Kahlus v1 EEG Baseline Plan#
What v1 Is#
Kahlus v1 is the EEG forecasting baseline lane:
EEG patch/window tokenizer
-> temporal encoder
-> latent state z_t
-> future-window decoder
This sprint makes the lane benchmarkable under subject-held-out splits with strong local baselines and an explicit evidence gate.
What v1 Is Not#
This sprint does not claim diagnosis, treatment, epilepsy detection, depression detection, foundation-model status, SOTA, Kahlus v2 success, or Kahlus v3 success. ResearchDock and Kahlus-Affect remain later device branches and are not implemented here.
Objective#
The v1 training objective is documented as:
L_v1 = lambda_1 MSE(Y_hat, Y)
+ lambda_2 MAE(Y_hat, Y)
+ lambda_3 L_smooth
with:
L_smooth = sum_t ||Y_hat_{t+1} - 2Y_hat_t + Y_hat_{t-1}||_2^2
The Sprint A implementation exposes the smoothness term for tests and reports the baseline metrics needed before any main-model claim.
Baseline Ladder#
The default runnable ladder is:
persistence
linear_ridge
autoregressive_ridge
MLP
TCN
Transformer
SSM fallback
NeuralStateSpaceTranslator ("neurotwin")
All requested models receive the same window length, forecast horizon, split, and training step budget. Baselines are results, not scaffolding.
Split Rule#
Subject-held-out is the default. The v1 audit writes split_audit.json with train/val/test
subjects, subject-overlap status, window-overlap status, and failure reasons. Session-held-out
is allowed by the gate shape but is not the default script path.
HBN-EEG External Path#
No public dataset is downloaded automatically. The HBN path is local-only and expects a
user-provided manifest.jsonl under --data-root; each row points to a .npy or .npz
signal file shaped [time, channels]. Missing local data fails clearly instead of falling
back to synthetic data.
A100 Scaling Boundary#
Sprint A/B are local validation sprints only. They must not launch A100 or cluster jobs. The 7x NVIDIA A100 80GB cluster is reserved for later work after local synthetic fixtures, subject-held-out audits, baseline ladders, and evidence gates pass from a clean merge/tag.
Future A100 handoff packaging must label the hardware honestly as 7xA100, include exact
commit hash, clean-worktree proof, runner tarball, configs, checksum manifest, CPU smoke,
DDP/torchrun command, evidence bundle writer, and audit script, and must exclude secrets,
checkpoints unless explicitly allowed, and raw participant data.
Outputs#
Each script run writes:
metrics.json
metrics.csv
baseline_table.json
baseline_table.csv
split_audit.json
evidence_gate.json
run_config.json
dataset_summary.json
failure_reasons.json
diagnostic_report.md
per_subject_metrics.csv
per_channel_metrics.csv
per_horizon_metrics.csv
baseline_verification.json
baseline_checksum_manifest.json
The narrow allowed claim scope is eeg_future_forecasting_benchmark_ready.
Verification#
PYTHONPATH=src python3 scripts/run_eeg_v1_baselines.py --dataset synthetic_fixture --out-dir /tmp/kahlus_v1_eeg_smoke --seed 0
Sprint C: Reviewer-Facing Autocorrelation Summary#
The baseline smoke command now prints the short-horizon, shuffled-target, long-horizon, and non-overlapping-window diagnostic MSE summary fields directly to stdout. The Markdown diagnostic report includes the same summary table so reviewers can see why a low ridge or persistence MSE may reflect autocorrelation rather than model understanding.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint A.3: First-Class SSM and Train-Only Shuffled Control#
The EEG v1 smoke now treats tiny_ssm as the first-class SSM baseline in the default ladder. The
older ssm_fallback executable remains available for compatibility, but the narrow
eeg_future_forecasting_benchmark_ready artifact gate requires tiny_ssm to be present in emitted
metrics.
The shuffled-target negative control is also required. It shuffles training targets only, keeps
validation/test targets unchanged, and records shuffle_boundary=train_split_only in the
autocorrelation diagnostic row. This prevents the control from proving anything by corrupting
held-out targets.
The diagnostic report now surfaces tiny_ssm_mse, shuffled_target_control_mse,
persistence_or_ridge_dominates, and shuffled_target_close_to_real_baselines so reviewers can see
whether low MSE is still dominated by simple continuation or broken target alignment.
