Revert "Feat/issue 271 preregistered study"

configs/drone_distribution_shift.yaml

@@ -1,8 +1,3 @@
-# Issue #265: primary moderate TEST-B condition. Training receives only the
-# nominal environment.parameters block. shift_parameters are held back until
-# after training and the shared pre-shift evaluation.
-name: "drone_online_adaptation"
-seed: 31001
 # Issue #105 / Issue #265: controlled distribution-shift benchmark for drone navigation.
 #
 # Protocol: the policy is trained ONLY on TRAIN conditions (8 static obstacles,
@@ -37,7 +32,6 @@ algorithm:
   batch_size: 128
   parameters:
     n_steps: 1024
-    n_epochs: 10
     clip_range: 0.2
     ent_coef: 0.01
 
@@ -46,18 +40,6 @@ environment:
   max_steps: 300
   parameters:
     bounds: [50.0, 50.0, 25.0]
-    num_obstacles: 8
-    num_lidar_rays: 16
-    linear_damping: 0.05
-    max_acceleration: 4.0
-    wind_speed: 0.5
-    gust_sigma: 0.15
-    gust_theta: 0.15
-    wind_direction: [1.5, 0.5, 0.0]
-
-training:
-  total_timesteps: 60000
-  checkpoint_freq: 0
     dynamic_obstacle_speed: 1.5
 
 training:
@@ -69,16 +51,6 @@ evaluation:
   eval_episodes: 15
   deterministic: true
 
-adaptation_benchmark:
-  protocol_version: "2.0"
-  scenario: "TEST-B"
-  shift_parameters:
-    num_obstacles: 12
-    wind_speed: 4.0
-    gust_sigma: 0.6
-
-output_dir: "artifacts/adaptation"
-log_dir: "artifacts/adaptation/logs"
 output_dir: "artifacts"
 log_dir: "artifacts/logs"
 

docs/research/TREATMENT_CARD.md

@@ -1,72 +1,3 @@
-# Issue #265 Treatment Card
-
-**Status: frozen before benchmark execution.** This card defines the adaptation
-treatment. Its SHA-256 is stored in every Issue #265 artifact. No benchmark
-outcome may be used to change these choices.
-
-## Fixed treatment choices
-
-- **Scope:** adapt the already trained PPO or SAC policy using only transitions
-  collected in completed post-shift episodes. Base training configuration,
-  checkpoint, and environment interaction schedule are identical between arms.
-- **Mechanism:** one cumulative, episode-bounded batch update at each boundary
-  B5 through B14. Each batch is the concatenation of all post-shift transitions
-  from episodes 1 through the boundary episode. The batch contains observation,
-  action, reward, next observation, terminated, and truncated values. No nominal
-  data is retained in or passed to the adaptation path.
-- **PPO update:** store the behavior-policy log-probability and value estimate
-  with each transition at collection time, plus the behavior value of a
-  truncated transition's next observation. At each block, construct a fresh
-  Stable-Baselines3 `RolloutBuffer` from the permitted cumulative prefix, use
-  the configured `gamma` and `gae_lambda`, bootstrap truncated transitions
-  from their recorded next observation, and run the native clipped PPO update
-  for the config's `n_epochs` and `batch_size`. The configured entropy and
-  value-function coefficients are unchanged. Previously seen prefix data is
-  reused at later blocks with its stored behavior quantities.
-- **SAC update:** construct a fresh Stable-Baselines3 `ReplayBuffer` for each
-  block from the permitted cumulative prefix; the training replay buffer is
-  never reused. Its capacity is `max(prefix_transition_count, configured
-  batch_size)`. Run native SAC training for the configured `gradient_steps`
-  and `batch_size`, retaining the configured learning rate, target update, and
-  entropy mechanism. A sampled minibatch may draw with replacement when it is
-  larger than the currently available transition count. Time-limit truncations
-  are marked according to SB3's timeout-mask convention.
-- **Randomness:** each block seeds Python, NumPy, and Torch with its one
-  preregistered `update` seed while sampling/updating, then restores the caller's
-  RNG states. No environment step occurs in either adapter.
-- **Learning rate, batch size, epochs/steps, and regularization:** all values
-  come from the base cell config; no additional optimizer or regularizer is
-  added. PPO's configured entropy/value terms and SAC's configured entropy
-  mechanism remain part of their respective native objectives.
-- **Minimum data:** at least one complete episode (the current block's newly
-  completed episode) and at least one valid transition are required.
-- **Buffer construction:** a fresh immutable snapshot is assembled at each
-  block boundary from the allowed prefix only. No shared/global replay buffer
-  is used. Transitions are never sampled from a future episode.
-- **Update timing:** only after episode termination/truncation and before the
-  next reset; no updates at B15. The policy is constant during every episode.
-- **Failure behavior:** any exception, non-finite loss/parameter/optimizer
-  value, invalid tensor, or shape change invalidates the replicate. The block is
-  rolled back before the failed replicate is recorded; execution does not
-  continue to another evaluation episode.
-- **Diagnostics:** record pre/post SHA-256 model fingerprints, update seed,
-  visible episode indices, transition count, status, and global L2 parameter
-  delta. Zero delta is recorded as a non-mutating update, not successful
-  adaptation.
-
-The Fixed arm is an independent clone of the frozen checkpoint. It performs no
-training or optimizer operation after the fork. Both arms use the same derived
-post-shift episode seeds. A smoke run validates machinery only and is not
-empirical evidence for the research hypothesis.
-
-## Implementation constraint
-
-Algorithm-specific recorded-data adapters are implemented for PPO and SAC.
-The complete benchmark experiment runner, environment shift integration,
-artifacts, and CLI are still required before the treatment may be run as an
-Issue #265 experiment. The runner must fail explicitly for any configuration
-that cannot provide the required PPO behavior quantities or recorded transition
-data without extra environment interaction.
 # AdaptiveRL Treatment Card: Online Fine-Tuning Specification
 
 **Protocol Version**: 2.0<br>