depth

search lanes

account

how a run is priced

Every question is planned first by the engine's own no-cost planner. The plan, the crew and the price band land in a confirm step before a single paid call fires. The receipt after the run reports the real number the server charged, not the estimate.

two axes, never collapsed

Where it came from is brought or derived. Brought means you put it in. Derived means a model wrote it. That axis says nothing about whether it is true.

Whether it holds up is unchecked, unverified or flagged, and it is only ever earned by a check that actually ran. Material you brought in starts unchecked exactly like material a model wrote, and it is scanned the moment it lands, because pasted internet and other models' output is the highest risk input on this page, not the lowest.

Nothing on this page can move anything to verified. Verified appears only on the engine's own class zero evidence, which carries a source and an access date, and even that stays marked unchecked because nothing here opens the page to confirm it.

what the tile actually draws

The anatomy strip is measured from the text that was really extracted, band by band down the document. It is never inferred from a filename or a file type. A document that could not be read shows a hollow strip and says so, because a confident looking picture of something nobody read is worse than a wrong sentence: it skips the reading.

what the scan checks

Accuracy, mechanically: links parse as real web addresses, a citation's domain field agrees with its link, duplicates, and how many numeric claims carry no link at all.

Laws and vision, mechanically: banned pattern strings, each reported against the section it breaks. Trials, urgency and funnel mechanics, up front annual charges, unlimited offers, margins under the floor, emoji in shipped copy, gradients on platform chrome, the shekels spelling.

Secrets and intellectual property, on value shape and never on variable name, so documentation naming a key does not cry wolf. Anything found is reported and never silently removed, because stripping a document lies about what the document contains.

what the scan does not check

Whether a citation says what the claim says. A browser cannot read another site's page, so every citation stays unchecked. Open it and judge it.

Whether a claim about our own architecture is true. That needs the repository and pricing.ts, and nothing on this page can read either, so an outside model's confident description of our router or our file layout is unchecked, never passing.

A second model's opinion. No cross family check runs from here. Where the material came from is recorded on every document so a check can never grade its own family, but that check is not wired.

knowledge bases

A base is a saved view, not a folder. A finding that belongs to two bases sits in both, one copy, and moving nothing. The base selected when you pin stamps that pin, and anything whose text matches an existing base's words joins it too. A base is never invented for you.

A base never picks a winner between documents that disagree. Where its records conflict, it says so and shows them.