Verification:
PYTHONPATH=src python3 -m unittest discover -s tests/eeg_v1 -v
PYTHONPATH=src python3 scripts/run_eeg_autocorr_diagnostics.py --dataset synthetic_fixture --out-dir /tmp/kahlus_v1_a3_autocorr --seed 0
Sprint D: HBN Local Manifest Validation#
The HBN-EEG local adapter now rejects relative manifest paths that escape --data-root and
rejects unsupported signal file extensions before NumPy loading. Absolute paths remain allowed
only because the adapter contract already permits user-provided local absolute paths; this is
local validation, not an automatic public-data loader or downloader.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint E: HBN Local Provenance Boundary#
HBN-style local manifest runs now write data_source and benchmark_status into stdout,
run_config.json, metrics.json, and the diagnostic report. Local manifest fixture runs are
explicitly marked local_manifest_not_public_hbn_benchmark so fixture smoke results cannot be
presented as public HBN benchmark evidence.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint F: HBN Numeric Manifest Validation#
The HBN-EEG local adapter now rejects non-finite or non-positive sampling_rate values before
time-vector construction and rejects signal arrays containing NaN or Inf before any windowing,
baseline, or report path can run. This is local-manifest validation only; it does not add public
dataset downloads, loaders, or benchmark claims.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint G: HBN Subject Metadata Validation#
The HBN-EEG local adapter now rejects manifest rows with missing or blank subject_id values
before split-manifest construction. Subject identity is required evidence for subject-held-out
claims, so malformed local rows fail clearly instead of producing raw key errors or empty-subject
splits.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint H: HBN Manifest Row Parsing Validation#
The HBN-EEG local adapter now rejects malformed JSONL rows and non-object JSON rows with
line-numbered ValueError messages before path, signal, or split logic runs. This keeps local
fixture failures auditable without adding any public-data loader or benchmark claim.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint I: HBN Nonempty Signal Shape Validation#
The HBN-EEG local adapter now rejects [time, channels] arrays with an empty time axis or empty
channel axis before time-vector, split, or baseline logic runs. Forecasting windows require at
least one sample and one channel, so empty local fixtures fail clearly instead of becoming
misleading benchmark artifacts.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint J: HBN Channel Count Consistency Validation#
The HBN-EEG local adapter now rejects local manifests whose recordings have inconsistent channel counts. This deliberately fails closed until explicit montage/channel-alignment metadata exists, because silently mixing different channel dimensions would make baseline comparisons ambiguous.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint K: HBN Unique Record Identity Validation#
The HBN-EEG local adapter now rejects duplicate record_id values before split manifests,
metrics, or evidence artifacts are built. Default generated IDs remain unique by row index; this
guard protects user-provided local manifest identifiers from silently aliasing two recordings.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint L: HBN Optional Channel Names Validation#
The HBN-EEG local adapter now validates optional channel_names metadata when local manifests
provide it. Names must be a JSON array of nonempty strings, match the signal channel count, and
stay consistent across recordings; otherwise the adapter fails closed rather than silently
pretending channel alignment is known.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint M: HBN Optional Text Metadata Validation#
The HBN-EEG local adapter now rejects blank provided session_id, site_id, and record_id
values. Omitted optional fields still receive deterministic defaults, but explicitly blank fields
fail closed so local evidence artifacts do not carry ambiguous recording metadata.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint N: HBN Sampling Rate Consistency Validation#
The HBN-EEG local adapter now rejects local manifests whose recordings use mixed
sampling_rate values. This fails closed until an explicit resampling path exists, because sample
windows at different time scales should not be compared as equivalent baseline evidence.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint O: Dataset Summary Evidence Artifact#
The EEG v1 baseline runner now writes dataset_summary.json with bounded, non-raw metadata:
dataset/source/status, split type, subject counts, train/val/test window counts, window settings,
and channel count. This gives reviewers a compact evidence check without exposing raw signals or
inflating benchmark claims.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint P: Invalid Task Configuration CLI Guard#
The EEG v1 baseline runner now reports impossible forecast task settings as a clear local configuration error instead of a Python traceback. Invalid window geometry is not a model result and must fail before baseline tables, evidence gates, or interpretation artifacts are written.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint Q: Baseline Evidence Gate Criteria#
The EEG v1 baseline evidence gate now records its decision criteria directly in
evidence_gate.json: minimum forecast horizon, allowed held-out split types, required split
audit, baseline table, finite metrics, calibration check, and the allowed narrow claim scope.
The diagnostic report renders the same criteria so reviewers can see that the gate is
benchmark-readiness only, not a model-understanding, clinical, or SOTA threshold.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint R: Baseline Report Run Config Summary#
The EEG v1 baseline diagnostic report now includes a Run Config section sourced from the
same bounded metadata written to run_config.json: seed, train-step budget, model list,
window settings, data source/status, selection policy, and claim scope. This makes the
human-readable report replayable without opening the JSON sidecar first.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint S: Baseline Report Split-Audit Failure Summary#
The EEG v1 baseline diagnostic report now includes a Split Audit Failures section when the
split audit records detailed leakage or split-validation failures. This surfaces the underlying
split evidence in the human-readable report instead of only showing the derived gate failure.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint T: Baseline Failure Reasons Split-Audit Sidecar#
The EEG v1 baseline failure_reasons.json artifact now includes split_audit_failures
alongside baseline and gate failures. This gives audit consumers the same structured
split-failure sidecar coverage already present in the few-shot adaptation evidence bundle.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint U: Baseline Report Baseline-First Method Order#
The EEG v1 baseline diagnostic report now includes a Method Order table before ranking.
Requested methods are tagged as baseline or main_model, so low-MSE baseline results are
visible before any comparison with NeuroTwin/Kahlus.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint V: Target Scale Context for Low MSE Interpretation#
The EEG v1 baseline bundle now writes target_scale_context.json and a matching diagnostic
report section. It reports target mean/std/variance/min/max and model RMSE/MSE ratios against
target scale, making clear that synthetic fixture MSE is in normalized fixture units rather
than raw EEG microvolts.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint W: Target Scale Context in Smoke Stdout#
The EEG v1 baseline runner now prints target units, target standard deviation, target variance, and the best baseline RMSE relative to target standard deviation directly in stdout. The smoke command therefore explains low normalized MSE without requiring the reviewer to open the JSON sidecar first.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint X: Metric Breakdown Summary in Diagnostic Report#
The EEG v1 diagnostic report now summarizes how many per-subject, per-channel, and per-horizon metric rows were written, and points to the detailed CSV sidecars. This keeps the Markdown report auditable while leaving detailed rows in machine-readable files.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint Y: Baseline Gap Summary in Diagnostic Report#
The EEG v1 diagnostic report now renders the existing persistence, ridge, and best-baseline gap
metrics with an explicit sign convention. This makes baseline wins or losses visible in the
Markdown report instead of requiring reviewers to inspect metrics.json.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint Z: Baseline Failure Details in Diagnostic Report#
The EEG v1 diagnostic report now lists requested baseline failures from the same payload written
to failure_reasons.json. Unknown or unavailable baselines remain failures; the Markdown report
now exposes them directly for audit.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint AA: Delta Prediction Summary Exposure#
The EEG v1 autocorrelation diagnostics already run a delta-prediction control where the target is future-minus-input instead of the raw future signal. The summary, CLI stdout, and diagnostic report now expose the delta-control best MSE gap against the short-horizon raw task so reviewers can see whether low raw MSE is just waveform continuation.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint AB: Autocorrelation Diagnostic Reasons in Report#
The EEG v1 diagnostic report now includes a reason column for each autocorrelation diagnostic.
