Vigil — Researches, Learns, and Shows Its Work

Vigil is an agnostic, non-parametric Proculmen multiplier: the summit advances through being reached.

Give Vigil a question, statement, claim, subject, document, or image. With permission, it searches for evidence, learns what the evidence actually separates, and shows every source behind what it found. Keep going and it learns. When evidence cannot carry something, Vigil returns the exact gap instead of guessing.

Vigil returns a structured web of evidence: the distinctions it finds, the data supporting them, and the relationships or gaps between them. Bring that result into the next exchange with new ground and Vigil recursively updates it through time. We call this evolving, inspectable structure the living map. You can inspect it, edit it, download it, and carry it forward; Vigil never keeps it from you.

For technically inclined readers: the map is a monadic, recursively updated relational matrix over accrued evidence. Each exchange receives the prior structure and new ground, then returns the updated structure together with what remains unresolved.

You do not need to arrive with labels or a codebook. Permit a web search or supply material of your own and Vigil proposes the first map, showing the source passages behind every distinction. On your own computer, sovereign Vigil can also read the files and drives you permit. The Apify edition cannot see those drives unless you explicitly upload or connect them.

If you already have a map, you may supply it directly — its labels and the phrases your material actually uses for them. Vigil reads your passages against it and returns three things: what carried, what didn't, and the question the leftovers raise.

Every coded row carries the verbatim sentence that produced it. Not a score, not a probability — the words in your text that the label was admitted on.

What it does with what it cannot place

A passage that reaches two of your labels equally is not an error and not a tie to be broken. It is the one part of your material your map does not yet separate, so it comes back naming which two, as the next question — generated from that pair, never composed.

A passage that reaches a label while carrying none of its phrases comes back too, asking whether the label is missing a phrase your material uses or does not belong to this text.

Nothing here selects what happens next. The leftovers do.

No map yet?

Run it with mode: derive and a question. Permit web gathering and Vigil searches for addressed ground, or offer documents/files directly when you want to choose the ground yourself. It splits what it receives by what the passages share and names each group from its own distinctive words, showing you which passages each distinction came from. Rename them, merge two that are really one, drop what isn't a distinction — then code with it.

Every proposed label is made of your words. Nothing is brought in.

What it will not do

• It will not invent a label or rename yours. It finds distinctions; naming them is yours.
• It will not quote a word you did not send. Checked on every response, not promised.
• It will not grade your material. What you say is evidence, not a claim to be assessed.
• It will not grade against nothing. Ask it to check text with no sources supplied and it says so, rather than returning something that reads like a finding.

A picture instead of typed text

Send a screenshot or a scan and it reads the text out of it. What it reads comes back labelled as a reading — never quoted as your words, never added to your map. The eye reads every surface it covers clearly; the label preserves whose words are whose. If you want the text received as your own direct evidence, type it, which supplies it in your voice rather than changing the provenance of the image.

Marks it cannot name are not dropped. They come back grouped by what recurs, so you can name one and have every member named with it.

It keeps nothing

No model and no weights file — the loop is plain Python. It makes an outbound request only when you explicitly permit web gathering; the received source addresses come home in the result. Your map arrives with the run, grows inside it, and goes home in the output. Nothing is stored between runs and nothing is keyed to you.

Every result also returns a complete application receipt. Paste that receipt into priorApplications on a later run and its unresolved spans re-enter as fresh ground, while its ID records which earlier application was received. The receipt includes the exact passages, bytes, and bits offered to that run. This is caller-carried recurrence, not a hidden account or server memory.

Across a session you can prove the thing that matters: the same frame, more ground. The frame tag on every response is cut to exclude accrued examples, so it stays identical while your evidence grows — if it changed, your map changed, and you would see it.

Pricing — in mouse arithmetic

You do not pay because the machine opened its door. You pay for what comes home in your hand.

What comes home	Price
One proposed distinction in a first lens, with its marker words and source passages	$1.00
One returned reading row — a quoted admission, named open address, or labelled image reading	$0.10
Optional source check: one sentence checked against evidence you supplied	$0.10
Optional source check: an additional catch where that evidence does not carry the sentence	$1.00

A first lens with four proposed distinctions costs $4.00. A later turn returning twelve rows costs $1.20. You can count both from the work returned. Starting, empty input, refusal, and a genesis that proposes nothing are free. The unknown material is not sold by size: an open-address row costs ten cents because the useful address formed, not because your remainder belongs to us.

And the large run, said out loud rather than left for your invoice. Genesis charges a dollar for every distinction it actually forms, and how many it forms is set by your material — nothing in the input caps it. A first lens over a large body can return thirty or forty distinctions and cost thirty or forty dollars. That is the map you asked for, and it is well under what the same first pass costs in anyone's time. But you cannot know the figure before the run, and a bill you cannot predict is a worse barrier than a bill that is merely large. So here is how to hold it: offer a slice first — one document, one interview — read what comes back, and widen only when the map is worth widening. Genesis is paid once per body of material; every later turn against that lens is rows at ten cents.

A distinction carrying no marker phrases costs the same dollar. Markers are the refusal instrument, and a label with none was admitted on agreement alone — the weaker guarantee. The run names those labels in the same breath, so you can see exactly which they are. The price does not drop for them.

Honest edges

• Genesis is the expensive step. Proposing a first map costs real compute and scales with how much ground is received; a later turn against an existing map is fast. You pay genesis once per body of material.
• The eye sees every surface it covers clearly and keeps provenance intact. Image readings are labelled and kept out of your map because Vigil reading a surface is not the same act as you supplying those words directly.
• Marker phrases are the refusal instrument. A label carrying none can still be admitted, on agreement alone — the run names those labels for you rather than letting the weaker guarantee pass unmentioned.