Microscopy AI that knows when it’s wrong.

Ask a question about a microscope image. VeriCell finds every cell or organoid, checks each one, and answers with an honest range, plus the few objects a scientist should double-check.

16 / 16 benchmark question types beat the published agent 0.012 calibration error 6 Meta model roles
recorded runintestinal organoids · brightfield
“Which organoid is the roundest?”High confidence
Loading a real run…
01Look & plan
02Find every object
03Check each one
04Answer + range

The best cell segmenter quietly gets about 1 in 5 cells wrong.

Almost every measurement in cell biology starts with AI outlining each cell or organoid. Drug screens, cancer research and neuroscience imaging all measure from those outlines. The errors are silent, an image can hold hundreds of objects, and nobody re-checks them all by hand.

1 in 5

cells missed or mis-drawn by Cellpose-SAM on our LIVECell test (recall 0.81), with no warning. In one dense image SAM 3.1 alone found 24 of 691 cells. Answers like “the largest organoid is 31,000 px²” come with no error bar, so a reader can’t tell a solid number from a lucky guess.

A fixed pipeline does the work. Muse Spark decides only where judgement is needed.

Plain code computes every number, so the model can’t invent one. You see each step as it happens.

01

Look & plan

Muse Spark reads the image, names the objects in words SAM understands, and turns your question into a strict plan.

02

Find every object

Cellpose-SAM at three scales plus SAM 3.1 on zoomed tiles. Two independent opinions on every object.

03

Check each one

A verifier turns 22 signals into an honest probability that each object is real. An agent decides what to double-check.

04

Answer with a range

Exact geometry gives the answer; a 90% range calibrated on held-out images says how sure it is. Your checks narrow it live.

Each model does what it’s best at, and each call is shown live.

In the lab, a chip lights up the moment a model is called, with a running count.

Muse Spark Meta API

Reads the image, plans the question, judges doubtful crops, and is the agent’s controller.

SAM 3.1 Meta API

Segments from a phrase. Zoom-tiling made it 3× better on dense images (F1 0.27 → 0.82).

Muse Voice Transcribe Meta API

Hold the mic and ask out loud.

Muse Spark Open open weights

The whole reasoning layer on open weights via Fireworks, one switch away.

Muse Glimmer 30B fine-tuned · experimental

LoRA fine-tuned on Fireworks as a dedicated crop judge, from our own labels. Not yet better than Muse Spark (61% vs 72%).

Cellpose-SAM open source

The strongest generalist cell segmenter, run at three scales so big organoids fit.

Lower error than the published organoid agent on all 16 question types.

MorphoOrgaVQA: 68 organoid images × 16 questions, median error (lower is better). Every number is out-of-fold: no image is scored by a model that saw it.

3.6%
largest-organoid area error, vs 16.8% published
0.866
AUROC spotting its own wrong cells, vs 0.823 for SAM’s confidence
0.012
calibration error: “80% sure” means right ~80% of the time
89–92%
of 90% ranges contain the truth on held-out images
QuestionMorphoOrgaAgentVeriCellError, side by side

It tells you which conclusions to trust, and which to check before you publish.

Measured on all 68 benchmark images, with verifier scores from models that never saw the image and the warning rules fixed before the test. Details in RESULTS.md, sections 6 and 7.

13% vs 49%
"Largest organoid" answers are wrong 13% of the time when VeriCell says the conclusion holds, and 49% when it warns it could change
83%
of wrong size conclusions are caught by the warning; checking about 1.5 named objects fixes 14 of them and breaks none
10 → 1
images where SAM 3.1 finds almost nothing, after we measured which words it actually sees ("organoid" alone: F1 0.00)
1 click
to stress-test a claim: "largest area at most 36,000 px²" came back 1-organoid fragile, with the one object to re-image

Drug and organoid screens

Size and count changes are the readout. A range and a fragility check say whether a 12% shrink is real or one bad outline.

Reviewers and PIs

Stress-test a figure's claim and get the objects that would break it, plus an evidence file anyone can re-check.

Automated microscopes

Re-image only the few spots that could change the answer: a z-stack plan with positions, instead of re-scanning the plate.

Imaging core facilities

Upload any brightfield or phase-contrast image; spend human checking time only where it moves the result.

It notices its own mistakes and decides what to check.

A harness keeps it honest: one tool at a time, a safety gate, budgets, and a stop rule.

“SAM found 0 objects versus 285 from Cellpose, so the prompt likely mismatched the biology word; trying a plain visual phrase.”the controller, on a dense colon-organoid image · it then found 181 objects

4 typed tools

Judge crops, re-prompt SAM, ask you, or finish.

Policy gate

Rejects invalid ids, repeats and over-budget calls.

Evaluator

Re-measures after every action; stops when the range is tight.

You, as a tool

Ranks the few objects that could move the answer most.

Swap the reasoning to Muse Spark Open, with the same answers.

Every reasoning step runs on hosted Muse Spark or on Muse Spark Open via Fireworks. Same benchmark, measured side by side.

CheckMuse Spark (hosted)Muse Spark Open
Planner, 16 benchmark questions16/1616/16
Planner time, 16 questions33 s21 s
Crop judge, 80 labelled crops72%70%
Agent controller, same situationjudge #150same decision

Reported on purpose, because honesty about uncertainty is the point.

  • On LIVECell cells, fusion doesn’t beat Cellpose-SAM’s F1 (0.851 vs 0.857); the gain there is knowing which cells are wrong.
  • The Muse Spark crop judge is right 70–72% of the time on “one organoid or several?”; a fine-tuned Muse Glimmer reached 61%.
  • Showing the judge labelled example crops didn’t help.
  • 90% ranges are wide on hard images: most answers are within a few percent, about 1 image in 10 fails badly.
  • The benchmark was rebuilt as 68 images (the paper used 69 and doesn’t list them).

Questions

Is this for diagnosis?

No. VeriCell is for research images only.

Does the AI calculate the numbers?

No. Muse Spark plans and judges; areas, counts and ranges are computed by plain code that matches the benchmark’s formulas exactly.

What does the 90% range mean?

On images it never trained on, the true answer fell inside the range 89–92% of the time. A wide range is the system honestly saying “check this”.

Can I run it on my own images?

Yes, locally: clone the repo, add your Meta API key, run make serve. The public demo uses four example images.

See it check its own work.