Skip to content

uipsf: M2 follow-up -- tilt, drift, score per height, h at 5 nm, a slide's flatness - #168

Merged
jries merged 2 commits into
mainfrom
claude/m2-followup
Nov 15, 2026
Merged

jries merged 2 commits into
mainfrom
claude/m2-followup

Conversation

@jries

@jries jries commented Nov 15, 2026

Copy link
Copy Markdown
Contributor

This runs PR #167's scripts on M2's paired data (12 Nov 2026) and writes up the results. The new section is "Follow-up (November 2026)" under "In situ against molecule stacks on M2" in docs/uipsf/architecture.md, and the in situ line of roadmap item 9 is updated.

Results

  • Score per height: the stack's pupil fits better below ~400 nm and the in situ pupils fit better above, by more the deeper the molecule. Per bin, from 0–200 nm to 800–1000 nm, in situ goes from +8.4 to −6.4 nats per molecule. So the difference depends on depth. The movie's surplus molecules cluster at 200–400 nm, which explains why that set favoured the stack.
  • h held at 5 nm:
    • The stack's phase stays within 0.4 nm rms, γ goes from 0.12 to 0.25, and the held-out score is unchanged (229.39 against 229.43).
    • In situ started from it lands where it did from h = 75 nm.
    • The gap is +5.20 / −1.73 nats per molecule, against +4.51 / −1.75 from h = 75.
    • So h is not part of the gap.
  • Tilt:
    • Most tiles of the movie lie under the nucleus (spike fraction 0), so the plane through all tiles measures the nucleus floor, not the coverslip.
    • The three tiles with glass molecules slope 23–29 nm per µm (1.3–1.7°) along y, while the pupil's tilt is along x.
    • The new stack_flatness.py measures the AB slide: level to 0.04° and flat to 4–5 nm rms, 30–70 times less than the movie's slope.
  • Drift:
    • Molecules on the glass appear only in the first ~2200 frames, so the fitted edge-vs-time line is not a drift.
    • The nucleus floor holds within ±20 nm over the last ~50 000 frames.
    • The first ~18 000 frames are undetermined.

Also in this PR

  • The order-6 cross-check commit, which was pushed after uipsf: in situ against molecule stacks on M2 (write-up, scripts, results) #166 was merged. It is cherry-picked here.
  • A fix in tilt_from_edge.py: the stack's plane is now moved off the GPU before converting to float64, which MPS lacks.
  • Results in benchmarks/uipsf/results/m2_pairs_followup.jsonl, plus new rows in m2_pairs_cross.jsonl and m2_pairs_runs.jsonl. No data are in git.

Differences from the brief

  • insitu_all_n1345_ab_n1345 is the existing model, which learnt its focus rather than taking it from the coverslip edge. It was used as it is.
  • The order-6 model is named insitu_start_o6_edge_n1345_ab_n1345.

The tests were not run, because the PR has only docs, benchmark scripts and results. ruff check --config benchmarks/uipsf/ruff.toml benchmarks/uipsf/real/m2_pairs is clean.

🤖 Generated with Claude Code

jries and others added 2 commits November 15, 2026 13:28
…above held from the molecule stack

insitu_start_o6_edge_n1345_ab_n1345: Z6 61.3 (stack 56.6, order 8 71.9), Z11 29.5 (24.2, 55.1), Z12 6.2,
x coma 24.7 / 14.4 (4.6 / 1.7), focus 537 nm (stack held 420, order 8 682).  Held-out: +2.32 nats per
molecule on the stack's (order 8 +4.51), -1.61 on the movie's held out (-1.75), +1.09 on its surplus
(+0.36).  Most of the low-order excess was the modes above order 6's; x coma and the gap remain.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… nm, a slide's flatness

PR #167's scripts run on M2's paired data (architecture.md "Follow-up (November 2026)", roadmap item 9):
- score per height: the stack's pupil fits better below ~400 nm, in situ's above (+8 ... -6 nats per
  molecule from 0-200 to 800-1000 nm): depth-dependent; the surplus crowds at 200-400 nm.
- h held at 5 nm: the stack's phase within 0.4 nm rms, γ 0.12 -> 0.25, the same held-out score; in situ
  from it lands where it did; the gap +5.20 / -1.73 nats per molecule (+4.51 / -1.75).
- tilt: most tiles lie under the nucleus (spike 0); the three glass tiles slope 23-29 nm per µm along y,
  the pupil's tilt is along x.  stack_flatness.py (new): a slide on M2 is level to 0.04°, flat to 4-5 nm
  rms -- 30-70 times less than the movie's.
- drift: glass molecules only in the first ~2200 frames, so the edge's line is not a drift; the floor
  holds within ±20 nm over the last ~50 000 frames.
tilt_from_edge.py: the stack's plane read off the device before float64 (MPS has none).  The order-6
cross-check's commit, cherry-picked, missed #166.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@jries
jries merged commit 13bb03e into main Nov 15, 2026
6 checks passed
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