Ask a question of more documents than any model can hold at once.
A real body of documents rarely fits in a context window. 1B Context Analysis reads every byte with a chain of model calls, extracts findings into a ledger that points back to the exact bytes each one came from, and answers with an explicit account of how much of the corpus it actually saw.
in developmentSummarizing a corpus that does not fit is how coverage quietly goes missing
Feed a model more documents than its context window holds and something has to give. Passed through a summarizer first, the parts that mattered can be cut before the model ever sees the question. The response reads complete. There is usually no way to tell what got left out.
Read every byte, then say how much was seen
Split into chunks
The corpus is split into overlapping chunks small enough for one model call each. Every byte of every document is assigned to at least one chunk before any question gets asked.
A ledger, not a summary
Each chunk is read by a model call whose findings land in a ledger entry that carries its own exact byte span back into the source document. Nothing is compressed away before it is recorded.
Coverage proof and receipt
The run reports how many bytes were read and how many chunks did not return a verdict, then seals the whole thing into a receipt, so a record dropped after the fact is detectable rather than merely absent.
What full coverage buys you
Every byte is accounted for
Coverage is measured, not assumed. A run reports the exact share of the corpus its chunks actually read, not the share it was asked to read.
Silence is counted, not assumed empty
A document that tells the reader to stay quiet produces a valid, empty answer that ordinary coverage metrics read as a blank page. Every segment must return a verdict, or it is counted as unaccounted for.
Every record gets judged
There is no similarity search deciding what the model gets to see. Every extracted record is judged against the question directly, so "we considered everything" is checkable instead of a claim.
A run commits to what it did
Every layer of the run seals into a receipt root, so a record dropped after the fact is detectable rather than merely missing from the final answer.
1B Context Analysis runs inside SafeRouter's infrastructure today. It opens to accounts as the ledger and receipt format finish hardening. Want a heads-up when it does? Write to hello@sumplus.xyz.