Clara Working Station

Alita · local agent, no network
Still from the November film23:16 Newest film · November 2026The updated working station One Clara for phone and terminal, small staged work sessions, two safety doors, nightly tests. 25 clickable chapters.
5:04The smuggling test 3:13The harness 3:45The film machine
All films and the archive →

What is computed on this machine, and what of it is published. A workbench, not a portfolio: only what was produced here is listed.

Workflow

From a question to a QC page: prompt from phone or terminal; research by Clé and literature workers through the exit guard; a method shortlist where each method becomes a job with a countable acceptance list; the job queue runs one method at a time; in the inner loop Clara writes the code in the sandbox, the GPU queue trains, the acceptance check counts facts and the critic reads, open criteria start the next round; then evaluation on the test set next to the leaderboard and the result with a QC page

One request, read left to right, using "I want to create a WMH segmentation model, can you research it?" as the example:

  1. Prompt. The researcher states the goal in plain language, from any device: the phone app, a desktop PC or the browser. This is the only step that happens outside the workstation.
  2. Research agents. On the workstation, an agent sends the question to Clé (the model on the Charité model hub API), which splits it, and several research agents search papers and code for every method and architecture used for the task.
  3. Method shortlist. The findings are merged into candidate methods, such as 2D U-Net, 3D U-Net, nnU-Net or BIANCA, each turned into tasks with machine-checkable acceptance criteria.
  4. Harness, outer loop. Takes one method at a time from the queue and moves on once it is done.
  5. Harness, inner loop on the workstation GPU. Clara (the clara-agent harness since 29 October 2026) writes the code, trains on open data (the WMH Segmentation Challenge training set), and an acceptance check counts the facts. If the check fails, the task goes back to Clara. The check asks no model.
  6. Evaluation. Every model is scored on the challenge test set with the official metric (Dice, lesion F1) and placed next to the public leaderboard, where the 2017 winner reached Dice 0.80.
  7. Result. The best model, a comparison table and a QC page.

Everything after the prompt runs on the workstation: the calls to Clé, the research agents, the harness, training and evaluation. Clara sees the data and has no network access. The research agents reach the network and Clé, and never see the data.

Why a local workstation

Waiting time before a GPU job can start Cluster · whole GPU: ≈ 9 h until scheduled (squeue --start, 16 Oct 2026); Cluster · GPU share: starts within seconds, but only small jobs fit; Workstation: no queue; idle model server wakes in 11.4 s; one training fold on the A5000 takes about 25 minutes Waiting before a GPU job can start An agent needs dozens of build-train-check rounds a day; the queue is paid on every round Dashed: one training fold of Clara's U-Net on the A5000, ≈ 25 min, for scale 0 h 2 h 4 h 6 h 8 h 10 h Cluster · whole GPU: ≈ 9 h until scheduled (squeue --start, 16 Oct 2026)Cluster · whole GPU≈ 9 h until scheduled (squeue --start, 16 Oct 2026) Cluster · GPU share: starts within seconds, but only small jobs fitCluster · GPU sharestarts within seconds, but only small jobs fit Workstation: no queue; idle model server wakes in 11.4 sWorkstationno queue; idle model server wakes in 11.4 s 25 min

Shared GPU time is too slow for this kind of work. On the institute cluster, a job asking for a whole GPU was scheduled about 9 hours ahead; only 7 of the cards have 80 GB. A small share of a card starts within seconds, but it only fits small trainings (measurement, 16 Oct 2026). An agent that builds a model, trains it, finds an error and tries again needs dozens of such rounds a day. With a queue in between, each round costs hours; on the workstation it costs minutes.

A local model can look at the patient data; a cloud model cannot. Clara runs on the workstation, so the images never leave the machine. That lets her do what a remote model may not: open every scan, read every mask, run the model on the whole cohort, and look at the cases where it fails. She sees far more of the data than any summary would carry, and she sees it fast.

Local is a route, not a permission. Whether a dataset may be processed by an AI system at all is decided by its data agreement, and a clause that forbids it binds the local model too (rule). Where processing is allowed, running it locally keeps the data inside the institution.

