run 1 · what alkor ships
With alkor
%
19 of 20 pass
The model filling alkor’s form, with the arithmetic done in code. The one miss named a category for a hypotension too brief to qualify.
Structuredsyndromic assessmentson your laptop
Give alkor a clinical note. It returns an assessment of shock, sepsis, volume overload and a focal lung process.Each verdict comes with the findings behind it and a list of what the note never documented. The AI models that do the reading run locally on your computer, so the note never leaves it.
an assessmentone synthetic note · four syndromes
What you get back
One note in, one assessment out. Each syndrome gets a verdict, the words in the note behind it, and the findings nobody documented. Where the note cannot support a verdict, the answer says so.
Medical take, second review.
Hypotensive for two and a half hours, currently 77 over 37, heart rate 128.
CORRECTION to my earlier entry. I recorded the peripheries as cool on first assessment. On review
with the registrar they are warm to the wrist and that earlier entry was wrong — she is warm
and vasodilated. Capillary refill brisk. Pulse bounding.
Jugular venous pressure normal to low. There are crackles at both bases.
the assessment
MAP 50shock index 1.66calculated in code from the documented vitals
- Shock
- Septic pattern
- Warm peripheries, brisk refill, bounding pulse, JVP normal to low, hypotension for 2.5 hours. The note’s correction is followed: the earlier “cool” is set aside.
- Nothing missing
- Sepsis screen
- qSOFA 1 · negative
- Systolic 77 scores the one point.
- Not documented: respiratory rate, mental status
- Volume status
- Indeterminate
- Only two of the seven findings were examined: JVP normal to low, and crackles at both bases.
- Not documented: S3, hepatojugular reflux, ankle oedema, hepatomegaly, pleural effusion
- Focal pulmonary
- Unsupported
- Crackles on both sides, and no imaging to place them.
- Not documented: imaging
A synthetic note written for testing, not a patient. The assessment is abridged for this page.
who it is forresearch · teaching · documentation
For reading notes, not for treating patients
alkor is a research and teaching tool. It is not a medical device and has no place in a patient’s care. These are the jobs it is built for.
Chart review
Screen a folder of notes for shock or sepsis criteria before a person reads them, with the evidence for every verdict attached.
Documentation audit
Every assessment lists what the note never documented. Across many notes, that shows what gets examined and what does not.
Teaching
A student writes their own reading of a case, then compares it with the rule’s, finding by finding.
Testing local AI
See how a small model on your own hardware handles your own notes, without sending them anywhere.
privacyno cloud · no account · nothing sent
The note never leaves your computer
Every model alkor uses runs on the machine in front of you. There is no cloud service behind it and no account to create. Once the models are downloaded, a run needs no internet connection.
- Runs on
- A laptop
- Built and measured on an 8-core laptop processor with no separate graphics card.
- Time per note
- Minutes
- Several models read and check in turn, one after another. The dashboard shows how far a run has got.
- Download
- 4.6 GB
- The main model, fetched once on first start.
- Cost
- Free
- Apache-2.0 at release. The clinical rules are CC BY 4.0.
measured · shockQwen3-4B · 20 synthetic cases · 2026-09-07
The same model, made to follow the rule
One small model, twenty shock cases, one answer key. Alone, it agreed with the rule 5 times; through alkor, 19. The only thing that changed was what alkor puts around it.
19 / 20
agree with the rule through alkor. The same model, alone: 5 of 20
run 2 · the model by itself
The model alone
%
5 of 20 fail
The case file, then the question. This is what downloading the model yourself gets you: 14 of 20 replies name two or three categories at once.
run 3 · the four answers listed
One word from a list
%
9 of 20 fail
The question plus the four allowed answers. It said “indeterminate” all twenty times, which is what that constant scores. With the list reversed it scored 4 of 20.
It says when the note is not enough
In nine of the twenty cases the rule cannot decide. Through alkor the model said so in eight of them. Alone it said so in none: it named a category every time, including for a normal blood pressure. The 5 of 20 is the generous grading, scored on its first answer; scored on its last, it is 4.
What this does not show
Twenty synthetic cases, not patients. Agreement means agreeing with one published bedside rule, which was right about 76% of the time in its own cohort. The rule alone, in code, scores 20 of 20; what alkor adds is the evidence and the reason attached to every answer. Runs 2 and 3 came from scripts no longer in the repository and cannot be reproduced. The current default model, Gemma 4 E4B, scores 20 of 20 through alkor and has not been run without it.
under the hoodseveral models · one machine
Small models, each with one job
A run is a series of small jobs, each handed to the model suited to it. alkor itself knows no medicine: the clinical rules live in files it loads, so another specialty can be added without rewriting it.
alkor runs on one machine and coordinates several local models: one reads the note, a second checks that reading against the note, and a small router model, kept loaded, picks the workflow. Rules in code pick the syndromes, compute, check and refuse.
-
Each step sees only what it needs
Every call to a model carries the note and the findings for that step, and nothing from earlier calls. A small model can manage a job that size.
-
Models take turns
Only the small routing model stays loaded. The others are loaded when their step comes and put away after, so four models never need to fit in memory at once.
-
What was measured is what runs
Testing and real use run the same code, so a score describes what you actually run. Every figure names the exact model file that produced it.
-
One note or a thousand
The dashboard is optional. A script can send a folder of notes one after another; each comes back as an assessment or a refusal, with an id to look up its full record later.
the dashboardthe run, as it happens
Watch every step, and see why a run stopped
A run takes minutes, and without a window into it those minutes are a closed box. alkor publishes what it is doing as it does it. The dashboard only watches and cannot change what the run decides.
The dashboard
The run, drawn live, for the minutes it takes
One command brings it up in your browser: the runs you have sent across the top, the steps in the middle turning over as each finishes, and a side panel with the note, the assessment and the event log. A run that stopped is obvious at a glance: the graph ends at the step that failed, with the reason beside it.
one command · in your browser
The test report
One file per test, readable offline
When alkor is tested against a set of cases, it writes a single page with every score and, beside each one, what that score does not show. It opens in any browser, with nothing to install.
one file · nothing to install
questionsasked first
Questions
Is alkor a medical device?
No. It is a research and teaching tool. It has not been reviewed or cleared by any regulator, it is not clinical decision support, and its output must not be used to diagnose or treat anyone.
Does any patient data leave my computer?
No. The models run on your own machine and alkor sends nothing anywhere. The only download is the models themselves, once.
What computer do I need?
It is built and measured on an ordinary laptop: an 8-core processor, no separate graphics card. A faster machine finishes sooner. The minimum memory will be stated with the release.
Which AI model does it use?
By default Gemma 4 E4B, an open model from Google small enough for a laptop. Other open models can be swapped in, and every figure alkor reports names the exact model file that produced it.
Can it be wrong?
Yes. A small model misses findings and invents plausible ones. alkor checks each reading against the note with a second model, does the arithmetic in code, and refuses the run when a check fails rather than returning it. What passes is still a model’s reading of a note, and a clinician has to review it.
Can it assess other syndromes?
Four today: shock, sepsis, volume overload and a focal lung process. The clinical rules are files alkor loads, so more can be added without changing alkor. Each will appear here once it has been measured.
availabilityalpha · free at release
Not public yet
alkor runs today on the machine it is built on. It will be released once the clinical rules and the refusal logic stop changing.
What the release will be
Everything on this page, installed with one command on your own computer, free under the Apache-2.0 licence. alkor runs its models on llama.cpp, a free AI engine you install alongside it.