code backup: README.md

README.md

@@ -739,6 +739,24 @@ reasoning.
 
 ## How the experiment is actually run, step by step
 
+### Step 0 — Execution environment (read this first)
+
+Everything below assumes `./setup.sh` has completed (see its section above) and the
+shared venv is active in EVERY shell that runs a command:
+
+```bash
+source /root/.venv/bin/activate        # required before any python command below
+cd /workspace
+python -m pytest tests -q              # MUST pass before burning any GPU time
+```
+
+If setup.sh fell back to split venvs, perception commands (`inference/`, `encoder/`)
+use `/root/.venv-perception` and VLM commands (`harness/`, `corruption/` VLM arm) use
+`/root/.venv-vlm`; everything else runs in either. After each step below finishes,
+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.
+
 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
@@ -871,15 +889,70 @@ H10, H11's 16-token half, and H18's base-protocol read.
 All on the one pre-registered ~600-question sample, extended protocol, ground-truth
 base codes; the solver runs every arm at zero GPU cost as the second reasoner:
 
+Substitute the frozen Step-1/2 winners for `<depth*> <tracking*> <selection*>
+<frames*>` throughout (depth/tracking are the production defaults `relative` /
+`tracking` unless Step 2 changed them):
+
 ```bash
-# H21: 4 corruption types x 4 magnitudes, BOTH 4B models (~19k generations),
-#      including the empirical-residual noise mode calibrated from Step 2's codes
-# H22: 2 chimera conditions (GT inventory x perceived geometry, and the reverse)
-# H24: 5 solver-certified answer-preserving re-parameterizations
-# H6 probe + wrong-scene control + serialization/wording robustness: 1 model each
-python -m corruption.launch ...   # see corruption/ docs once built
+# (a) generate + commit the pre-registered question sample (procedure and seed are
+#     already frozen in analysis/preregistration.md)
+python -m corruption.sample --depth <depth*> --tracking <tracking*> \
+  --input-selection <selection*> --frames <frames*>
+python backup.py --repo-id <you>/<repo> --target code   # timestamp the realized sample
+
+# (b) measure the real SAM3+DA3 residual distribution (H21's empirical mode)
+python -m corruption.empirical --depth <depth*> --tracking <tracking*> \
+  --input-selection <selection*> --frames <frames*>
+
+# (c) H21 grid: 4 synthetic corruption types x 4 magnitudes, BOTH 4B models,
+#     solver arm included automatically
+python -m corruption.launch --arm both --models qwen3.5-4b,internvl3.5-4b \
+  --transforms position-jitter,dimension-noise,drop-objects,hallucinate-objects \
+  --magnitudes 0.1,0.25,0.5,1.0 --sample analysis/corruption_sample.json
+
+# (d) H21 empirical mode (magnitude = scale; 1.0 = the real operating point)
+python -m corruption.launch --arm both --models qwen3.5-4b,internvl3.5-4b \
+  --transforms empirical --magnitudes 0.25,0.5,1.0,2.0 \
+  --residuals analysis/sam3_da3_residuals.json --sample analysis/corruption_sample.json
+
+# (e) H22 chimeras (needs the Step-2 config so it can load the perceived codes)
+python -m corruption.launch --arm both --models qwen3.5-4b \
+  --transforms chimera-gt-inventory,chimera-perceived-inventory --magnitudes 0 \
+  --depth <depth*> --tracking <tracking*> --input-selection <selection*> \
+  --frames <frames*> --sample analysis/corruption_sample.json
+
+# (f) H24 invariance: certify FIRST (only certified scene/transform pairs count),
+#     then run the certified transforms
+python -m corruption.run --arm certify --transform translate --magnitude 10 \
+  --scenes <comma-separated sample scenes>
+python -m corruption.launch --arm both --models qwen3.5-4b \
+  --transforms translate,rotate-z,reorder,round-precision --magnitudes 10,90,0,3 \
+  --sample analysis/corruption_sample.json
+# (magnitudes map per transform: translate=meters offset, rotate-z=degrees,
+#  reorder=ignored, round-precision=decimals -- run one transform per invocation if
+#  you want per-transform magnitudes, e.g. --transforms translate --magnitudes 10)
+
+# (g) H6 single-object probe + wrong-scene control (1 model each)
+python -m corruption.launch --arm both --models qwen3.5-4b \
+  --transforms single-object --magnitudes 0 --sample analysis/corruption_sample.json
+python -m corruption.run --arm vlm --model qwen3.5-4b --transform wrong-scene \
+  --magnitude 0 --scenes <scene> --wrong-scene <different scene> \
+  --sample analysis/corruption_sample.json
+
+# (h) robustness arm: paraphrased context line, then YAML serialization -- each
+#     changes exactly one surface property of B's prompt; explicit --results-dir
+#     keeps the arms isolated from the standard B results
+python -m harness.B.launch --model qwen3.5-4b --spatial-code-format explicit \
+  --input-selection <selection*> --frames <frames*> \
+  --results-dir "results/B/qwen3.5-4b/extended/robustness-paraphrase" --paraphrase-context
+python -m harness.B.launch --model qwen3.5-4b --spatial-code-format explicit \
+  --input-selection <selection*> --frames <frames*> \
+  --results-dir "results/B/qwen3.5-4b/extended/robustness-yaml" --serialization yaml
 ```
 
+Back up when done: `python backup.py --repo-id <you>/<repo> --target corruption` (and
+`--target B` for the robustness arms).
+
 *Hypotheses fed*: H21, H22, H24 (novel primary); H6 instantiation; the wrong-code
 and robustness controls that close the "does the model even read the code" and
 "is it a prompt artifact" objections.