How xscan works

Four steps, in order, for every tender. Cheap checks first; the AI court only when the cheap checks find something.

Collect
Real tenders from official APIs and from any website you paste.
Screen
Thirteen free rules score each tender from 0 to 1. Below the threshold it looks normal.
Argue
The prosecutor brings counts with verbatim quotes; the defense answers each one.
Rule and ask
The court rules count by count, sets a risk level, and the clerk writes to the buyer.

The three agents

Prosecutor

charges: 1–6
  quote     verbatim, original language
  category  one of 11 risk types
  severity  low | medium | high
injection_detected: boolean

Defense

rebuttals: one per charge
  stance    contest | partially_concede | concede
  counter_quote   verbatim, or empty
  would_clear_it  what the buyer could show

Court

rulings: one per charge
  ruling    SUSTAINED | DISMISSED | REFER_FOR_CLARIFICATION
  strength  prosecution 0–1, defense 0–1
risk: LOW | MEDIUM | HIGH | CRITICAL, score 0–1

Every quote is matched against the notice text: VERIFIED, APPROXIMATE or NOT_FOUND. A quote the model made up is struck through and the court gives it no weight.

The screening rules

Rule	Looks for	How
`short-window`	Too little time to bid	Days between publication and deadline, against the legal minimum (EU 15, elsewhere 10).
`delivery-window`	Delivery period too short	A delivery or execution period that is very short for the scope.
`brand-lock`	Brand with no equivalent	A named brand or model with no “or equivalent” wording.
`restrictive-spec`	Proprietary specification	Unusually exact or proprietary technical parameters.
`turnover-ratio`	Turnover far above value	Required turnover more than twice the contract value.
`eligibility`	Narrow eligibility	Eligibility tied to a manufacturer, a locality, or prior work for this buyer.
`penalties`	Heavy penalties	Severe or uncapped delay penalties.
`guarantee`	Large guarantees	Bid security or guarantees well above the usual share of value.
`amendment`	Suspicious amendment	A change that shortens the deadline or adds a new restrictive requirement.
`process`	Process gaps	No clarification channel, deferred justification, or price-only evaluation.
`missing-info`	Missing facts	Value, deadline or buyer not published.
`non-competed`	No competition	US awards made without full and open competition.
`injection`	Text aimed at AI	Instructions to automated reviewers inside the notice. Kept as evidence, never obeyed.

Sources

TED
Official EU API. Newest notices first; only the ones not on file are downloaded in full.
USAspending
Official US API. Federal contracts awarded without open competition.
SAM.gov
US tenders. With an API key, or by pasting a link, read in a headless browser.
zakup.gov.kz
Kazakhstan. Read from the portal's own public JSON API: exact official fields.
Any website
Paste a link. Plain pages are read directly, JavaScript pages in headless Chrome; the AI extracts the facts, the tender text stays verbatim.

For developers

GET  /api/v1/tenders                 every tender on file
POST /api/v1/tenders/:id/trials      open a hearing
GET  /api/v1/trials/:id/events       live stages (server-sent events)
POST /api/v1/import/one              { url } or { text }
POST /api/v1/import/listing          { url, max }
POST /api/v1/ingest                  { target, sources }

Models: Meta Muse Spark 1.3 with strict JSON schemas, validated and retried. Tender text is wrapped as untrusted data, so instructions inside a notice are never followed.

What xscan does not do

It never says a buyer is corrupt. It points out warning signs for a person to review, and drafts the question to ask. Every claim it makes can be traced to a quote in the notice.

Watch a hearing