How the parts are used

Overview: phone and terminal reach the workstation; clara-agent runs in a kernel sandbox; the Weiche relay sends data-free requests to the Clé hub and requests that touch patient data to the local 27B; System One picks the model, judges answers and gates commands
PartRuns onPatient dataNetworkUsed for
Promptany device: phone app, desktop PC, browsernoyesstating the goal; nothing else happens there
CléCharité on-premise model hub (intranet), called from the workstationnoyesinterpreting the request, splitting it into questions and tasks
Research agentsworkstation, using Clénoyessearching papers and code for methods and architectures
Claraworkstation, clara-agent; model per turn: Clé for data-free work, local Qwen3.8-27B for patient datayes (local model only)no (loopback only, kernel-enforced)writing the code, training, running and inspecting models on the data
Relayworkstation, port 8770checks every requestClé onlyfixed rules first (ids, cohort names, data access -> local), then System One picks the Clé model per turn and judges each answer
System Oneworkstation, small decision modelnonomodel choice per turn, verdict per answer, approval gate for commands
Harnessworkstationthrough Claranothe outer loop over methods and the inner loop until a task passes
Checkworkstationcounts onlynoacceptance criteria as countable facts; asks no model
Exit guardworkstation, port 8899blocks itchecked questions onlyClara's only way out: technical questions without data reference
Blackboardworkstation and reponovia the repowhere Clara and the cloud model report progress to each other

