Skip to content

Development Handoff

Last updated: 2026-05-29

For the concise current-facing research handoff, see current_research_next_steps.md. This file preserves the longer development history, local paper-workflow notes, and release-hardening log.

Current State

The repository is on main.

Recent commits (library / reporting; paper asset paths are local-only—see below):

  • 9da6dea - Paper workflow local-only; update CHANGELOG and handoff
  • 8a5edb4 - Add paper support paths to .gitignore
  • f40cea1 - Update development handoff
  • f147529 - Harden paper workflow cache handling
  • 7e2453f - Add paper benchmark workflow scaffolding
  • 9868d48 - Add advanced benchmark diagnostics and reporting

Tracked in Git: library source, configs/suites/ smoke and shared benchmarks, docs/ (including this file), CHANGELOG.md, tests under tests/ except the optional paper integration test listed under “Local-only paths”.

Clean-clone docs now include docs/paper_workflow.md, which explains the local-only paper kit without tracking draft paper artefacts.

This file (development_handoff.md) and CHANGELOG.md are meant to stay synced with the remote so you can continue planning and execution notes on any machine after git pull.

Local-only paths (not synced to the remote)

The following are listed in .gitignore and remain on your workstation only—copy or recreate them when you set up a new environment for publication runs:

  • paper_support/ — entire directory (runner package, optional artefacts/ staging).
  • configs/suites/paper/ — paper-oriented YAML manifests.
  • tests/integration/test_paper_workflow.py — optional smoke test for the local runner.
  • reports/ — all benchmark HTML/CSV/LaTeX outputs (including reports/paper/<run_id>/).
  • .lrdbench_cache/ — estimate cache and matplotlib config dir used by local paper runs.

Other common ignores: .codex, .venv/, etc., per .gitignore.

Completed recently (core library)

  • Publication plotting is a core reporting capability:
  • matplotlib and seaborn are core dependencies;
  • requested figures are not silently skipped when plotting imports fail;
  • disagreement and sensitivity heatmaps use seaborn.
  • Estimator disagreement metrics:
  • cross_estimator_dispersion;
  • pairwise_estimator_disagreement;
  • family_level_disagreement;
  • estimator_disagreement.csv.
  • Estimator parameter variants for scale/window sensitivity:
  • manifest variants;
  • parameter_variant_sensitivity;
  • max_variant_drift;
  • scale_window_sensitivity.csv.
  • Benchmark-level uncertainty:
  • aggregate bootstrap CIs;
  • paired bootstrap estimator differences;
  • raw result-store uncertainty scope;
  • benchmark_uncertainty.csv.
  • Report completeness exports:
  • estimator_metadata.csv;
  • failures.csv;
  • environment.json;
  • artefact_index.csv;
  • raw artefacts.csv.
  • Richer HTML and publication-oriented LaTeX tables for disagreement, sensitivity, benchmark uncertainty, and failures.
  • Opt-in report figures:
  • degradation_curve;
  • disagreement_heatmap;
  • sensitivity_heatmap;
  • benchmark_uncertainty_intervals;
  • false_positive_lrd.

Local paper workflow (reproduce on each machine)

When you maintain paper_support/ and configs/suites/paper/ locally, a typical layout is:

  • Manifests (examples): canonical_ground_truth.yaml, stress_contamination.yaml, null_false_positive.yaml, sensitivity_disagreement.yaml under configs/suites/paper/.
  • Runner: python -m paper_support.run_paper_suites <manifest> [...] from the repo root.
  • Writes: normal reports under reports/paper/<run_id>/; copies LaTeX and figures to paper_support/artefacts/; appends paper_support/artefacts/run_index.csv.
  • Cache/config: repo-root-relative estimate caches under .lrdbench_cache/...; set MPLCONFIGDIR to .lrdbench_cache/matplotlib in the runner to avoid unwritable home-config warnings.

Example command (paths exist only after you create or copy the local kit):

python -m paper_support.run_paper_suites configs/suites/paper/null_false_positive.yaml

Example outcomes from a prior local run (IDs will differ on your machine):

  • HTML report: reports/paper/<run_id>/html/report.html
  • Run index: paper_support/artefacts/run_index.csv
  • Typical staged count: several LaTeX tables plus figures (disagreement heatmap, benchmark uncertainty intervals, false-positive LRD plot when requested).

