Add continual-shift benchmark for retention vs adaptation

Author: OptimumAF

CHANGELOG.md

@@ -46,6 +46,11 @@ for versioning even while in research-stage development.
   - `.github/CODEOWNERS`
   - `docs/model-card.md`
   - `docs/figures/README.md`
+- Continual-shift comparison benchmark for retention vs adaptation:
+  - `src/app/continual_shift_benchmark.py`
+  - `scripts/run_continual_shift_benchmark.py`
+  - `tests/test_continual_shift_benchmark.py`
+  - shifted/rotated dataset support in `src/infra/datasets.py`
 
 ### Changed
 

README.md

@@ -164,6 +164,12 @@ Toy baseline with review-driven circadian controls:
 python predictive_coding_experiment.py --adaptive-sleep-trigger --adaptive-sleep-budget --reward-modulated-learning --reward-scale-min 0.8 --reward-scale-max 1.4
 ```
 
+Continual shift stress test (retention vs adaptation):
+
+```powershell
+python scripts/run_continual_shift_benchmark.py --profile strength-case --seeds 3,7,11,19,23,31,37
+```
+
 ResNet benchmark (all 3 models):
 
 ```powershell

docs/benchmarks/benchmark_continual_shift_strength_case_2026-02-28.txt

@@ -0,0 +1,10 @@
+Continual Shift Benchmark
+-------------------------
+Phase A trains on base distribution; phase B trains on shifted/rotated distribution.
+Seeds: [3, 7, 11, 19, 23, 31, 37]
+Phase B transform: rotation=40.0 deg, translation=(0.90, -0.70)
+Phase B train fraction: 0.14
+
+Backprop: A_pre=0.975+/-0.012, A_post=0.960+/-0.022, B_post=0.933+/-0.026, retention=0.985+/-0.026, balanced=0.946+/-0.016
+Predictive coding: A_pre=0.970+/-0.016, A_post=0.967+/-0.016, B_post=0.927+/-0.024, retention=0.997+/-0.020, balanced=0.947+/-0.008
+Circadian predictive coding: A_pre=0.975+/-0.015, A_post=0.968+/-0.009, B_post=0.930+/-0.023, retention=0.993+/-0.008, balanced=0.949+/-0.010, sleep_events=5.00, splits=5.00, prunes=0.00, hidden_end=17.00

scripts/run_continual_shift_benchmark.py

@@ -0,0 +1,113 @@
+"""Run continual-shift benchmark across backprop, predictive coding, and circadian PC."""
+
+from __future__ import annotations
+
+import argparse
+from pathlib import Path
+import sys
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+if str(REPO_ROOT) not in sys.path:
+    sys.path.insert(0, str(REPO_ROOT))
+
+from src.app.continual_shift_benchmark import (
+    ContinualShiftConfig,
+    format_continual_shift_benchmark,
+    run_continual_shift_benchmark,
+)
+from src.core.circadian_predictive_coding import CircadianConfig
+
+
+def build_parser() -> argparse.ArgumentParser:
+    """Build CLI parser for continual shift benchmark."""
+    parser = argparse.ArgumentParser(
+        description="Run phase-A/phase-B continual-shift benchmark for all three models."
+    )
+    parser.add_argument("--seeds", type=str, default="3,7,11,19,23,31,37")
+    parser.add_argument(
+        "--profile",
+        type=str,
+        choices=["baseline", "strength-case"],
+        default="strength-case",
+        help="Circadian profile: baseline uses defaults, strength-case emphasizes replay/splits.",
+    )
+    parser.add_argument("--sample-count-phase-a", type=int, default=500)
+    parser.add_argument("--sample-count-phase-b", type=int, default=500)
+    parser.add_argument("--phase-b-train-fraction", type=float, default=0.14)
+    parser.add_argument("--phase-a-epochs", type=int, default=110)
+    parser.add_argument("--phase-b-epochs", type=int, default=80)
+    parser.add_argument("--hidden-dim", type=int, default=12)
+    parser.add_argument("--phase-a-noise-scale", type=float, default=0.8)
+    parser.add_argument("--phase-b-noise-scale", type=float, default=1.0)
+    parser.add_argument("--phase-b-rotation-degrees", type=float, default=40.0)
+    parser.add_argument("--phase-b-translation-x", type=float, default=0.9)
+    parser.add_argument("--phase-b-translation-y", type=float, default=-0.7)
+    parser.add_argument("--sleep-interval-phase-a", type=int, default=40)
+    parser.add_argument("--sleep-interval-phase-b", type=int, default=8)
+    parser.add_argument("--output-file", type=str, default="")
+    return parser
+
+
+def main() -> None:
+    """Run CLI entrypoint."""
+    parser = build_parser()
+    args = parser.parse_args()
+
+    seeds = _parse_int_list(args.seeds)
+    circadian_config = (
+        _build_strength_case_circadian_config()
+        if args.profile == "strength-case"
+        else CircadianConfig()
+    )
+    config = ContinualShiftConfig(
+        sample_count_phase_a=args.sample_count_phase_a,
+        sample_count_phase_b=args.sample_count_phase_b,
+        phase_b_train_fraction=args.phase_b_train_fraction,
+        phase_a_epochs=args.phase_a_epochs,
+        phase_b_epochs=args.phase_b_epochs,
+        hidden_dim=args.hidden_dim,
+        phase_a_noise_scale=args.phase_a_noise_scale,
+        phase_b_noise_scale=args.phase_b_noise_scale,
+        phase_b_rotation_degrees=args.phase_b_rotation_degrees,
+        phase_b_translation_x=args.phase_b_translation_x,
+        phase_b_translation_y=args.phase_b_translation_y,
+        circadian_sleep_interval_phase_a=args.sleep_interval_phase_a,
+        circadian_sleep_interval_phase_b=args.sleep_interval_phase_b,
+        circadian_config=circadian_config,
+    )
+
+    result = run_continual_shift_benchmark(config=config, seeds=seeds)
+    formatted = format_continual_shift_benchmark(result)
+    print(formatted)
+    if args.output_file:
+        output_path = Path(args.output_file)
+        output_path.write_text(formatted + "\n", encoding="utf-8")
+
+
+def _build_strength_case_circadian_config() -> CircadianConfig:
+    """Build a practical circadian profile for retention/adaptation stress tests."""
+    return CircadianConfig(
+        use_reward_modulated_learning=True,
+        reward_scale_min=0.8,
+        reward_scale_max=1.3,
+        split_threshold=0.30,
+        prune_threshold=0.04,
+        max_split_per_sleep=1,
+        max_prune_per_sleep=0,
+        replay_steps=2,
+        replay_memory_size=8,
+        replay_learning_rate=0.03,
+        replay_inference_steps=10,
+        replay_inference_learning_rate=0.12,
+    )
+
+
+def _parse_int_list(raw_values: str) -> list[int]:
+    items = [item.strip() for item in raw_values.split(",") if item.strip()]
+    if not items:
+        raise ValueError("Expected at least one integer seed.")
+    return [int(item) for item in items]
+
+
+if __name__ == "__main__":
+    main()