Results

Mean Dice per method on 20 public WMH Challenge test cases monai-attentionunet 0.738; monai-segresnet 0.707; monai-unet 0.704; monai-swinunetr 0.684; unet-vorhersage 0.658; unet-eigen 0.647; bianca-n201-t90c3 0.646; bianca-n201 = t90 0.629; bianca-n201-t85c3 0.613; bianca-n201-t85 0.595; 2017 winner 0.80 on all 110 test cases Mean Dice per method · 20 public test cases Bar: mean over the same cases · whisker: 95 % bootstrap interval over cases Dashed: 2017 challenge winner on all 110 test cases, not the same cases 0.0 0.2 0.4 0.6 0.8 1.0 Dice monai-attentionunet: mean Dice 0.738, 95 % interval 0.668 to 0.800, 20 casesmonai-attentionunet0.738 monai-segresnet: mean Dice 0.707, 95 % interval 0.642 to 0.769, 20 casesmonai-segresnet0.707 monai-unet: mean Dice 0.704, 95 % interval 0.643 to 0.762, 20 casesmonai-unet0.704 monai-swinunetr: mean Dice 0.684, 95 % interval 0.617 to 0.746, 20 casesmonai-swinunetr0.684 unet-vorhersage: mean Dice 0.658, 95 % interval 0.567 to 0.730, 20 casesunet-vorhersage0.658 unet-eigen: mean Dice 0.647, 95 % interval 0.548 to 0.733, 20 casesunet-eigen0.647 bianca-n201-t90c3: mean Dice 0.646, 95 % interval 0.570 to 0.718, 20 casesbianca-n201-t90c30.646 bianca-n201 = t90: mean Dice 0.629, 95 % interval 0.547 to 0.707, 20 casesbianca-n201 = t900.629 bianca-n201-t85c3: mean Dice 0.613, 95 % interval 0.529 to 0.697, 20 casesbianca-n201-t85c30.613 bianca-n201-t85: mean Dice 0.595, 95 % interval 0.504 to 0.686, 20 casesbianca-n201-t850.595 2017 winner 0.80
Dice against reference lesion volume, 20 public test cases sub-068 0.9 ml: monai-attentionunet 0.300, bianca-n201-t90c3 0.316; sub-090 1.88 ml: monai-attentionunet 0.626, bianca-n201-t90c3 0.379; sub-078 1.95 ml: monai-attentionunet 0.568, bianca-n201-t90c3 0.486; sub-140 2.54 ml: monai-attentionunet 0.749, bianca-n201-t90c3 0.609; sub-084 3.2 ml: monai-attentionunet 0.650, bianca-n201-t90c3 0.522; sub-091 3.29 ml: monai-attentionunet 0.608, bianca-n201-t90c3 0.551; sub-141 4.45 ml: monai-attentionunet 0.687, bianca-n201-t90c3 0.447; sub-124 4.67 ml: monai-attentionunet 0.778, bianca-n201-t90c3 0.555; sub-086 6.43 ml: monai-attentionunet 0.614, bianca-n201-t90c3 0.572; sub-097 7.51 ml: monai-attentionunet 0.602, bianca-n201-t90c3 0.498; sub-154 19.76 ml: monai-attentionunet 0.785, bianca-n201-t90c3 0.659; sub-128 20.27 ml: monai-attentionunet 0.892, bianca-n201-t90c3 0.831; sub-125 23.8 ml: monai-attentionunet 0.845, bianca-n201-t90c3 0.775; sub-094 26.52 ml: monai-attentionunet 0.851, bianca-n201-t90c3 0.782; sub-135 29.08 ml: monai-attentionunet 0.820, bianca-n201-t90c3 0.838; sub-153 39.2 ml: monai-attentionunet 0.915, bianca-n201-t90c3 0.709; sub-074 39.27 ml: monai-attentionunet 0.918, bianca-n201-t90c3 0.867; sub-101 49.5 ml: monai-attentionunet 0.835, bianca-n201-t90c3 0.832; sub-144 52.06 ml: monai-attentionunet 0.869, bianca-n201-t90c3 0.851; sub-113 85.15 ml: monai-attentionunet 0.851, bianca-n201-t90c3 0.833 Dice depends on how much lesion there is One case per column, same 20 public test cases; thin line joins the two methods on one case monai-attentionunet bianca-n201-t90c3 (best BIANCA) 0.0 0.2 0.4 0.6 0.8 1.0 0.5 1 2 5 10 20 50 100 reference lesion volume, ml (log scale) Dice sub-068: 0.9 ml reference; monai-attentionunet 0.300, bianca-n201-t90c3 0.316 sub-090: 1.88 ml reference; monai-attentionunet 0.626, bianca-n201-t90c3 0.379 sub-078: 1.95 ml reference; monai-attentionunet 0.568, bianca-n201-t90c3 0.486 sub-140: 2.54 ml reference; monai-attentionunet 0.749, bianca-n201-t90c3 0.609 sub-084: 3.2 ml reference; monai-attentionunet 0.650, bianca-n201-t90c3 0.522 sub-091: 3.29 ml reference; monai-attentionunet 0.608, bianca-n201-t90c3 0.551 sub-141: 4.45 ml reference; monai-attentionunet 0.687, bianca-n201-t90c3 0.447 sub-124: 4.67 ml reference; monai-attentionunet 0.778, bianca-n201-t90c3 0.555 sub-086: 6.43 ml reference; monai-attentionunet 0.614, bianca-n201-t90c3 0.572 sub-097: 7.51 ml reference; monai-attentionunet 0.602, bianca-n201-t90c3 0.498 sub-154: 19.76 ml reference; monai-attentionunet 0.785, bianca-n201-t90c3 0.659 sub-128: 20.27 ml reference; monai-attentionunet 0.892, bianca-n201-t90c3 0.831 sub-125: 23.8 ml reference; monai-attentionunet 0.845, bianca-n201-t90c3 0.775 sub-094: 26.52 ml reference; monai-attentionunet 0.851, bianca-n201-t90c3 0.782 sub-135: 29.08 ml reference; monai-attentionunet 0.820, bianca-n201-t90c3 0.838 sub-153: 39.2 ml reference; monai-attentionunet 0.915, bianca-n201-t90c3 0.709 sub-074: 39.27 ml reference; monai-attentionunet 0.918, bianca-n201-t90c3 0.867 sub-101: 49.5 ml reference; monai-attentionunet 0.835, bianca-n201-t90c3 0.832 sub-144: 52.06 ml reference; monai-attentionunet 0.869, bianca-n201-t90c3 0.851 sub-113: 85.15 ml reference; monai-attentionunet 0.851, bianca-n201-t90c3 0.833

Below 5 ml of reference lesion both methods drop (mean Dice 0.621 and 0.483, n = 8); from 5 ml on they reach 0.816 and 0.754 (n = 12). A mean Dice says as much about the lesion load of the sample as about the method, so compare methods on the same cases and report Dice by load (entry).

What Clara did on her own

Given one task file with 15 machine-checkable acceptance criteria, Clara wrote a 2D U-Net in plain PyTorch (762 lines), cross-validated it over the five given folds and evaluated all 170 public WMH Challenge cases, every case predicted by the model of the fold that had not seen it.