Verification (CI vs local)

On the remote / CI, the following are representative:

python -m ruff check .
python -m pytest

Optional local check after (re)creating paper_support and the paper integration test:

python -m pytest tests/integration/test_paper_workflow.py

Full pytest pass counts move with the test suite; run pytest after each pull for the current number.

Local verification on 2026-04-22 after installing .[all]:

  • python -m ruff check . — passed.
  • python -m pytest — passed: 56 tests, 80% total coverage in a clean clone; 57 tests, 80% total coverage when the local-only tests/integration/test_paper_workflow.py is present.
  • mkdocs build --strict — passed after adding the tracked paper workflow docs page and nav entry.
  • python -m pytest tests/integration/test_paper_workflow.py — passed after recreating the local-only paper runner scaffold on this machine.

Note: SPECIFICATION.md refers to lrdbench-design-specifications.pdf as the frozen design authority when present locally. That PDF is not currently tracked in this clean repository.

Local paper kit recreated on this machine:

  • paper_support/run_paper_suites.py;
  • configs/suites/paper/canonical_ground_truth.yaml;
  • configs/suites/paper/stress_contamination.yaml;
  • configs/suites/paper/null_false_positive.yaml;
  • configs/suites/paper/sensitivity_disagreement.yaml;
  • tests/integration/test_paper_workflow.py.

These paths remain ignored by Git.

Canonical paper smoke run on 2026-04-22:

  • Command: python -m paper_support.run_paper_suites configs/suites/paper/canonical_ground_truth.yaml
  • Run ID: 2d321e2e-3cb2-4c41-a43b-a5caaa09fe5e
  • HTML report: reports/paper/2d321e2e-3cb2-4c41-a43b-a5caaa09fe5e/html/report.html
  • Staged artefacts: 7 files under paper_support/artefacts/2d321e2e-3cb2-4c41-a43b-a5caaa09fe5e/.
  • Key table counts: per_stratum_metrics.csv 151 rows, benchmark_uncertainty.csv 48 rows, estimator_disagreement.csv 76 rows, failures.csv 42 rows.

Additional local paper suite runs on 2026-04-22:

  • stress_contamination.yaml
  • Run ID: 10cfa083-3278-4f1b-9862-f489580bcc09
  • HTML report: reports/paper/10cfa083-3278-4f1b-9862-f489580bcc09/html/report.html
  • Staged artefacts: 8 files.
  • Key table counts: per_stratum_metrics.csv 296 rows, stress_metrics.csv 160 rows, benchmark_uncertainty.csv 96 rows, estimator_disagreement.csv 74 rows, failure_map.csv 48 rows, failures.csv 48 rows.
  • Local manifest correction made before the successful run: outliers uses rate/amplitude; polynomial_trend uses order/strength.
  • null_false_positive.yaml
  • Run ID: f811d6eb-5008-4f2c-b2b4-3d1f69b71fae
  • HTML report: reports/paper/f811d6eb-5008-4f2c-b2b4-3d1f69b71fae/html/report.html
  • Staged artefacts: 7 files.
  • Key table counts: per_stratum_metrics.csv 42 rows, estimator_disagreement.csv 60 rows, failures.csv 14 rows.
  • false_positive_lrd_rate balanced-global rows were produced for RS, DFA, and GPH; all were 0.0 in this small smoke-sized null run.
  • sensitivity_disagreement.yaml
  • Run ID: f7b513aa-829c-43ec-a9e1-428a73c4e101
  • HTML report: reports/paper/f7b513aa-829c-43ec-a9e1-428a73c4e101/html/report.html
  • Staged artefacts: 7 files.
  • Key table counts: per_stratum_metrics.csv 66 rows, estimator_disagreement.csv 56 rows, scale_window_sensitivity.csv 14 rows, failures.csv 26 rows.

Local paper kit restored again on 2026-04-25 from the historical scaffold in commits 7e2453f and f147529:

  • paper_support/run_paper_suites.py;
  • configs/suites/paper/canonical_ground_truth.yaml;
  • configs/suites/paper/stress_contamination.yaml;
  • configs/suites/paper/null_false_positive.yaml;
  • configs/suites/paper/sensitivity_disagreement.yaml;
  • tests/integration/test_paper_workflow.py.

