The handoff watchman. A resident mind on your own machine that reads the sign-out as it is being written, checks every sentence against the living donor record — and says one word, wait, before the stale fact is ever sent.
FREE FOREVER. NO ACCOUNT, NO CLOUD, NO TELEMETRY, NO PHONE-HOME. YOU CLONE IT, YOU COMPILE IT, IT RUNS ON YOUR SILICON WITH YOUR WEIGHTS — AND NOTHING IT SEES EVER LEAVES THE BUILDING.
Eleven hours into the shift, the coordinator starts the sign-out. Forty minutes ago the chart moved — the pressor dose changed, the echo was re-read, the OR window shifted. The coordinator is excellent. The coordinator is also human, and the note in their head is the note from eleven o'clock.
The sentence being typed: "Hemodynamically stable on levophed at 8, echo pending, tentative OR 0600."
The record, forty minutes ago: levophed titrated to 14. Echo resulted. OR moved to 0500.
The next team doesn't act on the chart. The next team acts on the note.
VIGIL, mid-sentence, at the boundary of the clause: wait — levophed is 14 as of 02:31, and the OR moved. This sentence is stale.
Handoffs are where continuity breaks — every safety body that has studied clinical communication lands on the same finding. The industry's answer is checklists and read-backs: rules written in advance, for the errors someone already imagined. But the 3 a.m. error is precisely the one nobody pre-named — the fact that changed while the sentence was being written.
You cannot checklist the fact that changed forty minutes ago. You can only be in the room when it's written down wrong.
VIGIL is not something you ask questions. It has no chat box, no prompt, no send key — and you will mostly forget it is there. It is a resident: a small open-weight AI that lives on your own Windows machine, holds the donor record and the forming sign-out in one continuous attention state, and judges — at every completed thought, silently — one question: does what was just written contradict what the record establishes?
Almost always the answer is hold, and you never hear from it. When the answer is wait, a thin banner appears at the seam — before send, citing the record line it's standing on, with one key to dismiss. That's the entire product. The only word it knows is "wait," and the only power it has is to ask a human to look.
The sign-out / handoff note as it is composed, and the live donor record beside it — vitals, meds, labs, timelines, logistics. Word by word, as they land.
The sentence that contradicts the record: the stale dose, the moved window, the resulted lab still described as pending, the number that changed mid-composition.
Nothing. It writes to no chart, files to no system, sends nothing anywhere. Draft-only by construction — it can prevent an error; it cannot cause one.
Everything you have used before checks the note after it exists — on submit, on a timer, on request. VIGIL is built on FUSOR-1, a substrate for minds that don't take turns, and that changes what is physically possible:
Every word of the record and every word of the forming note appends to one persistent attention state the moment it exists. Nothing is ever re-read; there is no refresh, no sync interval, no "checking now." Being current is not something VIGIL does — it's what VIGIL is.
A language model computes, for free, the probability that the current words complete a thought. VIGIL fires its one question exactly there — at the end of each clause, mid-composition — which is where a human colleague's judgment fires too, and for the same reason: that's where a finished claim exists to be checked.
"Stale" is not a vibe. The note says 8; the record says 14 as of 02:31. Every wait cites the specific record line it stands on, so the coordinator's next act isn't trusting an AI — it's glancing at one line of their own chart.
VIGIL is trained default-silent, with a contracted interruption budget: a handful of interruptions per shift, or it isn't doing its job. Every hold and every wait is logged with its confidence margin — so "it stayed quiet" is a decision on the record, not an absence.
Rules catch the errors someone already imagined. VIGIL's entire reason to exist is the other kind — the disqualifying fact that didn't exist when the rule was written:
"Stable on levophed at 8" — titrated to 14 forty minutes ago. The classic. The one the next team medicates against.
"Tentative OR 0600" — moved to 0500 while this paragraph was being written. Every downstream team plans against the wrong hour.
"Echo pending" — it resulted at 02:31, and it changes the conversation. The note ships the old world.
The dose that's right in mg and written in mcg; the weight in pounds entering a kg field. Deterministic checks catch the enumerable ones; VIGIL reads the ones hiding in prose.
Paragraph two contradicts paragraph five of the same note. No system compares a document against itself while it's still being written. A resident holds both in one attention state, so the collision is immediate.
The sequence of events as written couldn't have happened in that order. Not wrong in any single field — wrong in the shape of the story. That's a reading catch, and reading is what the mind is for.
Credibility in this field is mostly about what a tool refuses to do. VIGIL's exclusions are not a roadmap of features coming later. They are the design, permanent — the software has no code paths to these places:
And one boundary that runs the other direction: a human can always overrule it in one keystroke. Dismissals aren't friction — they're how VIGIL learns your OPO's judgment, locally, on your machine, in a record you own.
