What 7 of 8 does and does not mean
On one locked synthetic fixture, Noru's selector surfaced relevant prior evidence for 7 of 8 answerable queries and abstained on 2 of 2 no-answer queries. We did not test agent task outcomes.
A promising number is most useful when its boundary stays attached.
We evaluated Noru's contextual-recall selector on a fixed synthetic corpus. The purpose was to test a mechanism: can the selector surface labelled prior evidence without filling the agent's context with an unlimited history dump?
It was not a test of whether an agent completed more work.
What we ran
The fixture contained 10 unique retrieval cases and 32 synthetic capture bundles. Eight queries had relevant prior evidence. Two no-answer queries were designed to test abstention. Five repeated selector passes produced 50 case-runs for local latency and determinism checks; those repeats were not treated as independent evidence.
| Measure | Result |
|---|---|
| Relevant evidence in the top three, answerable cases | 7 / 8 |
| Correct no-answer abstentions | 2 / 2 |
| Returned-candidate precision after the temporal fix | 9 / 12 |
| Future candidates after the temporal fix | 0 / 12 |
The fixture was synthetic. The embedding configuration was deterministic. The sample was small. Those facts are part of the result, not footnotes to remove when the number becomes a headline.
What failed
The missed answerable case relied on weak OCR and visual similarity. The implemented selector uses visual distance to reorder candidates admitted through text; it does not yet provide an independent visual candidate path. That limitation explains the miss and defines a concrete next experiment.
The initial run also returned one future-timestamped candidate at rank one. We added a strict query-time cutoff and reran the same locked fixture. Future candidates fell from one to zero, Hit@3 remained 7 of 8, and correct abstentions remained 2 of 2. Because the fixture had already been seen, this is regression evidence for the fix, not independent confirmation.
What we did not run
We planned paired agent runs across current perception, manual history retrieval, and automatic hints. Zero paired agent runs were executed. The safe isolation path needed to combine fresh observation with an ephemeral synthetic history corpus was not yet qualified, and weakening the production guard would have made the comparison less trustworthy.
So the public decision remains simple: do not claim that Noru improves agent task performance from this evaluation. We measured whether a selector found labelled items in one controlled fixture. We did not measure completed tasks, intervention rates, or improvement across clients and models.
The supported conclusion
On this locked synthetic fixture, Noru's selector returned relevant prior evidence among at most three candidates for 7 of 8 answerable queries and correctly abstained on 2 of 2 no-answer queries. After the temporal cutoff, it returned zero future candidates in the same fixture.
A result becomes more credible, not less ambitious, when it says exactly what was measured and stops there.