Verification on 2026-04-25:

  • python -m ruff check paper_support tests/integration/test_paper_workflow.py — passed.
  • python -m pytest tests/integration/test_paper_workflow.py — passed: 2 tests.
  • Full local paper command used PYTHONPATH=src because the package was not installed in this shell:
PYTHONPATH=src python -m paper_support.run_paper_suites \
  configs/suites/paper/canonical_ground_truth.yaml \
  configs/suites/paper/stress_contamination.yaml \
  configs/suites/paper/null_false_positive.yaml \
  configs/suites/paper/sensitivity_disagreement.yaml

Local paper suite runs on 2026-04-25:

  • canonical_ground_truth.yaml
  • Run ID: 65a814d3-f659-456e-a569-85a0699228c7
  • HTML report: reports/paper/65a814d3-f659-456e-a569-85a0699228c7/html/report.html
  • Staged artefacts: 7 files.
  • Key table counts: per_stratum_metrics.csv 1726 rows, benchmark_uncertainty.csv 961 rows, estimator_disagreement.csv 2542 rows, failures.csv 649 rows.
  • stress_contamination.yaml
  • Run ID: 013396d5-2a31-444d-a5cf-d13112b78231
  • HTML report: reports/paper/013396d5-2a31-444d-a5cf-d13112b78231/html/report.html
  • Staged artefacts: 8 files.
  • Key table counts: per_stratum_metrics.csv 9516 rows, stress_metrics.csv 10801 rows, benchmark_uncertainty.csv 4376 rows, estimator_disagreement.csv 10576 rows, failures.csv 2561 rows.
  • null_false_positive.yaml
  • Run ID: 5600a62a-cb4d-42a8-a308-17351eb8cf6b
  • HTML report: reports/paper/5600a62a-cb4d-42a8-a308-17351eb8cf6b/html/report.html
  • Staged artefacts: 8 files.
  • Key table counts: per_stratum_metrics.csv 236 rows, benchmark_uncertainty.csv 49 rows, estimator_disagreement.csv 631 rows, failures.csv 77 rows.
  • Observed false_positive_lrd_rate aggregate range: 0.0 to 0.1.
  • sensitivity_disagreement.yaml
  • Run ID: d5e94487-44e4-425e-a274-73fd15e6053c
  • HTML report: reports/paper/d5e94487-44e4-425e-a274-73fd15e6053c/html/report.html
  • Staged artefacts: 8 files.
  • Key table counts: per_stratum_metrics.csv 1118 rows, benchmark_uncertainty.csv 574 rows, estimator_disagreement.csv 2201 rows, scale_window_sensitivity.csv 367 rows, failures.csv 577 rows.

Run index: paper_support/artefacts/run_index.csv. The local estimate cache now contains complete fit caches for these runs under .lrdbench_cache/paper_*, which should make repeated local paper runs substantially faster.

Planning note from inspecting the 2026-04-25 outputs: the current paper leaderboards should not be treated as final paper rankings. Some truth-free aggregate metrics produce synthetic estimator names such as __all_estimators__, and variant sensitivity metrics are grouped by base estimator while accuracy rows are emitted per variant. Use the exported disagreement/sensitivity tables and figures for interpretation until leaderboard semantics are tightened or the paper manifests remove those metrics from leaderboard components.

Follow-up on 2026-04-25:

  • Tracked library fix: WeightedRankLeaderboardBuilder now ranks only estimator names declared in the manifest. Synthetic aggregate diagnostic names such as __all_estimators__ remain in metric exports, but cannot appear as leaderboard rows.
  • Added a unit regression test for this behavior: tests/unit/test_leaderboard.py.
  • Local paper manifest adjustment: removed group-level disagreement/sensitivity metrics from leaderboard component lists while leaving those metrics in reports and exported tables.
  • Added local-only paper smoke manifests under configs/suites/paper/smoke_*.yaml.
  • Added --smoke to python -m paper_support.run_paper_suites, selecting the smoke manifest set.
  • Verification:
  • python -m ruff check . — passed.
  • python -m pytest — passed: 60 tests, 80% total coverage.

  • Local paper smoke command:

    PYTHONPATH=src python -m paper_support.run_paper_suites --smoke \
  --export-root reports/paper_smoke \
  --artefact-root paper_support/artefacts_smoke \
  --index-path paper_support/artefacts_smoke/run_index.csv

Smoke paper runs on 2026-04-25:

  • smoke_canonical_ground_truth.yaml
  • Run ID: 071d5d24-2433-407d-90e2-e19b2b34f689
  • HTML report: reports/paper_smoke/071d5d24-2433-407d-90e2-e19b2b34f689/html/report.html
  • Staged artefacts: 7 files.
  • smoke_stress_contamination.yaml
  • Run ID: d28f1de0-ab83-407e-a865-6c9dc087ce6c
  • HTML report: reports/paper_smoke/d28f1de0-ab83-407e-a865-6c9dc087ce6c/html/report.html
  • Staged artefacts: 8 files.
  • smoke_null_false_positive.yaml
  • Run ID: 17f564ab-060d-4501-bbcb-08aeb1fd246e
  • HTML report: reports/paper_smoke/17f564ab-060d-4501-bbcb-08aeb1fd246e/html/report.html
  • Staged artefacts: 8 files.
  • smoke_sensitivity_disagreement.yaml
  • Run ID: 63970c6b-792f-4546-a815-da880658eafd
  • HTML report: reports/paper_smoke/63970c6b-792f-4546-a815-da880658eafd/html/report.html
  • Staged artefacts: 8 files.

Refreshed full paper reports after the leaderboard fix:

  • canonical_ground_truth.yaml
  • Run ID: cad42a6a-79bb-450a-87c0-d084171cbf31
  • HTML report: reports/paper/cad42a6a-79bb-450a-87c0-d084171cbf31/html/report.html
  • Leaderboard rows checked: 6; no synthetic estimator rows.
  • sensitivity_disagreement.yaml
  • Run ID: 1ad150af-055f-4300-8f96-c0f8f68af393
  • HTML report: reports/paper/1ad150af-055f-4300-8f96-c0f8f68af393/html/report.html
  • Leaderboard rows checked: 9; no synthetic estimator rows.

Grid-tuning pass on 2026-04-25:

  • Current output summary before tuning:
  • stress degradation was strongest for polynomial_trend, moderate for outliers, weaker for heavy_tail_noise, and effectively invariant for level_shift;
  • null false-positive aggregate rows ranged from 0.0 to 0.05 in the then-current full run;
  • sensitivity run had 24 invalid estimates, all from WaveletOLS::balanced_band at n=512 (insufficient_signal_for_wavelet_ols).
  • Full local paper manifest changes:
  • canonical replicates increased from 3 to 5;
  • stress replicates increased from 3 to 5;
  • stress level_shift reduced to one negative-control level: shift: [0.75];
  • stress polynomial_trend strengths expanded to [0.25, 0.75, 1.25];
  • stress outliers rates expanded to [0.01, 0.05, 0.1];
  • stress heavy_tail_noise scales expanded to [0.25, 0.75, 1.25];
  • null false-positive replicates increased from 10 to 30;
  • sensitivity/disagreement replicates increased from 3 to 5;
  • sensitivity WaveletOLS::balanced_band changed from j_drop_low: 2, j_drop_high: 2 to j_drop_low: 1, j_drop_high: 1.
  • Validation:
  • python -m pytest tests/integration/test_paper_workflow.py — passed: 3 tests.
  • All configs/suites/paper/*.yaml manifests loaded successfully with PYTHONPATH=src.

  • Paper smoke preflight passed with:

    PYTHONPATH=src python -m paper_support.run_paper_suites --smoke \
  --export-root reports/paper_smoke \
  --artefact-root paper_support/artefacts_smoke \
  --index-path paper_support/artefacts_smoke/run_index.csv

Latest smoke paper runs after grid tuning:

  • smoke_canonical_ground_truth.yaml
  • Run ID: ba3d36b8-1dc6-48fc-88a7-70bc39fb325b
  • HTML report: reports/paper_smoke/ba3d36b8-1dc6-48fc-88a7-70bc39fb325b/html/report.html
  • smoke_stress_contamination.yaml
  • Run ID: 32e65c9b-4851-4230-8dc2-e1c1cfd60c9b
  • HTML report: reports/paper_smoke/32e65c9b-4851-4230-8dc2-e1c1cfd60c9b/html/report.html
  • smoke_null_false_positive.yaml
  • Run ID: 420d3440-5bc3-4bac-814f-34713ef00152
  • HTML report: reports/paper_smoke/420d3440-5bc3-4bac-814f-34713ef00152/html/report.html
  • smoke_sensitivity_disagreement.yaml
  • Run ID: 0f09e394-09c1-4cc0-af59-c25ce8d3ec29
  • HTML report: reports/paper_smoke/0f09e394-09c1-4cc0-af59-c25ce8d3ec29/html/report.html

Next paper-campaign command, when ready for a longer local run:

PYTHONPATH=src python -m paper_support.run_paper_suites \
  configs/suites/paper/canonical_ground_truth.yaml \
  configs/suites/paper/stress_contamination.yaml \
  configs/suites/paper/null_false_positive.yaml \
  configs/suites/paper/sensitivity_disagreement.yaml

Tuned full paper campaign completed on 2026-04-25 with the command above.