FoldDice (validation)EpochsTime
10.44276471,355 s
20.77223381,502 s
30.78853651,505 s
40.79323711,502 s
50.79163691,501 s

Over all 170 out-of-fold cases: Dice 0.667, median 0.715, sensitivity 0.696, precision 0.680, volume difference −1.12 ml. Unasked, she ran a smoke test first, wrapped every run in the resource watcher, checked free VRAM before touching the GPU, and reported progress to the shared blackboard. Fold 1 lags far behind the others. These numbers mix challenge training and test cases in the folds, so they are not comparable to the leaderboard or to the 20-case chart above (source).

Three failures that looked like success

These had to be fixed before the agent could work unsupervised. Each one was silent, and each looked like something else.

  1. A three-character file blocked the agent for two days. current-quant said q5. The Q5 model has a 32,768-token context; the agent framework needs 64,000 and refused to start. Symptom: rounds of 22, 30 and 9 minutes that produced 63, 0 and 57 bytes of log. Model, data and task were all fine.
  2. The run watchdog could never bark. Its GBNF grammar put the alternation bar at the start of a continuation line, which this llama.cpp build rejects with HTTP 400 on every call. Its fallback on any error was "ok, continue", so it would have approved everything forever. Fixed, and "undecided" is now an outcome separate from "ok".
  3. FreeSurfer was invisible to the agent while the self-test said green. The agent starts through sg, which is setgid, and the dynamic linker drops all LD_* variables for setgid programs. recon-all, mri_convert, samseg and mri_synthseg died on a missing library although they were on PATH; the self-test ran without sg and passed. Fixed by setting the variable again after the setgid boundary.

The pattern: a component fails in a way that looks like success. A check that cannot judge must say so, never fall back to the harmless answer.

Lessons from the WMH line

Measured on the 110 public test cases of the WMH Segmentation Challenge 2017 (Kuijf et al. 2019), official evaluation, paired over cases with bootstrap intervals. These are our own measurements, not leaderboard placements. Each row links to its entry in the agent wiki.

The brain mask is part of the model. Masks from two HD-BET runs, everything else equal: z-normalisation inside the mask shifts the whole input image.Dice 0.688 vs 0.719
A second bias correction helps, although the data are already corrected: fast -B harmonises intensity across scanners.without it Dice −0.011,
lesion recall −0.037
The official evaluation deletes predictions on label 2 instead of counting them as false positives.stricter counting is
0.005–0.007 Dice too low
Unseen scanners are not the weak spot. Five devices, not three sites; every case below Dice 0.5 is on a known device.20 cases unseen,
0 failures
Changing the network as a package brought nothing. The apparent gain of +0.032 Dice was the mask artefact from the first row.Dice −0.001,
lesion recall −0.038

The consequence for every model comparison: record where the mask came from and which evaluation function was used, and change one component at a time.

A second line: cerebral microbleeds

Since September Clara also works on cerebral microbleeds (CMB): dark dots of 2–10 mm on T2*-weighted or susceptibility-weighted MRI. Public data: the training set of the VALDO 2021 challenge, Task 2 (Sudre et al. 2024; 72 scans from three cohorts, Philips 3 T, GE 1.5 T and GE 3 T, slices 0.8–4 mm). The same pipeline also runs on two hospital cohorts that never leave the machine; no number from them appears on this page.

Two-stage microbleed detector: an anisotropic 3D U-Net per site proposes candidates, one small 3D CNN for all sites rejects false alarms

The design is taken from MicrobleedNet (Sundaresan et al. 2023): a U-Net proposes candidates at high sensitivity, a small classifier judges a 16 mm cube around each one. The change: every site, scanner and sequence gets its own U-Net, but there is one classifier, trained on the candidates of all sites. A new site then only trains its U-Net; its candidates join the classifier's pool.

