detent
Ecosystem/07 of 08

Label Pipeline and the Corpus

Why it is in the ecosystem. The ground truth factory. Agent security has no natural chargeback, so we manufacture one: every receipt is stamped with exactly one of five outcome labels, and the labelled corpus is what trains the watchdog and prices the risk.

labels REALcorpus REALwho pays: nobody directly. this is the asset the other seven cards feed

This is the asset the other seven cards feed. No classifier is trained on it yet and the page says so: the watchdog is rules and statistics until one auditable receipt shows a learned signal beating a held-out baseline.

The five labels

REAL
  1. 01confirmed lossvalue left and did not come back. the deny was wrong, or there was no deny. the only label that costs money.counting
  2. 02human overridea STEP-UP was answered by a person. the machine asked, the owner decided. demonstrated live in tour 3.counting
  3. 03simulation mismatchthe pre-execution simulation and the intent disagreed: the fill would not have been what the human named.counting
  4. 04policy breachthe signed policy was broken: cap, allowlist, window, shape, or a definition that changed after approval.counting
  5. 05counterparty disputethe other side disputed the outcome after the fact. the label that turns a payment into a case.counting

one slot per receipt, empty at issuance, stamped once. a receipt can never carry two.

The corpus, live

Labelled receiptsREAL
reading the corpus
source · receipt_labels + tour_labels
By ownersREAL
reading
source · signed-in accounts
By visitorsREAL
reading
source · tour replays, session-bound
Visitor sessionsREAL
reading
source · distinct sessions that stamped
share by label

no labels yet. run a replay in tour 2 and the first bar appears here.

Recent labels

newest first · click a row to follow it
WhenLabelByVerdictTop reasonReceipt
reading the pipeline

Follow one receipt

REAL

pick a row above. the trace shows the verdict, the label that judged the verdict, and what the label moved.

the aha: Every refusal becomes training data with a verdict on the verdict.