Final tuned paper runs:

  • canonical_ground_truth.yaml
  • Run ID: 53f3f969-4f42-4233-82df-e0137ceeb69b
  • HTML report: reports/paper/53f3f969-4f42-4233-82df-e0137ceeb69b/html/report.html
  • Staged artefacts: 7 files.
  • Key table counts: per_stratum_metrics.csv 1726 rows, benchmark_uncertainty.csv 961 rows, estimator_disagreement.csv 3886 rows, failures.csv 649 rows.
  • Estimates: 960 total, 0 invalid.
  • Leaderboard rows checked: 6; no synthetic estimator rows.
  • stress_contamination.yaml
  • Run ID: c4f56a44-1d66-4808-8705-28c2b20383f9
  • HTML report: reports/paper/c4f56a44-1d66-4808-8705-28c2b20383f9/html/report.html
  • Staged artefacts: 8 files.
  • Key table counts: per_stratum_metrics.csv 11916 rows, stress_metrics.csv 22801 rows, benchmark_uncertainty.csv 5496 rows, estimator_disagreement.csv 20416 rows, failures.csv 3201 rows.
  • Estimates: 6000 total, 0 invalid.
  • Leaderboard rows checked: 5; no synthetic estimator rows.
  • Stress pattern: level_shift remains a useful negative control (estimate_drift 0.0, degradation ratio 1.0). Degradation increases clearly with stronger polynomial trends, outlier rates/amplitudes, and heavy-tail noise scales. Highest mean degradation ratios in this run were polynomial trend strength 1.25 (order=2: 7.8751, order=1: 7.3011), followed by high-rate/high-amplitude outliers and scale 1.25 heavy-tail noise.
  • null_false_positive.yaml
  • Run ID: 9e55852c-d670-4e11-a1c5-8df238feb187
  • HTML report: reports/paper/9e55852c-d670-4e11-a1c5-8df238feb187/html/report.html
  • Staged artefacts: 8 files.
  • Key table counts: per_stratum_metrics.csv 236 rows, benchmark_uncertainty.csv 49 rows, estimator_disagreement.csv 1751 rows, failures.csv 77 rows.
  • Estimates: 600 total, 0 invalid.
  • Leaderboard rows checked: 5; no synthetic estimator rows.
  • Balanced-global false-positive LRD rates: RS 0.0167, DFA 0.0333, GPH 0.0667, Periodogram 0.0167, ModifiedLocalWhittle 0.05.
  • sensitivity_disagreement.yaml
  • Run ID: f5aeb79c-c1d6-4219-92ab-e68d2e591783
  • HTML report: reports/paper/f5aeb79c-c1d6-4219-92ab-e68d2e591783/html/report.html
  • Staged artefacts: 8 files.
  • Key table counts: per_stratum_metrics.csv 1184 rows, benchmark_uncertainty.csv 628 rows, estimator_disagreement.csv 3721 rows, scale_window_sensitivity.csv 559 rows, failures.csv 625 rows.
  • Estimates: 720 total, 0 invalid.
  • Leaderboard rows checked: 9; no synthetic estimator rows.
  • Balanced-global parameter_variant_sensitivity: DFA 0.0477, DMA 0.0640, WaveletOLS 0.0284.
  • Balanced-global max_variant_drift: DFA 0.1114, DMA 0.1518, WaveletOLS 0.0673.

The tuned full campaign writes the current paper_support/artefacts/run_index.csv and completes the immediate publication-run preflight. Next recommended work is interpretive: inspect the latest HTML reports and staged figures/tables, choose the subset of results to carry into the manuscript, and decide whether the stress narrative should emphasize polynomial-trend nonstationarity first or present heavy tails/outliers as separate robustness sections.

Manuscript-facing selection pass on 2026-04-25:

  • Created local-only selection bundle under paper_support/artefacts/selected/.
  • Selection note:
  • paper_support/artefacts/selected/manuscript_selection.md
  • Candidate main figures:
  • paper_support/artefacts/selected/fig1_stress_degradation_curve.png
  • paper_support/artefacts/selected/fig2_null_false_positive_lrd.png
  • paper_support/artefacts/selected/fig3_scale_window_sensitivity_heatmap.png
  • Candidate supplementary figure:
  • paper_support/artefacts/selected/fig_supp_canonical_disagreement_heatmap.png
  • Main-text numeric anchors recorded in the selection note:
  • canonical balanced-global MAE;
  • stress balanced-global degradation ratio;
  • null balanced-global false-positive LRD rate;
  • scale/window sensitivity and max variant drift;
  • selected stress-severity degradation ratios.

