Science#
Ingence makes claims about files, not about science. This page says where the work around it stands: the technical report, how the engine is evaluated, the datasets planned, and what a demonstration is and is not.
The technical report#
A technical report on how Ingence works and how it is evaluated is planned. It will be an open preprint, a paper posted openly before journal review, under a Creative Commons Attribution license (CC BY 4.0), which lets anyone reuse it if they credit the author. It has not been posted. When it is, this page will link to it.
How the engine is evaluated#
The engine is checked on invented laboratories with a written answer key. A scorer counts what each rule concluded against the key, for each kind of question. It reports counts with their totals and an interval around each proportion. Saying nothing is not counted as a mistake, and it is not hidden: how many answers were found is shown beside how many were right.
There is no held-out measurement on real laboratories yet, and this site claims no figure for how often Ingence is right on a real drive. What the evaluation says and does not say gives the method, the limits and a command to run the public check yourself.
What Ingence does and does not claim#
Every conclusion Ingence reaches names a rule, written in plain words, and can be rebuilt from four inputs: what was observed, the rules, the Dictionary of words and your decisions (How Ingence thinks). No language model writes a conclusion, and none is needed to read one.
What it produces are proposals and readings of how files relate. It does not claim that what it finds is scientifically valid, and it offers no scientific interpretation.
Datasets#
Two datasets are planned. Neither has been published.
The sample collection, under CC0 1.0, which lets anyone use it for any purpose. It is a shared drive of an invented laboratory over several years, in three sizes, each with its answer key, the generator’s version and seed, and a datasheet. Everything in it is invented. A command that ships with Ingence generates it. It is described on the evaluation page.
De-identified graph outputs, also under CC0 1.0: the graphs Ingence builds over real folders, with names replaced by codes and dates and sizes blurred. This is planned and not built: Ingence cannot yet write one. The plan has two limits. The shape of the folder tree and the counts of files are still published, so someone who holds the same files could recognize them. And the protection for the people whose folders are in it is consent, not the coding: nothing from a real laboratory is released unless every member whose work shows in it agrees, or their folder is left out. Until then, no real laboratory’s data is public.
Demonstrations are not scored results#
A demonstration shows what Ingence does: a recorded tour, a worked example, a public archive read from its list of files. It is not a measurement. A demonstration is chosen to be shown, and nobody wrote an answer key for it first. Only a run scored against a key, reported with its counts and intervals, is a result. Where this site shows a result, it says what it was scored against.
Reproducing a run#
Give Ingence the same four inputs and it reaches the same conclusions. Every run records the version of the engine that made it, and its exact commit where the engine came from a checkout or a build that carries one. The pool of invented files that the public check reads is written from a seed, and the same seed gives the same files each time (the evaluation page).
To cite Ingence, see About.