Skipped, blocked, or not-applicable controls are therefore visible in the human-readable report
instead of only in autocorrelation_diagnostics.csv.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint AC: HBN Sampling Rate Provenance#
The v1 future-window task now carries a single finite dataset sampling rate into run_config.json,
dataset_summary.json, and the diagnostic Markdown report. This makes local HBN-style fixture
windows auditable in seconds/Hz terms after the adapter has already rejected mixed-rate manifests.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint AD: Baseline Evidence Checksum Manifest#
The EEG v1 baseline artifact bundle now writes baseline_checksum_manifest.json with SHA-256
digests and byte counts for the emitted JSON/CSV/Markdown evidence artifacts. The manifest
excludes itself to avoid circular hashing and gives reviewers a local integrity check before
trusting baseline, autocorrelation, split-audit, or gate evidence.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint AE: Baseline Checksum Audit Script#
The EEG v1 baseline lane now includes scripts/audit_eeg_v1_baseline_checksums.py, a local
JSON-emitting verifier for baseline_checksum_manifest.json. It fails closed on missing manifests,
invalid JSON, unsupported schemas or algorithms, invalid artifact rows, duplicate paths, missing
artifacts, byte-count changes, and checksum mismatches, so the baseline checksum manifest becomes
a runnable evidence gate.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint AF: Baseline Checksum Audit Command in Smoke Stdout#
The EEG v1 baseline runner now prints the baseline_checksum_manifest.json path and a concrete
checksum_audit_command using the actual --out-dir value from the smoke run. This makes the
baseline checksum gate copy/pasteable from stdout while preserving the local-only lane and leaving
training, ranking, gates, and claim scope unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint AG: Required Baseline Artifact Manifest Coverage#
The EEG v1 baseline checksum audit now requires baseline_checksum_manifest.json to include every
core baseline evidence artifact that the writer always emits: metrics, tables, split audit, gate,
run config, dataset summary, target-scale context, failure reasons, report, and per-subject,
per-channel, and per-horizon metrics. Optional autocorrelation sidecars remain outside this sprint’s
required-entry contract because direct writer calls can omit that diagnostic payload.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint AH: Unexpected Baseline Checksum Manifest Entry Rejection#
The EEG v1 baseline checksum audit now rejects validly check-summed but unexpected entries in
baseline_checksum_manifest.json. The allowlist is bounded to the required core evidence artifacts
plus optional autocorrelation diagnostic JSON/CSV sidecars, so a manifest cannot silently bless extra
files outside the declared local baseline evidence bundle.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint AI: Baseline Verification Sidecar#
The EEG v1 baseline artifact bundle now writes baseline_verification.json, a machine-readable
sidecar declaring the local-only execution lane, a100_jobs_launched=false, the checksum manifest
name, and the exact checksum-audit command for the emitted artifact directory. The sidecar is covered
by baseline_checksum_manifest.json, so local verification instructions are auditable evidence
rather than stdout-only prose.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint AJ: Baseline Verification Sidecar Contract Audit#
The EEG v1 baseline checksum audit now validates baseline_verification.json semantically in
addition to checking its SHA-256 digest. The audit fails if the sidecar no longer declares the
local-only execution lane, a100_jobs_launched=false, the expected checksum manifest, or the exact
checksum-audit command for the artifact directory.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint AK: Baseline Report Artifact Index and Checksum Instructions#
The EEG v1 baseline diagnostic report now includes an Artifact Index table and a Checksum Audit
section near the top of the report. The index names the bounded JSON/CSV/Markdown evidence artifacts,
including baseline_verification.json and baseline_checksum_manifest.json, while the checksum
section gives the local audit command with an explicit <artifact-dir> placeholder.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint AL: Baseline Verification Sidecar in Smoke Stdout#
The EEG v1 baseline runner now prints the baseline_verification.json path and
a100_jobs_launched=false directly in smoke stdout. This keeps the one-command local smoke output
connected to the checksum-covered verification sidecar without changing model training, ranking,
gates, or claim scope.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint AM: Stimulus/Task Label Split Audit Diagnostic#
The EEG v1 autocorrelation diagnostic now audits stimulus/task label overlap when labelled recording metadata exists. It reports observed label keys, train/val/test label sets, overlap status, and failure reasons for the current split; it does not invent a new stimulus-held-out split builder or claim a labelled public benchmark result.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint AN: Stimulus/Task Split Audit Report Visibility#
The EEG v1 diagnostic report now adds a dedicated Stimulus/Task Split Audit section when
labelled stimulus/task metadata was audited. The section surfaces observed label keys,
train/val/test label sets, overlap status, leakage status, and failure reasons in Markdown while
preserving the existing JSON sidecar and generic autocorrelation diagnostics table.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint AO: Stimulus/Task Split Audit Evidence Gate Blocker#
When labelled stimulus/task metadata exists and the stimulus/task split audit detects label overlap, the EEG v1 artifact writer now adds a diagnostic gate failure and blocks the narrow benchmark-readiness claim. Unlabelled synthetic fixture runs remain unchanged; this only tightens the evidence gate when a completed labelled split diagnostic proves overlap.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint AP: Failure Reasons Sidecar Contract Audit#
The EEG v1 baseline checksum audit now validates the structure of failure_reasons.json, requiring
baseline_failures, gate_failures, split_audit_failures, and diagnostic_failures to exist as
lists. This prevents a checksum-updated bundle from silently dropping diagnostic gate-failure
evidence while still keeping the audit local and schema-focused.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint AQ: Failure Reasons Gate Consistency Audit#
The EEG v1 baseline checksum audit now cross-checks failure_reasons.json against
evidence_gate.json by requiring failure_reasons.gate_failures to exactly match
evidence_gate.failure_reasons. This prevents checksum-updated evidence bundles from fabricating or
dropping gate failures in the human-facing failure sidecar.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint AR: Failure Reasons Split-Audit Consistency Audit#
The EEG v1 baseline checksum audit now cross-checks failure_reasons.json against
split_audit.json by requiring failure_reasons.split_audit_failures to exactly match
split_audit.failure_reasons. This prevents checksum-updated evidence bundles from hiding split
audit failures in the human-facing failure sidecar.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a -v
Sprint A.4: Shuffled-Control Separation Gate#
