code backup: README.md

README.md

@@ -757,6 +757,32 @@ back up its results (additive, never destructive):
 `python backup.py --repo-id <you>/<repo> --target <step's target>` -- targets are
 listed in backup.py's own docstring; `--target all` covers everything.
 
+**Step 0.5 (optional pilot) — reasoning-budget calibration.** Before committing the
+main sweeps to the pre-registered 2048-token budget, a small pilot (e.g. 4 H100s)
+sweeps OPERATOR-CHOSEN budgets on an OPERATOR-CHOSEN question list and format —
+nothing about the pilot's scope is baked in; you list the budgets, the exact
+question ids, and the spatial-code format:
+
+```bash
+python -m calibration.run --models all --budgets 256,512,1024 \
+  --questions 12,34,56,789 --spatial-code-format explicit \
+  --depth metric --tracking tracking --input-selection selective --frames 64
+# or --questions-file pilot_questions.json instead of --questions
+python -m calibration.report        # per-cell recommendation
+```
+
+Budgets, question ids, spatial-code format, AND the full perceived-code config
+(depth / tracking / input selection / frame count) are all required operator
+choices -- nothing is defaulted. Scenes are derived from the question ids
+automatically (unknown ids fail loudly); each cell writes to
+`results/calibration/<model>/<format>/<depth>/<tracking>/<input>/<frames>/<budget>/<scene>/<question_id>.json`,
+so pilots at different configs can never collide.
+
+The recommendation rule is fixed in `calibration/report.py`'s docstring (smallest
+budget within tolerance of best AND with an acceptable forced rate). If it moves the
+budget off 2048, record the new value in `analysis/preregistration.md` BEFORE Step 1
+and pass `--reasoning-budget` on every extended run thereafter.
+
 Rather than a full factorial sweep across every axis, configuration is narrowed in
 stages — each stage's winner freezes the next stage's config — because the model axis
 is never collapsed (every stage always sweeps all 3 models) but frame count, input