# KS1H+ induced-fit control

This transfers `/root/me344-dock`'s **pre-expanded rigid receptor cross-docking**
procedure to the author S5.4 KS1 monocation. It does not perform protein relaxation.
Read `docs/analysis/ks1-inducedfit-2026-09-13.md` before interpreting scores.

```sh
/root/girdin-dock/venv/bin/python scripts/ks1_inducedfit/prepare.py \
  --out data/ks1-inducedfit/NEW_RUN
setsid nohup taskset -c 0-3 nice -n 10 \
  /root/girdin-dock/venv/bin/python scripts/ks1_inducedfit/run.py \
  --root data/ks1-inducedfit/NEW_RUN > data/ks1-inducedfit/NEW_RUN/run.stdout 2>&1 < /dev/null &
/root/girdin-dock/venv/bin/python scripts/ks1_inducedfit/analyse.py \
  --root data/ks1-inducedfit/NEW_RUN
/root/girdin-dock/venv/bin/python scripts/ks1_inducedfit/viewer.py \
  --root data/ks1-inducedfit/NEW_RUN
```

The absolute source paths are intentional: the original receptor preparation,
legacy typed PDBQT and source-pinned gallery are authoritative inputs, not files
to regenerate from molecule names. Preparation copies and hashes them; preserve
that input bundle for reuse elsewhere. Raw data remains ignored. NumPy and RDKit
are required; installed versions are recorded with the final evidence.

`analyse.py --partial` only inspects completed, sealed receipts and writes to a
separate `analysis-partial` directory. Final analysis requires exactly the full
20 KS1 jobs plus the four declared controls. Every output's SHA-256 and manifest
association are checked. Zero returned modes or malformed outputs are preserved
as failed calculations, never dropped from the plan.

Pose QC checks atom names/types, all fixed-core distances, proper core alignment
and the six anchor bonds. Clustering uses nonaligned 55-heavy-atom RMSD with
aryl symmetry (2 Å threshold). A few modes are not proof of a snug or invalid fit:
energy-window and RMSD output filters also limit counts. Inspect the actual
contacts and scores. The AD8 control separately measures best sampled and
top-ranked crystallographic RMSD.

Tests: `python scripts/ks1_inducedfit/test_geometry.py` uses the prepared dated
input. It challenges coordination breakage, rigid transforms, aryl symmetry and
the distinction between pose RMSD and aligned-conformer RMSD.

`shard.py` is a prepared optional transport worker for the two 1N45 variants.
Remote transfer/execution requires explicit authorization; its preparation is
not evidence that it ran. Collect only sealed, hash-verified results before the
local controller reaches those jobs, and never overwrite an active attempt.

Observed primary poses at the edge of the 22-A box motivated two adaptive
30-A/seed42 checks in the expanded structures. `box_sensitivity.py --root RUN`
writes a separate plan, logs and summary under `box30/`; it never rewrites the
primary matrix. Treat favorable surface poses from a wider box as a search-region
sensitivity, not validation of the original distal site. It uses the same four
physical CPU affinity as the primary queue if run concurrently.