The EEG v1 evidence gate now treats the train-only shuffled-target negative control as a required
separation check, not just a reported diagnostic. The narrow
eeg_future_forecasting_benchmark_ready claim is blocked when the shuffled-target control does not
degrade or when it remains too close to the real short-horizon baseline performance.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_a4_gate_fails_when_shuffled_target_control_stays_too_close -v
Sprint A.5: Model-Win Claim Status#
The EEG v1 evidence bundle now separates benchmark-readiness from a model-performance win. A run can
pass the narrow benchmark-readiness gate while model_win_claim_allowed remains false when
persistence/ridge/autoregressive baselines dominate or Kahlus does not beat the best baseline.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_script_writes_expected_artifacts -v
Sprint A.6: Model-Win Checksum Audit Consistency#
The EEG v1 baseline checksum audit now cross-checks model_win_claim_allowed,
model_win_status, and model_win_claim_failure_reasons between metrics.json and
evidence_gate.json. This prevents a checksum-updated evidence bundle from silently changing the
human-facing metrics artifact into a model-win claim while the gate artifact still blocks it.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_model_win_metrics_gate_mismatch -v
Sprint A.7: Autocorrelation Gate Checksum Consistency#
The EEG v1 baseline checksum audit now cross-checks the gate-driving shuffled-control fields in
autocorrelation_diagnostics.json against evidence_gate.json. A checksum-updated diagnostics
artifact cannot mark shuffled targets unsafe unless the gate also records the matching failure.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_autocorr_gate_mismatch -v
Sprint A.8: Passing-Gate Autocorrelation Manifest Requirement#
The EEG v1 baseline checksum audit now requires autocorrelation_diagnostics.json to be listed in
baseline_checksum_manifest.json whenever evidence_gate.json allows the benchmark-readiness claim
and declares required negative controls. Blocked gates may still omit optional diagnostics.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_requires_autocorr_manifest_entry_when_gate_passes -v
Sprint A.9: Required Negative-Control Row Audit#
The EEG v1 baseline checksum audit now verifies that a passing gate’s required negative controls are
present as completed rows inside autocorrelation_diagnostics.json. A diagnostics file cannot keep
its checksum valid while dropping shuffled_target_control.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_missing_required_autocorr_control -v
Sprint A.10: Train-Only Shuffled-Control Evidence Audit#
The EEG v1 baseline checksum audit now verifies that the required shuffled_target_control row
states shuffle_boundary=train_split_only, train targets shuffled, and validation/test targets not
shuffled. A checksum-updated diagnostics artifact cannot relabel the negative control as all-split
shuffling while the gate still passes.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_non_train_only_shuffled_control -v
Sprint A.11: Shuffled-Control Seed Provenance Audit#
The EEG v1 baseline checksum audit now verifies that the required shuffled_target_control row
includes deterministic seed provenance for the train-only target shuffle. A checksum-updated
diagnostics artifact cannot drop shuffle_seed or shuffle_seed_source while the gate still passes.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_missing_shuffled_control_seed_provenance -v
Sprint A.12: Shuffled-Control Seed Contract Audit#
The EEG v1 baseline checksum audit now cross-checks the required shuffled_target_control
shuffle_seed against run_config.seed + 1701 whenever the row declares
shuffle_seed_source=diagnostic_seed_plus_1701. A checksum-updated diagnostics artifact cannot
replace the deterministic train-shuffle seed while keeping the benchmark-readiness gate passing.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_shuffled_control_seed_run_config_mismatch -v
Sprint A.13: Autocorrelation Summary MSE Integrity Audit#
The EEG v1 baseline checksum audit now cross-checks the displayed autocorrelation summary MSE fields
against the completed diagnostic rows that generated them. A checksum-updated diagnostics artifact
cannot alter short_horizon_best_mse, tiny_ssm_mse, shuffled_target_best_mse, or
shuffled_target_control_mse while leaving the row-level evidence unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_autocorr_summary_mse_row_mismatch -v
Sprint A.14: Autocorrelation Summary Delta Integrity Audit#
The EEG v1 baseline checksum audit now recomputes the reviewer-facing autocorrelation delta fields
from completed diagnostic rows. A checksum-updated diagnostics artifact cannot alter
long_horizon_delta_vs_short, non_overlap_delta_vs_short, or delta_prediction_delta_vs_short
without changing the row-level MSE evidence that produced them.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_autocorr_summary_delta_row_mismatch -v
Sprint A.15: Autocorrelation Summary Boolean Integrity Audit#
The EEG v1 baseline checksum audit now recomputes the reviewer-facing autocorrelation boolean fields
from completed diagnostic rows. A checksum-updated diagnostics artifact cannot alter
shuffled_control_degrades, persistence_or_ridge_dominates, or
shuffled_target_close_to_real_baselines while leaving the row-level evidence unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_autocorr_summary_boolean_row_mismatch -v
Sprint A.16: Model-Win Claim Recompute Audit#
The EEG v1 baseline checksum audit now recomputes model_win_claim_allowed, model_win_status, and
model_win_claim_failure_reasons from the underlying metrics and autocorrelation diagnostics. A
checksum-updated bundle cannot edit both metrics.json and evidence_gate.json consistently to
claim a model win when ridge/persistence-style baselines still dominate.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_coordinated_false_model_win_claim -v
Sprint A.17: Best-Baseline Summary Recompute Audit#
The EEG v1 baseline checksum audit now recomputes best_baseline, best_baseline_mse,
best_baseline_gap, and kahlus_beats_best_baseline from metrics_by_model. A checksum-updated
bundle cannot alter derived baseline-win fields to make the model-win gate look earned while the
underlying model MSEs still show a baseline winning.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_best_baseline_summary -v
Sprint A.18: Baseline Ranking Recompute Audit#
The EEG v1 baseline checksum audit now recomputes baseline_ranking from metrics_by_model. A
checksum-updated bundle cannot reorder or rewrite reviewer-facing ranking rows while leaving the
underlying model MSE evidence unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_baseline_ranking -v
Sprint A.19: Baseline Table JSON Recompute Audit#
The EEG v1 baseline checksum audit now recomputes baseline_table.json rows and ranking from
metrics_by_model. A checksum-updated bundle cannot rewrite the reviewer-facing baseline table JSON
while leaving the underlying model metrics unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_baseline_table_json -v
Sprint A.20: Baseline Table CSV Recompute Audit#
The EEG v1 baseline checksum audit now recomputes baseline_table.csv rows from
metrics_by_model. A checksum-updated bundle cannot rewrite the reviewer-facing CSV table while
leaving the underlying model metrics unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_baseline_table_csv -v
Sprint A.21: Metrics CSV Recompute Audit#
The EEG v1 baseline checksum audit now recomputes metrics.csv rows from metrics_by_model.