Every judgment — every wait, every hold, every dismissal, with timestamps and the record line cited — lands on an append-only, hash-chained local log. Not marketing analytics: an audit artifact. Your QAPI team can replay any shift and see exactly what was caught, what was raised, and what a human decided.
Beneath the mind sits a deterministic rule floor — a signed, human-readable policy file your compliance officer can open, read, and edit without touching a model: hard checks that always run, patterns that always flag, and the interruption budget as a written contract.
VIGIL installs in watch-only mode: it surfaces nothing for its first weeks and simply writes the ledger. Your team reviews what it would have said — the morning brief — and turns the banner on only when the record has earned it. Trust is a receipt, not a promise.
Every dismissal and every accepted wait becomes a labeled example — on your machine, in your corpus. The disposition is periodically re-tuned from your own record, so VIGIL converges on your OPO's sense of what warrants an interruption. That record is an asset you own, and it never leaves.
VIGIL rides on the FUSOR-1 resident substrate. These are the substrate's bench ratings — measured, dated, receipted in the reference build's public runs ledger:
| Quantity | Rating | Receipt |
|---|---|---|
| Thought-boundary detection (fire the check exactly where a claim completes) | 0.97 vs 0.02 | ingest logits, free |
| Lag — completed thought → judgment rendered | ~327 ms | workday replay, 1× |
| Honest recall across a working context | ≥ 98,304 tok | planted multi-hop spans |
| Abort a forming output when evidence lands | 13 µs | branch kill |
| Three seats, one attention state | 1.208× | worst-case co-decode |
| Memory fold, self-authored (a day held resident) | 14.8× | all planted facts survive |
SUBSTRATE RATINGS, REFERENCE BENCH: ONE RTX 4070 Ti SUPER (16 GB), 9B-CLASS OPEN WEIGHTS, WINDOWS. DATED RECEIPTS IN THE FUSOR-1 REFERENCE BUILD'S RUNS LEDGER. VIGIL HAS NO CLINICAL PERFORMANCE NUMBERS YET, AND THIS PAGE WILL NOT PRETEND IT DOES — THE SHADOW-MODE LEDGER AT YOUR OWN OPO IS DESIGNED TO PRODUCE THEM, ON YOUR DATA, UNDER YOUR REVIEW.
FUSOR-1 is the open, MIT-licensed resident-kernel substrate — the loop, the bus, and the laws for minds that don't take turns. VIGIL is a fit: the substrate pointed at one seam, with a disposition tuned for it. The engine, the laws, and the receipts live at fusor1.com.
VIGIL is open source under MIT, and deliberately ships as source you build: in a domain where the software reads PHI, "trust us" is not an installer. Clone it, read it, compile it on the machine it will run on, and bring your own open weights. No license server, no account, no vendor in the loop — the same posture as the rest of the AORTA toolkit.
Fitting it to your OPO: VIGIL reads two lanes — the composing note and the record feed — through small, swappable adapters (a watched folder, a text pane, an export stream). The adapters are the part each OPO fits; the engine never changes. The repo ships with a synthetic-donor sandbox — realistic, entirely fabricated records — so you can run the full experience, catches and all, before it ever touches a real workflow.
Everything above is the real fit: a service on your silicon, reading your record, built by you. That is the version that ships. But the argument it rests on — a mind that is already reading, and almost always decides to stay quiet — is not something a screenshot can settle.
So there is a preview that needs no OS, no GPU requirement you have to look up, no build. Open a tab and a small model downloads into your browser and runs on your own graphics card. Type anything. Watch it judge every completed thought and choose silence, and watch the counter that records the silences.
STATUS: REV 0. THE SUBSTRATE IS PUBLIC AND MEASURED; THE VIGIL FIT IS BEING EXTRACTED FROM THE REFERENCE BUILD FOR RELEASE. EVERY DEMO ON THIS SITE IS PROOF-OF-MECHANISM ON SYNTHETIC DATA — THERE IS NO LIVE OPO INSTALLATION TODAY, AND THIS PAGE WILL SAY SO UNTIL THERE IS. STAR THE REPO OR CHECK BACK — EVERY CLAIM ON THIS PAGE STAYS PINNED TO WHAT THE TAPE CAN BACK.
Organ procurement runs on handoffs written by exhausted experts at exactly the hours judgment is hardest. The record almost always contains the truth. VIGIL is the smallest possible thing that closes that gap: a watchman on your own machine, reading along, whose only word is wait.
Built by OPO people. Free because the mission is the point. Honest because the ledger is the product.