Selection recommendation: use the stress degradation result as the first main results figure, then use null false-positive and scale/window sensitivity as separate results panels/figures. Keep the canonical disagreement heatmap and full audit-style LaTeX/CSV exports in supplement unless the paper needs an early estimator-disagreement framing figure. The current metrics_summary.tex exports are too broad for main text; derive compact manuscript tables from CSV exports or from manuscript_selection.md.

Public research release roadmap documented on 2026-04-25:

  • Tracked roadmap: docs/public_release_roadmap.md.
  • MkDocs nav entry: Meta → Public release roadmap.
  • Roadmap separates local manuscript infrastructure from public library readiness.
  • All alpha, beta, and release-candidate phases have been completed. The stable v1.0.2 package is now published on PyPI with the unchanged 1.0.0 public output contract.

Hardening pass: 2026-05-11

Focus: design quality, mathematical accuracy, and onboarding friction.

Design & CLI improvements

  • Added --dry-run to lrdbench run and BenchmarkRunner.preview() for grid-size preview.
  • Added machine-readable manifest JSON Schema at configs/contracts/manifest_schema.json.

Documentation & onboarding

  • Rewrote docs/architecture.md into a full contributor guide.
  • Expanded CONTRIBUTING.md with validation commands and smoke-test verification.
  • Added docs/faq.md with troubleshooting for common failures.
  • Added docs/parameter_glossary.md covering all major estimator parameters.
  • Fleshed out all four Jupyter notebooks with explanatory text and new code cells.
  • Fixed terminology drift: all canonical manifests now use family: temporal consistently.

API documentation

  • Added comprehensive docstrings to all abstract base classes in interfaces.py.
  • Added docstrings to BenchmarkRunner.run(), run_manifest_path(), and key schema dataclasses.
  • Expanded bootstrap module docstrings with methodology notes.

Mathematical hardening

  • Deduplicated GPH / Periodogram regression cores into _log_periodogram_regression_d().
  • Fixed GPH bandwidth handling so params.m is passed through to the shared regression core.
  • Added optional cosine-bell spectral tapering (params.taper: cosine) to GPH and Periodogram.
  • Implemented Anis-Lloyd finite-sample correction for RS (use_anis_lloyd_correction).
  • Corrected RS to estimate the log-log R/S scaling slope across subseries lengths instead of using a single full-record R/S ratio.
  • Documented GHE flat_slope_tol heuristic explicitly; noted how to disable it.
  • Documented RS finite-sample bias and the Anis-Lloyd correction in estimator status.
  • Documented fOU as an Euler–Maruyama approximation in simulate_fou docstring.
  • Documented bootstrap block-length rationale in docs/benchmark_protocol.md.

Verification

  • python -m pytest tests/ -q --tb=short -o "addopts=" — passed: 145 tests.
  • Notebook integration tests (test_research_notebooks.py) — all 4 passed.
  • Dry-run smoke checks confirmed for ground-truth and stress-test modes.
  • Spectral tapering and Anis-Lloyd correction produce plausible estimates in manual checks.

Research Direction

The project has two linked roles:

  1. Provide a comprehensive, reproducible LRD benchmarking framework.
  2. Support a benchmark paper showing that classical second-order LRD estimators are unstable outside their intended stationary finite-variance regimes and cannot reliably separate genuine LRD from nonstationarity, heavy tails, intermittency, and related mechanisms.

This benchmark paper is intended to precede a theory paper arguing that the failure is structural rather than estimator-specific.

  1. On a machine where the local paper kit exists, run the canonical ground-truth manifest:
python -m paper_support.run_paper_suites configs/suites/paper/canonical_ground_truth.yaml
  1. Inspect the canonical report, figures, and paper_support/artefacts/run_index.csv.
  2. Run stress_contamination.yaml and sensitivity_disagreement.yaml from the same local kit.
  3. Compare tables and figures across runs; adjust manifest grids, estimators, or report.figure_set before longer benchmark campaigns.

When working only from a clean clone without the local paper kit, use lrdbench run with tracked manifests under configs/suites/ (e.g. smoke suites) and the same reporting options supported by docs/benchmark_protocol.md.