A checksum-updated bundle cannot rewrite the reviewer-facing metric CSV while leaving the structured
metrics artifact unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_metrics_csv -v
Sprint A.22: Granular Metrics CSV Recompute Audit#
The EEG v1 baseline checksum audit now recomputes per_subject_metrics.csv,
per_channel_metrics.csv, and per_horizon_metrics.csv from the corresponding structured fields
in metrics.json. A checksum-updated bundle cannot rewrite reviewer-facing granular metric tables
while leaving the structured metrics artifact unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_granular_metric_csvs -v
Sprint A.23: Autocorrelation Diagnostics CSV Recompute Audit#
The EEG v1 baseline checksum audit now recomputes autocorrelation_diagnostics.csv from
autocorrelation_diagnostics.json. A checksum-updated bundle cannot rewrite the reviewer-facing
autocorrelation diagnostic table while leaving the structured diagnostic artifact unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_autocorr_diagnostics_csv -v
Sprint A.24: Diagnostic Report Claim-Line Audit#
The EEG v1 baseline checksum audit now checks diagnostic_report.md against metrics.json and
evidence_gate.json for the report’s visible claim scope, scientific-claim status, best baseline,
and Kahlus-vs-best-baseline result. A checksum-updated bundle cannot rewrite those human-facing
report lines while leaving the structured evidence unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_diagnostic_report_claim_lines -v
Sprint A.25: Target Scale Context Recompute Audit#
The EEG v1 baseline checksum audit now recomputes each model’s target_scale_context.json RMSE,
RMSE-relative-to-target-std, and MSE-relative-to-target-variance fields from metrics.json plus the
recorded target scale. A checksum-updated bundle cannot make normalized EEG MSE look smaller relative
to target scale while leaving the structured metrics unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_target_scale_context -v
Sprint A.26: Dataset Summary Consistency Audit#
The EEG v1 baseline checksum audit now checks dataset_summary.json against metrics.json,
split_audit.json, and run_config.json for dataset/source/status, split type, subject counts, and
run window settings. A checksum-updated bundle cannot relabel a synthetic fixture as a public
benchmark or alter held-out split context while leaving the structured evidence unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_dataset_summary -v
Sprint A.27: Run Config Consistency Audit#
The EEG v1 baseline checksum audit now checks run_config.json against metrics.json,
evidence_gate.json, and dataset_summary.json for dataset/source/status, claim scope, configured
model coverage, selection policy, and run window settings. A checksum-updated bundle cannot hide a
required completed baseline, broaden the claim scope, or relabel the run configuration while leaving
the structured evidence unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_run_config -v
Sprint A.28: Split Audit Internal Consistency#
The EEG v1 baseline checksum audit now recomputes subject overlap from split_audit.json subject
lists and checks that subject_overlap and leakage_passed agree with the listed subjects and
failure reasons. A checksum-updated bundle cannot hide train/validation/test subject overlap inside
the split audit artifact while keeping the same subject counts.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_split_audit_subject_overlap -v
Sprint A.29: Target Scale Variance Consistency Audit#
The EEG v1 baseline checksum audit now checks that target_scale_context.json keeps
target_variance consistent with target_std ** 2. A checksum-updated bundle cannot distort the
normalized-MSE scale denominator while also rewriting the dependent per-model variance ratios.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_target_scale_variance_std_mismatch -v
Sprint A.30: Evidence Gate Finite-Metrics Consistency Audit#
The EEG v1 baseline checksum audit now recomputes the finite_metrics gate field from
metrics.json. A checksum-updated bundle cannot leave evidence_gate.json claiming a clean
benchmark-readiness gate while disagreeing with the actual metric finiteness payload.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_gate_finite_metrics_mismatch -v
Sprint A.31: Evidence Gate Baseline-Table Consistency Audit#
The EEG v1 baseline checksum audit now recomputes the baseline_table_present gate field from
metrics.json. A checksum-updated bundle cannot make the gate disagree with whether completed
baseline results exist.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_gate_baseline_table_present_mismatch -v
Sprint A.32: Evidence Gate Split-Audit Consistency Audit#
The EEG v1 baseline checksum audit now checks evidence_gate.json["split_audit_passed"] against
split_audit.json["leakage_passed"]. A checksum-updated bundle cannot make the gate disagree with
the subject-held-out leakage audit result.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_gate_split_audit_passed_mismatch -v
Sprint A.33: Evidence Gate Dataset Consistency Audit#
The EEG v1 baseline checksum audit now checks evidence_gate.json["dataset"] against
metrics.json["dataset"]. A checksum-updated bundle cannot relabel a synthetic fixture gate as a
different dataset while leaving the structured metrics unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_gate_dataset_mismatch -v
Sprint A.34: Evidence Gate Branch Consistency Audit#
The EEG v1 baseline checksum audit now checks evidence_gate.json["branch"] == "v1". A
checksum-updated bundle cannot relabel a v1 EEG benchmark-readiness gate as another branch.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_gate_branch_mismatch -v
Sprint A.35: Evidence Gate Claim-Scope Consistency Audit#
The EEG v1 baseline checksum audit now checks evidence_gate.json["claim_scope"] against
eeg_future_forecasting_benchmark_ready. A checksum-updated bundle cannot relabel the v1 EEG
benchmark-readiness gate as diagnosis, treatment, or another broader claim scope.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_gate_claim_scope_mismatch -v
Sprint A.36: Evidence Gate Allowed-Claim-Scope Consistency Audit#
The EEG v1 baseline checksum audit now checks
evidence_gate.json["gate_criteria"]["allowed_claim_scope"] against
eeg_future_forecasting_benchmark_ready. A checksum-updated bundle cannot keep the top-level v1
claim narrow while broadening the nested gate criteria shown to reviewers.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_gate_allowed_claim_scope_mismatch -v
Sprint A.37: Evidence Gate Required-Control Criteria Consistency Audit#
The EEG v1 baseline checksum audit now checks
evidence_gate.json["gate_criteria"]["required_first_class_baselines"] and
