feat(review): add fail-closed local review cascade

NeuralComposeEEG/ROADMAP.md

@@ -22,6 +22,36 @@ The supporting [mathematics, physics, and methods
 scope](../docs/scoping/eeg-mathematics-physics-methods-scope.md) separates the
 linear-algebra, signal-processing, and evaluation foundations needed now from
 later electroquasistatic, sensor-fusion, optimization, and shadow-policy work.
+The [SVD decision memo](../docs/research/svd-four-channel-eeg-decision-memo_v1.md)
+and [proposed SVD experiment roadmap](../docs/scoping/svd-eeg-experiment-roadmap.md)
+classify singular-spectrum diagnostics and train-only feature reduction as
+separate, non-runtime studies. At D0 they are foundational only: they do not
+modify the fixed encoder conditions or authorize a physical-data claim.
+Their contracts are [SVD diagnostics](experiments/EXP-NC-SVD-DIAG-001.md),
+[M0 feature reduction](experiments/EXP-NC-SVD-M0-001.md),
+[representation analysis](experiments/EXP-NC-SVD-REP-001.md), and the deferred
+[synthetic inverse study](experiments/EXP-NC-SVD-INV-001.md).
+
+The separate [`EXP-FUNC-SYN-000` decision
+memo](../docs/research/function-space-foundations-decision-memo-v0.md) and
+[contract](../configs/function-space-foundations-v0.json) define a
+synthetic-only D0 mathematical-foundations package. Its empirical-measure,
+derivative, Sobolev, convergence, operator, and mixed-integration fixtures do
+not read EEG, alter preprocessing, change M0-M4, or delay physical Muse
+capture. Every result remains `insufficient_evidence` and `not_eligible`.
+
+The [JEPA decision
+memo](../docs/research/jepa-four-channel-eeg-decision-memo-v0.md) and proposed
+[`EXP-NC-EEG-JEPA-001`](experiments/EXP-NC-EEG-JEPA-001.md) define a separate
+Pass 1.5 encoder-objective comparison. At D0 this is documentation and
+deterministic diagnostic-contract work only. It does not alter M0-M4, authorize
+Laya download or training, or create a live JEPA path. D2 may permit only a
+separately preregistered tiny pipeline-evidence run; D3 is the first scientific
+comparison and still requires an approved pretraining-data gate. The registered
+[collapse diagnostics](../docs/scoping/jepa-collapse-diagnostics-v0.json)
+measure representation geometry directly, but cannot support an encoder
+without the existing grouped predictive, calibration, artifact, and robustness
+evidence.
 
 Gate to encoder confirmation: protocol-complete sessions on multiple days,
 then M0, M1, M2, M3, and eventually M4 on the same session-grouped folds.
@@ -33,6 +63,34 @@ without the shared fixed-compute configuration hash and retains runtime evidence
 separately from predictive outcomes. Core ML conversion is deferred until a
 compact candidate is selected by this offline evidence.
 
+## EXP-NC-EEG-JEPA-001
+
+Proposed Pass 1.5 work only. It compares matched reconstruction, latent
+prediction, anti-collapse, and bounded hybrid objectives after the ordinary
+encoder program reaches the applicable gate. D2 can establish pipeline
+execution only under a separate preregistration. A scientific comparison waits
+until D3 and a separately reviewed corpus is adequate for self-supervised
+learning. Short pilot captures are not enough merely because their labels can
+be hidden. Every local pretraining fit remains inside its outer
+training-session partition.
+
+This is not the existing action-conditioned `WorldModel/` JEPA spike. An
+encoder representation does not become a transition model, MPC objective, or
+dialogue controller. Every JEPA result remains `insufficient_evidence`,
+`promotion_status: not_eligible`, and `runtime_change: none`.
+
+### Synthetic JEPA rehearsal
+
+[`EXP-NC-EEG-JEPA-SYN-000`](experiments/EXP-NC-EEG-JEPA-SYN-000.md) is the
+separate executable path for deterministic S0-S6 objective, control, collapse,
+grouping, and artifact checks. Its
+[generator contract](../docs/scoping/jepa-synthetic-generators-v0.json)
+prohibits fallback acquisition data and physical claims. The optional
+[mode-stratified extension](experiments/EXP-NC-EEG-JEPA-SYN-MODE-000.md)
+treats mirror, focus, reflective, and contemplative only as externally assigned
+synthetic regimes. Neither rehearsal changes the D0 status of the physical
+JEPA experiment or sets physical-study thresholds.
+
 ## EXP-NC-ARC-XFER-001
 