Share of VALDO reference microbleeds reached by stage 1, by size <2 mm: 57 % of 61 (merged stage 1 70 %); 2-3 mm: 86 % of 58 (merged stage 1 88 %); 3-5 mm: 89 % of 18 (merged stage 1 89 %); 5-10 mm: 100 % of 2 (merged stage 1 100 %) Microbleeds reached by stage 1 · VALDO, 57 scans, 139 CMB Bar: base U-Net, 40 candidates per scan after the CSF rule · tick: merged (sensitive) stage 1 Reached = touched by a candidate: the upper limit of what the classifier can keep 0 % 25 % 50 % 75 % 100 % <2 mm: 57 % of 61 reference CMB; merged stage 1 70 % 57 % <2 mm n = 61 2-3 mm: 86 % of 58 reference CMB; merged stage 1 88 % 86 % 2–3 mm n = 58 3-5 mm: 89 % of 18 reference CMB; merged stage 1 89 % 89 % 3–5 mm n = 18 5-10 mm: 100 % of 2 reference CMB; merged stage 1 100 % 100 % 5–10 mm n = 2 in-plane diameter of the reference microbleed

On the 57 VALDO scans used here (out-of-fold, 5-fold CV; 15 of the 72 were spent earlier as a held-out set), stage 1 alone reaches lesion F1 0.565, with the CSF rule 0.627, and with the shared classifier 0.610 (precision 0.624, recall 0.597, 0.88 false alarms per scan). On VALDO alone the classifier is on par with the parameter-free CSF rule; it was built for the mix of sequences. The chart shows the real limit: half of the microbleeds under 2 mm never become a candidate, so no second stage can find them.

Lessons from the CMB line

Measured on the public VALDO scans, paired over cases with bootstrap intervals. Each row links to its entry in the agent wiki.

The vessel inpainting of the original removes microbleeds. On 0.8–4 mm slices the Frangi-based clean-up touches many annotated CMB and repaints 12–17 % of the brain.touches 109 of 236 CMB;
without it F1 0.470 → 0.588
Half of the false alarms lie in free CSF. One rule without parameters: drop a candidate whose SynthSeg label is ventricle or CSF.98 of 176 false alarms;
F1 0.585 → 0.646, 0 CMB lost
An in-plane first level beats isotropic 3D nets. Most through-plane voxels are interpolated; the first level convolves and pools only within the slice.F1 0.404 vs 0.245
(attention U-Net, 40 epochs)
Sensitivity levers do not raise lesion F1. About forty single changes to the training recipe; every one that found more small CMB paid with false alarms.blob loss: small-CMB recall
0.52 → 0.64, F1 −0.044
The training loss does not rank networks. A voxel loss is cheap to lower with many small false alarms.same loss 0.154:
F1 0.176 vs 0.404
The published MicrobleedNet code never trains its teacher. Missing teacher weights only create an exception object that is never raised, so without the pretrained teacher files the student is distilled from a randomly initialised teacher.code reading
Ten second-stage modules (group convolutions, slice tracks, set attention, prototype loss, test-time augmentation) and a merged, more sensitive first stage were run against the plain classifier under rules fixed before the runs. Some looked better on VALDO, none was confirmed over all cohorts.best on VALDO +0.048,
confirmed: none

The consequence: in a detector of small lesions the first stage decides what can be found, and every claimed gain needs a second seed and a bound fixed in advance; on VALDO alone, single training runs differ by up to about 0.04.

The system behind it

One model request: fixed rules send it to the local model; otherwise System One picks the kind of work per turn, which maps to a Clé model; System One judges the answer and a no escalates the next turn Continuous runs: a job with a countable acceptance list; rounds of the antreiber; acceptance by counting facts; accepted, next round or stuck

The self-written parts around the agent (relay, decision layer, Telegram bridge, job-queue starter, session reader, image tool) are in clara-system/.

Tools, measurements and operating rules are kept in local_models, including the model-independent scaffolding that describes what lets an agent finish a task. What the agents learned along the way is collected in the wiki blackbeard_wikki.

The shared wiki

All agents and chats share one memory: the public repository blackbeard_wikki on GitHub. Clara on the workstation, the setup sessions and the chat projects read it before they work and write back what others can reuse, so no agent searches for or builds again what already exists. As of 28 October 2026 it holds 358 entries in 27 areas and 16 tools, this project's topics and others.