evidence_gate.json["gate_criteria"]["required_negative_controls"] against the required TinySSM
baseline and shuffled-target control. A checksum-updated bundle cannot hide the baseline/control
requirements from the reviewer-facing gate criteria.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_gate_required_control_criteria_mismatch -v
Sprint A.38: Evidence Gate Shuffled-Control Criteria Consistency Audit#
The EEG v1 baseline checksum audit now checks
evidence_gate.json["gate_criteria"]["requires_shuffled_target_degradation"] and
evidence_gate.json["gate_criteria"]["requires_shuffled_target_not_close_to_real_baselines"]
against the required shuffled-target safeguards. A checksum-updated bundle cannot soften the
negative-control criteria while keeping the report internally consistent.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_gate_shuffled_control_criteria_mismatch -v
Sprint A.39: Evidence Gate Core Criteria Consistency Audit#
The EEG v1 baseline checksum audit now checks the core reviewer-facing gate criteria: minimum forecast horizon, allowed split types, split-audit requirement, baseline-table requirement, finite-metrics requirement, and calibration requirement. A checksum-updated bundle cannot soften these criteria while keeping the report internally consistent.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_gate_core_criteria_mismatch -v
Sprint A.40: Diagnostic Report Gate-Criteria Consistency Audit#
The EEG v1 baseline checksum audit now checks the gate-criteria lines in diagnostic_report.md
against evidence_gate.json. A checksum-updated bundle cannot keep structured gate JSON intact
while showing softened criteria to reviewers in the Markdown report.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_diagnostic_report_gate_criteria -v
Sprint A.41: Diagnostic Report Model-Win Consistency Audit#
The EEG v1 baseline checksum audit now checks the model-win lines and model-win failure reasons
in diagnostic_report.md against evidence_gate.json. A checksum-updated bundle cannot keep the
structured gate blocked while making the Markdown report claim that Kahlus beat the baselines.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_diagnostic_report_model_win -v
Sprint A.42: Diagnostic Report Target-Scale Consistency Audit#
The EEG v1 baseline checksum audit now checks the target-scale lines in diagnostic_report.md
against target_scale_context.json. A checksum-updated bundle cannot keep the structured target
scale intact while showing misleading normalized-MSE scale context in the Markdown report.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_diagnostic_report_target_scale -v
Sprint A.43: Diagnostic Report Target-Units Consistency Audit#
The EEG v1 baseline checksum audit coverage now explicitly tampers target_units in
diagnostic_report.md and requires the audit to reject the bundle against
target_scale_context.json. A checksum-updated bundle cannot make normalized fixture MSE look like
raw microvolt-scale evidence in the Markdown report.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_diagnostic_report_target_scale -v
Sprint A.44: Diagnostic Report Autocorrelation Summary Consistency Audit#
The EEG v1 baseline checksum audit now checks reviewer-facing autocorrelation summary lines in
diagnostic_report.md against autocorrelation_diagnostics.json. A checksum-updated bundle cannot
hide ridge/persistence dominance or make shuffled-target controls look safer in the Markdown report.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_diagnostic_report_autocorr_summary -v
Sprint A.45: Diagnostic Report Autocorrelation Verdict Consistency Audit#
The EEG v1 baseline checksum audit now also checks the autocorrelation control deltas,
shuffled_control_degrades, and verdict lines in diagnostic_report.md against
autocorrelation_diagnostics.json. A checksum-updated bundle cannot preserve the structured
diagnostics while softening the reviewer-facing autocorrelation verdict.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_diagnostic_report_autocorr_summary -v
Sprint A.46: Diagnostic Report Autocorrelation Row Consistency Audit#
The EEG v1 baseline checksum audit now checks each autocorrelation diagnostic table row in
diagnostic_report.md against autocorrelation_diagnostics.json. A checksum-updated bundle cannot
keep the structured shuffled-target control evidence intact while softening the Markdown row that
reviewers read.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_diagnostic_report_autocorr_row -v
Sprint A.47: Diagnostic Report Artifact-Index Consistency Audit#
The EEG v1 baseline checksum audit now checks the artifact-index rows in diagnostic_report.md.
A checksum-updated bundle cannot keep autocorrelation diagnostics in the machine artifacts while
hiding them from the reviewer-facing artifact list.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_diagnostic_report_artifact_index -v
Sprint A.48: Diagnostic Report Verification-Lane Consistency Audit#
The EEG v1 baseline report now prints the local-only execution lane and a100_jobs_launched value
from baseline_verification.json, and the checksum audit rejects Markdown-only tampering. A
checksum-updated bundle cannot make the reviewer-facing report imply that A100 work was launched.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_diagnostic_report_verification_lines -v
Sprint A.49: Diagnostic Report Method-Order Consistency Audit#
The EEG v1 baseline checksum audit now checks the Method Order rows in diagnostic_report.md
against run_config.json. A checksum-updated bundle cannot make the reviewer-facing report show
the main model before baselines while the replay config remains baseline-first.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_diagnostic_report_method_order -v
Sprint A.50: Diagnostic Report Baseline-Ranking Consistency Audit#
The EEG v1 baseline checksum audit now checks the Baseline Ranking rows in diagnostic_report.md
against metrics.json. A checksum-updated bundle cannot make the reviewer-facing ranking claim a
different best model while the structured baseline ranking remains unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_diagnostic_report_baseline_ranking -v
Sprint A.51: Diagnostic Report Baseline-Gap Consistency Audit#
The EEG v1 baseline checksum audit now checks the Baseline Gap Summary lines in
diagnostic_report.md against metrics.json. A checksum-updated bundle cannot soften the
reviewer-facing persistence/ridge/best-baseline gap while keeping structured metrics unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_diagnostic_report_baseline_gaps -v
Sprint A.52: Diagnostic Report Duplicate Protected-Line Audit#
The EEG v1 baseline checksum audit now rejects duplicate protected lines in diagnostic_report.md.
