Skip to content

uipsf: in situ against molecule stacks on M2 (write-up, scripts, results) - #166

Merged
jries merged 1 commit into
mainfrom
claude/m2-insitu-vs-stacks
Nov 15, 2026
Merged

jries merged 1 commit into
mainfrom
claude/m2-insitu-vs-stacks

Conversation

@jries

@jries jries commented Nov 15, 2026

Copy link
Copy Markdown
Contributor

The comparison of 13–14 November, written up so that a session without the data can build on it.

  • docs/uipsf/architecture.md, a new section, "In situ against molecule stacks on M2 (November 2026)". It covers which files were paired (12 Nov 2026, iXon, 638 nm / 690 nm, lens state), each run's settings, a table of Noll 5/6/7/8/11/12/16/22/37, γ, σ, the focus and h, held-out cross scores, what was ruled out, the open hypotheses, the control, z linearity and the NPC rings. Noll 4 is not a learnt mode; the focus stands in for it.
  • docs/uipsf/roadmap.md item 9: a one-line pointer.
  • benchmarks/uipsf/real/m2_pairs/: the scripts. The data path comes from $M2_DATA; no data are in git. cross_score.py is new and scores each model on the other's held-out molecules.
  • benchmarks/uipsf/results/m2_pairs_{runs,scores,cross,npc_rings}.jsonl: the results.

Main finding, in nats per molecule, relative to the stack model (lower is better):

model stack held-out (553) movie held out (223) movie surplus, not drawn (4105)
in situ, edge-pinned, n 1.345 +4.51 −1.75 +0.36
in situ, free, n 1.33 +8.43 −2.75 +0.50

Each model wins on its own sample's held-out molecules. A coverslip tilt takes most of the difference in the low orders, but it needs 1.6–2.9°.

One more cross-check is still running and will follow as a commit on this branch: in situ with γ and the phase above order 6 held from the stack model.

No code outside the benchmarks changed. The tests were not run because this PR has only docs and benchmark scripts. ruff check --config benchmarks/uipsf/ruff.toml benchmarks/uipsf/real/m2_pairs is clean.

🤖 Generated with Claude Code

…runs, held-out cross scores

M2, 12 November 2026: beads, the BG-AF647 nucleus movie and AF647 molecules on the glass in z sweeps,
one camera ROI, one evening.  architecture.md "In situ against molecule stacks on M2 (November 2026)"
and a pointer in roadmap item 9: the files, each run's settings, a table of the low orders, γ, σ, the
focus and h, and each model scored on the other's held-out molecules (cross_score.py, new): on the
stack's 553 the in situ pupils lose 4.5 and 8.4 nats per molecule; on the movie's 223 held out they win
by 1.75 and 2.75, on its 4105 surplus lose by 0.36-0.50.  What was ruled out, the open hypotheses, the
simulated control, z linearity and the NPC rings.

The scripts in benchmarks/uipsf/real/m2_pairs/ (data from $M2_DATA, not in git), the results in
benchmarks/uipsf/results/m2_pairs_*.jsonl.  Benchmark scripts and docs only; no tests run.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@jries
jries merged commit 307a750 into main Nov 15, 2026
6 checks passed
@jries

jries commented Nov 15, 2026

Copy link
Copy Markdown
Contributor Author

Cross-check added (181eaee): in situ at order 6, with γ and the phase above order 6 held from the stack model. Z6 61.3 (stack 56.6, order 8 71.9); Z11 29.5 (stack 24.2, order 8 55.1); focus 537 nm (stack held 420, order 8 682). Held-out scores in nats per molecule against the stack model: +2.32 on the stack's held-out molecules (order 8 +4.51), −1.61 on the movie's held-out molecules (order 8 −1.75), +1.09 on the movie's surplus (order 8 +0.36). Most of the low-order gap closes; the x coma and the stack's lead on its own molecules remain.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant