Ignore observational reads in replay receipt validation

research/experiments/mechanism_replay.py

@@ -371,7 +371,11 @@ def verify_replay_submission(runtime,episode,history,submission):
             (runtime.campaign_id,work_id)).fetchall()
     recorded=[json.loads(row[0]) for row in rows]
     completed={row['action_work_id']:row for row in recorded}
-    submitted_actions=[action for entry in history for action in entry.get('actions',[]) if isinstance(action,Mapping)]
+    submitted_actions=[action for entry in history for action in entry.get('actions',[])
+        if isinstance(action,Mapping) and not (
+            action.get('action_name')=='read_context_segment' or
+            isinstance(action.get('result'),Mapping) and
+            action['result'].get('operation')=='read_context_segment')]
     forged=[action.get('action_work_id') for action in submitted_actions if action.get('state')=='completed' and
         (action.get('kind')!='replay_transition' or completed.get(action.get('action_work_id'))!=action)]
     performed={row['result']['record_id'] for row in recorded if row['result']['record_id'] is not None}

tests/test_mechanism_replay.py

@@ -185,6 +185,19 @@ def test_unexecuted_and_forged_operations_cannot_certify_repair(tmp_path):
     assert verify_replay_submission(runtime,episode,history,submission)['valid'] is False
 
 
+def test_observational_context_reads_do_not_count_as_forged_scientific_actions(tmp_path):
+    runtime,episode,_=runtime_fixture(tmp_path)
+    history,submission=repair(runtime,episode)
+    history.append({'actions':[{'kind':'analysis','action_work_id':'episode--replay-fixture--read-context--page-1',
+        'state':'completed','result':{'operation':'read_context_segment','segment_id':'public-context',
+            'pointer':'','offset':0,'next_offset':32,'content':'observational context only'}}]})
+    verified=verify_replay_submission(runtime,episode,history,submission)
+    assert verified['forged_history_actions']==[]
+    assert verified['valid'] is True
+    assert verified['scientific_metric_values_generated'] is False
+    assert verified['gpu_actual_seconds']==0
+
+
 def test_source_artifacts_are_reconciled_after_edit_not_implicitly_reused(tmp_path):
     runtime,episode,prepared=runtime_fixture(tmp_path)
     binding=prepared['action_bindings'][0]

tests/test_study_runtime.py

@@ -1553,6 +1553,40 @@ def test_large_model_context_interning_preserves_exact_json_and_tail_text():
     assert stats[0]["output_characters"] < stats[0]["input_characters"]
 
 
+def test_aide_artifact_observation_payload_bloat_is_losslessly_shared_before_codex_transport():
+    """Keep full evidence while sharing duplicated AIDE fields in the actual wire request."""
+    from research.experiments.study_context import restore_messages
+    from research.experiments.study_runtime import _intern_delivered_messages
+    from research.agents.codex_transport import request_payload
+
+    manifest = {"train_ids": [f"sample-{index:06d}" for index in range(10_000)],
+        "validation_ids": [f"validation-{index:06d}" for index in range(2_000)],
+        "dataset_contract": "registered-dataset-manifest," * 4_300}
+    training_config = {"optimizer": "AdamW", "learning_rate": 0.0003,
+        "batch_size": 64, "updates": 1200, "normalization": {"mean": [0.1] * 64,
+            "scale": [0.9] * 64}}
+    observations = [{"artifact_id": f"run-{index}", "dataset_manifest": manifest,
+        "training_configuration": training_config,
+        "validation_log": (f"validation-run-{index}:" + ("observed-value," * 6_500))}
+        for index in range(6)]
+    aide_context = {"view": "native_agent_public_artifact_index",
+        "artifacts": observations, "observations": observations}
+    original = [{"role": "user", "content": "AIDE public evidence:\n" +
+        json.dumps(aide_context, ensure_ascii=False, separators=(",", ":"))}]
+
+    delivered, stats = _intern_delivered_messages(original)
+    wire = request_payload(delivered, json_mode=True)
+    wire_characters = len(json.dumps(wire, ensure_ascii=False, separators=(",", ":")))
+
+    assert stats and stats[0]["input_characters"] > 5_000_000
+    assert stats[0]["output_characters"] < stats[0]["input_characters"]
+    # This is the observed Codex CLI input-envelope capacity, not an Agent
+    # reasoning quota. The repeated immutable evidence is still losslessly
+    # represented in the request, below that transport's real limit.
+    assert wire_characters < 1_050_000
+    assert restore_messages(delivered) == original
+
+
 def test_downstream_executor_error_preserves_actual_model_response(tmp_path, monkeypatch):
     campaign = tmp_path / "campaign"
     campaign.mkdir()