A checksum-updated bundle cannot keep the correct structured line while appending a second
reviewer-facing claim or baseline line that creates ambiguity.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_duplicate_diagnostic_report_claim_lines -v
Sprint A.53: Diagnostic Report Model-Win Reason List Audit#
The EEG v1 baseline checksum audit now checks the rendered model_win_claim_failure_reasons
bullet list in diagnostic_report.md exactly against evidence_gate.json. A checksum-updated
bundle cannot append extra reviewer-facing model-win reasoning while keeping the gate unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_extra_diagnostic_report_model_win_reason -v
Sprint A.54: Diagnostic Report Target-Scale Model Row Audit#
The EEG v1 baseline checksum audit now checks each target-scale model row in diagnostic_report.md
against target_scale_context.json. A checksum-updated bundle cannot make normalized-MSE scale
context look better in the reviewer-facing report while keeping the structured artifact unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_diagnostic_report_target_scale_model_row -v
Sprint A.55: Diagnostic Report Target-Scale Row-List Audit#
The EEG v1 baseline checksum audit now checks the full target-scale model-row list in
diagnostic_report.md against target_scale_context.json. A checksum-updated bundle cannot append
extra reviewer-facing target-scale rows while keeping the structured artifact unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_extra_diagnostic_report_target_scale_model_row -v
Sprint A.56: Diagnostic Report Metric-Breakdown Count Audit#
The EEG v1 baseline checksum audit now checks Metric Breakdown Summary row counts in
diagnostic_report.md against metrics.json. A checksum-updated bundle cannot hide or inflate the
reviewer-facing per-subject/channel/horizon sidecar coverage counts while keeping structured
metrics unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_diagnostic_report_metric_breakdown_counts -v
Sprint A.57: Diagnostic Report Metric Sidecar List Audit#
The EEG v1 baseline checksum audit now checks the detailed_sidecars line in
diagnostic_report.md. A checksum-updated bundle cannot point reviewers away from the
per-subject/channel/horizon metric sidecars while keeping those artifacts unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_diagnostic_report_metric_sidecars -v
Sprint A.58: Diagnostic Report Run-Config Line Audit#
The EEG v1 baseline checksum audit now checks the Run Config section in
diagnostic_report.md against run_config.json. A checksum-updated bundle cannot change
reviewer-facing seed, model list, windowing, data-source, selection-policy, or claim-scope lines
while keeping the structured run config unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_diagnostic_report_run_config_lines -v
Sprint A.59: Diagnostic Report Method-Order Row-List Audit#
The EEG v1 baseline checksum audit now checks the full Method Order table in
diagnostic_report.md against run_config.json. A checksum-updated bundle cannot append extra
reviewer-facing model rows while keeping the configured model list unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_extra_diagnostic_report_method_order_row -v
Sprint A.60: Diagnostic Report Baseline-Ranking Row-List Audit#
The EEG v1 baseline checksum audit now checks the full Baseline Ranking table in
diagnostic_report.md against metrics.json. A checksum-updated bundle cannot append extra
reviewer-facing ranking rows while keeping the structured metrics unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_extra_diagnostic_report_baseline_ranking_row -v
Sprint A.61: Diagnostic Report Artifact-Index Row-List Audit#
The EEG v1 baseline checksum audit now checks the full Artifact Index table in
diagnostic_report.md against the canonical baseline artifact list. A checksum-updated bundle
cannot append extra reviewer-facing artifact rows while keeping the actual evidence manifest
unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_extra_diagnostic_report_artifact_index_row -v
Sprint A.62: Diagnostic Report Autocorrelation Row-List Audit#
The EEG v1 baseline checksum audit now checks the full autocorrelation diagnostics table in
diagnostic_report.md against autocorrelation_diagnostics.json. A checksum-updated bundle
cannot append extra reviewer-facing autocorrelation/control rows while keeping the structured
diagnostics unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_extra_diagnostic_report_autocorr_row -v
Sprint A.63: Diagnostic Report Metric-Breakdown Section Audit#
The EEG v1 baseline checksum audit now checks the full Metric Breakdown Summary section in
diagnostic_report.md against metrics.json. A checksum-updated bundle cannot append extra
reviewer-facing sidecar/count lines while keeping the structured metrics unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_extra_diagnostic_report_metric_breakdown_line -v
Sprint A.64: Diagnostic Report Gate-Criteria Section Audit#
The EEG v1 baseline checksum audit now checks the full Evidence Gate Criteria section in
diagnostic_report.md against evidence_gate.json. A checksum-updated bundle cannot append extra
reviewer-facing gate criteria while keeping the structured gate unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_extra_diagnostic_report_gate_criteria_line -v
Sprint A.65: Diagnostic Report Target-Scale Header Audit#
The EEG v1 baseline checksum audit now checks the target-scale header bullets in
diagnostic_report.md against target_scale_context.json, including the scale note. A
checksum-updated bundle cannot append extra reviewer-facing target-scale interpretation lines while
keeping the structured scale context unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_extra_diagnostic_report_target_scale_header_line -v
Sprint A.66: Diagnostic Report Checksum-Audit Section Audit#
The EEG v1 baseline checksum audit now checks the full Checksum Audit section in
diagnostic_report.md against baseline_verification.json and the canonical local audit command.
