistency with design docs.

CausalStress 0.1.0

  • First MVP:
    • Synthetic DGPs: synth_baseline, synth_heavytail.
    • Estimators: oracle_att, lm_att, ipw_att, grf_dr_att.
    • Registry system for DGPs and estimators.
    • Runners: cs_run_single(), cs_run_seeds().
    • Contracts + tests for DGPs, estimators, runners, truth helpers.