 Deferred until the encoder experiment selects a *fixed* shadow-only state

NeuralComposeEEG/configs/jepa-synthetic-v0.json

@@ -0,0 +1,140 @@
+{
+  "schema_version": "nc-eeg-jepa-synthetic-config-v0",
+  "experiment_id": "EXP-NC-EEG-JEPA-SYN-000",
+  "mode_experiment_id": "EXP-NC-EEG-JEPA-SYN-MODE-000",
+  "seed": 42,
+  "device": "cpu",
+  "source": {
+    "schema_version": "nc-eeg-jepa-synthetic-source-v0",
+    "source_type": "deterministic_synthetic_fixture",
+    "device_id": "synthetic-generator",
+    "participant_id": null,
+    "physical_capture_eligible": false,
+    "fallback_capture_allowed": false
+  },
+  "data": {
+    "sample_rate_hz": 64,
+    "channel_count": 4,
+    "window_samples": 64,
+    "windows_per_session": 24,
+    "sessions_per_generator": 6,
+    "generator_ids": ["S0", "S1", "S2", "S3", "S4", "S5", "S6"]
+  },
+  "split": {
+    "unit": "complete_synthetic_session",
+    "outer_test_sessions": 2,
+    "inner_validation_sessions": 1,
+    "outer_fold_count": 2,
+    "normalization": "fit_training_sessions_only",
+    "adjacent_window_split_allowed": false
+  },
+  "mask": {
+    "fraction": 0.25,
+    "geometry": "contiguous_temporal_span",
+    "same_schedule_across_conditions": true
+  },
+  "model": {
+    "architecture_id": "synthetic-four-channel-jepa-v0",
+    "encoder_hidden_dim": 16,
+    "latent_dim": 8,
+    "predictor_hidden_dim": 24,
+    "mode_context_dim": 4,
+    "decoder_present_in_every_condition": true
+  },
+  "training": {
+    "optimizer": "adamw",
+    "learning_rate": 0.003,
+    "weight_decay": 0.0001,
+    "batch_size": 32,
+    "steps": 12,
+    "ema_tau": 0.95,
+    "gradient_clip_norm": 5.0,
+    "conditions": ["T0", "T1", "T2", "T3", "T4", "T5"],
+    "matched_fields": [
+      "encoder_architecture",
+      "parameter_count",
+      "generated_samples",
+      "batch_size",
+      "steps",
+      "optimizer",
+      "seed",
+      "context_target_geometry",
+      "mask_schedule",
+      "probe_budget"
+    ]
+  },
+  "loss": {
+    "T0": {
+      "objective": "seeded_random_encoder",
+      "train": false
+    },
+    "T1": {
+      "objective": "masked_reconstruction",
+      "reconstruction_weight": 1.0
+    },
+    "T2": {
+      "objective": "latent_prediction",
+      "prediction_weight": 1.0
+    },
+    "T3": {
+      "objective": "latent_prediction_plus_sigreg",
+      "prediction_weight": 1.0,
+      "sigreg_weight": 0.2,
+      "regularization_scope": "projector_embedding"
+    },
+    "T4": {
+      "objective": "latent_prediction_plus_vicreg",
+      "prediction_weight": 1.0,
+      "vicreg_variance_weight": 0.2,
+      "vicreg_covariance_weight": 0.02,
+      "regularization_scope": "projector_embedding"
+    },
+    "T5": {
+      "objective": "latent_prediction_plus_bounded_reconstruction",
+      "prediction_weight": 1.0,
+      "reconstruction_weight": 0.2,
+      "maximum_reconstruction_fraction": 0.25
+    }
+  },
+  "controls": [
+    "temporal_target_permutation",
+    "different_sequence_target",
+    "channel_order_shuffle",
+    "channel_coordinate_shuffle",
+    "mask_location_shuffle",
+    "constant_embedding",
+    "nonfinite_input"
+  ],
+  "probe": {
+    "type": "logistic_regression",
+    "regularization_c": 1.0,
+    "max_iterations": 200,
+    "fit_scope": "training_sessions_only"
+  },
+  "diagnostics": {
+    "variance_floor": 1e-06,
+    "singular_value_floor": 1e-08,
+    "energy_rank_threshold": 0.99,
+    "constant_output_tolerance": 1e-07,
+    "pairwise_sample_limit": 256
+  },
+  "mode_extension": {
+    "enabled": true,
+    "modes": ["mirror", "focus", "reflective", "contemplative"],
+    "sessions_per_mode": 4,
+    "controls": ["C0", "C1", "C2", "C3", "C4", "C5"],
+    "training_condition": "T3",
+    "split_unit": "complete_synthetic_session",
+    "leave_one_mode_out": true,
+    "mode_is_externally_assigned": true,
+    "cognitive_mode_inference_allowed": false
+  },
+  "artifact": {
+    "schema_version": "nc-eeg-jepa-synthetic-rehearsal-v0",
+    "decision": "pipeline_evidence_only",
+    "promotion_status": "not_eligible",
+    "runtime_change": "none",
+    "physical_eeg_used": false,
+    "scientific_transfer_claim_allowed": false
+  }
+}