Filters bound, hints suggest
--memory-type, --subject, --tag, --since and --until restrict what is eligible. They are the difference
between a useful answer and a ranking where “what happened in May” competes with “define a Fourier series”.
--mode is a hint. It narrows the admissible plans; it never selects one. --mode semantic still admits a term scan,
because no index in this protocol is authoritative and a hint must not become a way to make one so.
One filter rule is inherited from the protocol and surprises people: a block with no occurred_at fails a time
window. Not “passes by default” — fails. --since over a module of undated semantic blocks returns nothing, and
that is correct.
Both query search and query explain accept YYYY-MM-DD calendar bounds as well as protocol timestamps.
--since 2026-05-01 expands to 2026-05-01T00:00:00Z, and --until 2026-05-31 includes the whole day through
2026-05-31T23:59:59Z. Bounds are inclusive in UTC; a reversed interval is rejected.
DateScan when the selected modules have that index. Results are
ordered newest first across modules, with block identity breaking timestamp ties. Additional filters, accessibility,
and verification still apply before the result limit. Without BTree coverage, the planner can scan the blocks.
What comes back is an evidence bundle
scoreis agreement between retrieval strategies. Not a probability, not a confidence. A string in the JSON, at the protocol’s precision — the ordering was decided on the full float before rendering, so two matches showing1.00are not tied.truncatedistruewhenever a candidate that passed every filter was discarded, including one cut by a generator’sk. Check it before ever saying “the brain does not contain X”.superseded_bymeans something replaced it. Cite the successor, or say you are quoting a superseded claim.resolvable: falsemeans redacted or not installed — not corrupt. The block is still a verifiable member.
verified: false
makes it look like a low-quality result.
The CLI will never write your answer
There is noanswer field, and there will not be one. The brain returns evidence; synthesising prose is the caller’s
job. A CLI that summarised for you would have quietly become the model, and the summary would carry none of the
verification the bundle exists to provide.