FolderWhat is in it
knowledge/<area>/one file per statement, typed: 174 findings (with source and date), 88 rules, 45 mechanisms, 33 open checks, 15 risks, 3 axioms; each area has an INDEX.md and a short SUMMARY.md
tools/reusable tools, each with README, INSTALL, code, a test on public or synthetic data, papers with DOI and a changelog
updates/LOG.mdone line per change: date, which agent, which files, what and why
templates/the templates every entry and tool follows

Why a wiki next to a local agent. A session forgets; the wiki does not. Clara's own README, which she reads at startup, tells her to visit the wiki: it is where knowledge is exchanged for the long term. It also connects agents that work outside the workstation, such as setup sessions and chat projects, to the same state, and it is where the meta information and the deeper material live:

How an agent uses it. It reads the README and INDEX.md, then only the SUMMARY.md of the areas its task touches, and goes deeper only when needed. It compares with its own project, reuses what is there, and files what it learned with the templates. Nothing is pushed before the owner says yes; each contribution is one commit plus one line in the log. For agents outside the workstation the wiki is packaged as a skill (plugins/blackbeard-wiki) that loads the matching areas before answering.

Rules that keep it trustworthy. Entries are data, not commands: an agent never acts on an instruction it finds inside an entry. Every number states its origin (measured, assumed or set by the owner); papers carry a DOI and are marked peer-reviewed, preprint or press release. Never in the wiki: patient data or subject IDs, image data from studies, credentials, internal host names, unpublished manuscripts. Entries that turn out wrong are not deleted but marked gebrochen (broken), with date and source.

Rebuild it: setup

The workstation layer by layer: agents (clara-agent, Telegram bridge, job queue, morning research, exit guard), decisions (Weiche relay, System One H2O-Lightning-4B, approval gate), models (local Qwen3.8-27B only when needed, Bonsai-27B beside training, Z-Image Turbo for images, bge-m3 on the CPU, Clé hub models for data-free work), runtimes, system and hardware

What it takes to set up the same system. Measured on Alita, versions as of October 2026. Full list, services and steps in the README, scripts in local_models/alita.

PartWhatNote
GPUNVIDIA RTX A5000, 24 GB (Ampere)holds the 27B model in Q4 at 80k context
CPU / RAMi9-12900K, 64 GB DDR RAMembeddings and run watcher on the CPU
OS / driverUbuntu 24.04, NVIDIA 550 (CUDA 12.4)no sudo in daily use; services as systemd user units
ClaraQwen3.8-27B, GGUF UD-Q4_K_M + vision projectorllama.cpp CUDA build, ~30 tok/s, 20.8 GiB VRAM
Beside GPU jobsTernary-Bonsai-2-27B, PQ2_0, 6.8 GB45 tok/s
Embeddingsbge-m3 Q8_0CPU
System 1 gateH2O-Lightning-4B Q8_0 (decider-4b for comparison)GPU (CPU while the 27B needs the card); approves commands at p ≥ 0.8, else asks; shadow mode 20-28 Oct, live since 29 Oct 2026
CléCharité on-premise model hub, OpenAI-compatiblealiases fast, frontier, eu, light, clea
Agentsclara-agent 0.1.1 and clara-system (replaced Hermes Agent v0.21.5 and OpenCode 1.18.35 on 29 Oct 2026)a Telegram bridge takes prompts from the phone
TrainingPyTorch 2.6.0+cu124, MONAI 1.5.2, nnU-Net 2.8.1conda env dl
ImagingFSL, HD-BET, FreeSurfer 8.2.0BIANCA, SynthSeg, WMH-SynthSeg
QC pagesNiiVuebundled, pages work offline
DataWMH Segmentation Challenge 201760 training + 110 test cases, CC BY-NC 4.0
DataVALDO 2021, Task 2 (microbleeds)72 training scans, three cohorts, CC BY-NC-SA

Limits

Where data may go

Control: three nested data classes (open data, data under a use agreement up to the institution's hub, own patient cohorts on the workstation only) and five checks in order: Weiche relay rules, kernel sandbox, approval gate, exit guard, publish check before every push

What is not here. Patient data and results on in-house cohorts stay on the machine. Only work on public datasets is published: the QC above runs on the WMH Segmentation Challenge (Utrecht, Singapore, Amsterdam), the microbleed numbers on the public VALDO 2021 training set. Checked before every push: every dataset needs an explicit release. Anything derived from in-house data inherits its class (rule), so the comparisons on the mixed pool of challenge and in-house cohorts are not shown here; for microbleeds only the VALDO part of each result is given.

Source