Cross-session · Verification · 01

Network-Grounded Truth Sieve

A language model predicts fluent words, not true ones. This sieve ignores fluency entirely: it strips a passage down to its hard factual anchors — proper nouns, dates, numbers — then runs a live audit against Wikipedia to ask one thing of each anchor. Does it actually exist?

Big idea: existence is the first checkpoint — a confident sentence full of invented names fails before truth even comes up.

⚠ Needs internet — this tool runs live Wikipedia lookups. Best for solo or studio investigation; not ideal as a live screen-share on shaky wifi.

0
Isolated anchors
0%
Reality verification
Idle
Network pipeline

fig. 1 — token sieve map

Awaiting raw claims submission…
anchor · entityfiller · fluency

fig. 2 — knowledge-graph cross-reference log

Dispatched query findings map down here…

Field note · why we check the pantry, not the chef

We are not asking the chef whether the dish is good. The sieve takes the names printed on the label — the isolated anchors — and sends a courier to the supply warehouse (Wikipedia) to confirm each vendor actually exists. If the warehouse says "item not found," the chef invented an ingredient. Fluency is not evidence; existence is only the first thing worth checking — for whether real entities were combined truthfully, run the Relational Co-Occurrence Sieve.