diff --git a/research/classifier/measurements-meta.json b/research/classifier/measurements-meta.json
index 35e9b81..6037ef5 100644
--- a/research/classifier/measurements-meta.json
+++ b/research/classifier/measurements-meta.json
@@ -1,5 +1,5 @@
 {
-  "_note": "Hand-recorded from the v0.2.0 collection run done on the author's laptop before the audit fields were added to collect_measurements.py. Future runs will populate this file automatically, with per-cell tokens + wall-clock aggregated by the collector. Fields with `null` are unknown for this specific run; everything else came from the PR description (#1) and the row count of measurements.csv.",
+  "_note": "Hand-recorded from the v0.2.0 collection run done on the author's laptop before the audit fields were added to collect_measurements.py. Future runs will populate this file automatically, with per-cell tokens + wall-clock aggregated by the collector. Token totals remain `null` because the OpenAI-compat `usage` block was not persisted during this run; the started_at / ended_at / wall_clock_s fields below are *reconstructed* from the modification timestamps of the run's stdout log file (`/tmp/mnemoscope-full.log` on the author's machine), not measured by the collector — they bound the wall-clock but include any time the orchestrator spent idle between cells.",
   "model": "gemma4-ctx:latest",
   "model_base": "gemma4:26b",
   "model_quantization": "Q4_K_M",
@@ -7,11 +7,11 @@
   "endpoint": "http://localhost:11434/v1",
   "endpoint_kind": "ollama",
   "offline": false,
-  "started_at": null,
-  "ended_at": null,
-  "wall_clock_s": null,
+  "started_at": "2026-04-29T04:11:20-04:00",
+  "ended_at": "2026-04-30T03:13:50-04:00",
+  "wall_clock_s": 82950,
   "wall_clock_s_grading_only": null,
-  "wall_clock_human": "multi-hour, single-machine local run",
+  "wall_clock_human": "23h02m30s end-to-end (reconstructed from log file mtime, not measured by collector)",
   "variants_attempted": 50,
   "variants_kept": 50,
   "variants_skipped_no_needles": 0,
diff --git a/research/classifier/model.json b/research/classifier/model.json
index b602930..cadf0bd 100644
--- a/research/classifier/model.json
+++ b/research/classifier/model.json
@@ -23,7 +23,7 @@
   },
   "rows_train": 40,
   "rows_test": 10,
-  "onnx_round_trip_max_delta": 1.5894571991914574e-08,
+  "onnx_round_trip_max_delta": 1.5894571880892272e-08,
   "data_source": "classifier/measurements.csv",
   "seed": 42,
   "grader_models": [
@@ -31,7 +31,7 @@
   ],
   "offline_rows": 0,
   "dataset_collection": {
-    "_note": "Hand-recorded from the v0.2.0 collection run done on the author's laptop before the audit fields were added to collect_measurements.py. Future runs will populate this file automatically, with per-cell tokens + wall-clock aggregated by the collector. Fields with `null` are unknown for this specific run; everything else came from the PR description (#1) and the row count of measurements.csv.",
+    "_note": "Hand-recorded from the v0.2.0 collection run done on the author's laptop before the audit fields were added to collect_measurements.py. Future runs will populate this file automatically, with per-cell tokens + wall-clock aggregated by the collector. Token totals remain `null` because the OpenAI-compat `usage` block was not persisted during this run; the started_at / ended_at / wall_clock_s fields below are *reconstructed* from the modification timestamps of the run's stdout log file (`/tmp/mnemoscope-full.log` on the author's machine), not measured by the collector \u2014 they bound the wall-clock but include any time the orchestrator spent idle between cells.",
     "model": "gemma4-ctx:latest",
     "model_base": "gemma4:26b",
     "model_quantization": "Q4_K_M",
@@ -39,11 +39,11 @@
     "endpoint": "http://localhost:11434/v1",
     "endpoint_kind": "ollama",
     "offline": false,
-    "started_at": null,
-    "ended_at": null,
-    "wall_clock_s": null,
+    "started_at": "2026-04-29T04:11:20-04:00",
+    "ended_at": "2026-04-30T03:13:50-04:00",
+    "wall_clock_s": 82950,
     "wall_clock_s_grading_only": null,
-    "wall_clock_human": "multi-hour, single-machine local run",
+    "wall_clock_human": "23h02m30s end-to-end (reconstructed from log file mtime, not measured by collector)",
     "variants_attempted": 50,
     "variants_kept": 50,
     "variants_skipped_no_needles": 0,