A checksum-updated bundle cannot append extra reviewer-facing cluster-readiness or verification
claims while keeping the structured verification sidecar unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_extra_diagnostic_report_checksum_audit_line -v
Sprint A.67: Diagnostic Report Claim-Boundaries Section Audit#
The EEG v1 baseline checksum audit now checks the full Claim Boundaries section in
diagnostic_report.md against the narrow v1 benchmark-readiness claim boundary. A
checksum-updated bundle cannot append reviewer-facing clinical, SOTA, foundation-model, v2, or v3
claims while keeping the structured evidence artifacts unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_extra_diagnostic_report_claim_boundary_line -v
Sprint A.68: Diagnostic Report Summary-Header Section Audit#
The EEG v1 baseline checksum audit now checks the top summary block in diagnostic_report.md
against metrics.json, evidence_gate.json, and split_audit.json. A checksum-updated bundle
cannot append extra reviewer-facing model-win or claim-scope lines before the artifact index while
keeping the structured evidence artifacts unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_extra_diagnostic_report_summary_header_line -v
Sprint A.69: Diagnostic Report Baseline-Gap Section Audit#
The EEG v1 baseline checksum audit now checks the full Baseline Gap Summary section in
diagnostic_report.md against metrics.json. A checksum-updated bundle cannot append extra
reviewer-facing calibrated-gap or model-win lines while keeping the structured baseline gaps
unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_extra_diagnostic_report_baseline_gap_line -v
Sprint A.70: Diagnostic Report Autocorrelation-Summary Row Audit#
The EEG v1 baseline checksum audit now checks the full autocorrelation Summary table in
diagnostic_report.md against autocorrelation_diagnostics.json. A checksum-updated bundle cannot
append extra reviewer-facing autocorrelation-cleared rows while keeping the structured diagnostics
unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_extra_diagnostic_report_autocorr_summary_row -v
Sprint A.71: Diagnostic Report Autocorrelation-Dominance Bullet Audit#
The EEG v1 baseline checksum audit now checks the Baseline Dominance bullets in
diagnostic_report.md against autocorrelation_diagnostics.json. A checksum-updated bundle cannot
append extra reviewer-facing autocorrelation-cleared dominance lines while keeping the structured
diagnostics unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_extra_diagnostic_report_autocorr_dominance_line -v
Sprint A.72: Diagnostic Report Autocorrelation Intro Audit#
The EEG v1 baseline checksum audit now checks the opening autocorrelation warning and caveat in
diagnostic_report.md against autocorrelation_diagnostics.json plus the fixed claim-hygiene
caveat. A checksum-updated bundle cannot append reviewer-facing autocorrelation-cleared prose before
the summary table while keeping the structured diagnostics unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_extra_diagnostic_report_autocorr_intro_line -v
Sprint A.73: Diagnostic Report Model-Win Section Audit#
The EEG v1 baseline checksum audit now checks the full Model Win Claim Status section in
diagnostic_report.md against evidence_gate.json. A checksum-updated bundle cannot append extra
reviewer-facing calibrated model-win override lines while keeping the structured gate unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_extra_diagnostic_report_model_win_line -v
Sprint A.74: Failure-Reasons Sidecar Consistency Audit#
The EEG v1 baseline checksum audit now checks failure_reasons.json baseline and diagnostic failure
rows against the authoritative metrics and autocorrelation artifacts. A checksum-updated bundle
cannot hide baseline runner failures or missing-control diagnostic failures while leaving the source
artifacts unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_failure_reasons_baseline_and_diagnostic_mismatch -v
Sprint A.75: Autocorrelation Row Metric Consistency Audit#
The EEG v1 baseline checksum audit now checks each completed autocorrelation diagnostic row so its
published persistence_mse, linear_ridge_mse, tiny_ssm_mse, best_model, and best_mse fields
match the nested metrics_by_model values. A checksum-updated bundle cannot make TinySSM or the
winning autocorrelation baseline look different in summary fields than in the structured row metrics.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_autocorr_row_metric_field_mismatch -v
Sprint A.76: Diagnostic Report Failure-Section Audit#
The EEG v1 baseline checksum audit now checks the Gate Failures, Split Audit Failures, and
Baseline Failures sections in diagnostic_report.md against the structured gate, split-audit, and
metrics failure payloads. A checksum-updated bundle cannot append reviewer-facing failure lines that
do not exist in the source artifacts.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_extra_diagnostic_report_failure_lines -v
Sprint A.77: Diagnostic Report Stimulus/Task Split Audit#
The EEG v1 baseline checksum audit now checks the Stimulus/Task Split Audit section in
diagnostic_report.md against the structured stimulus_task_held_out_split diagnostic row. A
checksum-updated bundle cannot add reviewer-facing stimulus/task leakage overrides while keeping the
structured diagnostic unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_extra_diagnostic_report_stimulus_task_audit_line -v
Sprint A.78: HBN Local Boundary Report Audit#
The EEG v1 baseline checksum audit now checks the HBN Local Path Boundary section in
diagnostic_report.md whenever benchmark_status is local_manifest_not_public_hbn_benchmark.
A checksum-updated bundle cannot turn local HBN-style manifest evidence into reviewer-facing public
HBN benchmark evidence.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_hbn_local_boundary -v
Sprint A.79: Dataset Summary Granular Count Audit#
The EEG v1 baseline checksum audit now checks dataset_summary.json test-window and channel counts
against per_subject_metrics.csv and per_channel_metrics.csv. A checksum-updated bundle cannot
inflate reviewer-facing test-window or channel counts while leaving granular metric sidecars unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_dataset_summary_granular_count_mismatch -v
Sprint A.80: Diagnostic Report Dataset Summary Audit#
The EEG v1 diagnostic report now includes a bounded Dataset Summary section sourced from
dataset_summary.json, and the checksum audit checks that section exactly. A checksum-updated bundle
cannot inflate reviewer-facing subject, window, channel, or window-geometry counts in Markdown while
leaving the structured dataset summary unchanged.
Verification:
PYTHONPATH=src python3 -m unittest tests.eeg_v1.test_eeg_v1_sprint_a.EEGV1SprintATests.test_baseline_checksum_audit_rejects_tampered_diagnostic_report_dataset_summary -v