Noru Flow Research & Notes https://noruflow.com/research/ Research syntheses, product notes, and Noru evaluations on seeing, memory, evidence, and control. en-us Wed, 12 Aug 2026 12:00:00 GMT A screenshot is evidence, not permission https://noruflow.com/research/a-screenshot-is-evidence-not-permission/ https://noruflow.com/research/a-screenshot-is-evidence-not-permission/ Wed, 12 Aug 2026 12:00:00 GMT Noru Flow Boundaries Letting an agent observe the screen should not quietly grant it the right to act. Noru keeps observation local, visible, and separate from control. Why old context cannot answer “now” https://noruflow.com/research/why-old-context-cannot-answer-now/ https://noruflow.com/research/why-old-context-cannot-answer-now/ Wed, 12 Aug 2026 12:00:00 GMT Noru Flow Evidence History can explain what happened before. It cannot prove what is on screen now, and a failed fresh observation must never be replaced with an old one. Memory should return evidence, not confidence https://noruflow.com/research/memory-should-return-evidence-not-confidence/ https://noruflow.com/research/memory-should-return-evidence-not-confidence/ Wed, 12 Aug 2026 12:00:00 GMT Noru Flow Memory Noru Pro memory is retrieval over local captures, not training. It can suggest a small number of prior observations, preserve their provenance, and return nothing when nothing earns its place. What 7 of 8 does and does not mean https://noruflow.com/research/what-seven-of-eight-does-and-does-not-mean/ https://noruflow.com/research/what-seven-of-eight-does-and-does-not-mean/ Wed, 12 Aug 2026 12:00:00 GMT Noru Flow Evidence 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. Does the model actually need to see it? https://noruflow.com/research/does-the-model-need-to-see-it/ https://noruflow.com/research/does-the-model-need-to-see-it/ Mon, 01 Jun 2026 12:00:00 GMT Noru Flow Seeing In one GPT-4o ChartQA comparison, accuracy was 85.7% with the chart and 30.0% without it. When an answer depends on visual relationships, the image carries evidence a description may omit. Why the screenshot needs a text layer https://noruflow.com/research/why-the-text-layer-earns-its-place/ https://noruflow.com/research/why-the-text-layer-earns-its-place/ Mon, 01 Jun 2026 12:00:00 GMT Noru Flow Seeing At constrained image budgets, OCR can restore characters the pixels lose while the screenshot preserves layout. The evidence supports a paired payload, not OCR as ground truth. How much screenshot is enough? https://noruflow.com/research/theres-a-right-size-for-a-screenshot/ https://noruflow.com/research/theres-a-right-size-for-a-screenshot/ Mon, 01 Jun 2026 12:00:00 GMT Noru Flow Seeing In one Qwen2-VL resolution sweep, InfographicVQA rose from 28.9% to 77.3%, while fine-OCR peaked earlier. More pixels can help, but the useful budget depends on the model and task. The same context, in a better order https://noruflow.com/research/label-it-and-put-it-in-order/ https://noruflow.com/research/label-it-and-put-it-in-order/ Mon, 01 Jun 2026 12:00:00 GMT Noru Flow Seeing In VisualWebArena's tested GPT-4V baseline, adding screenshots raised success from 7.3% to 15.1%; Set-of-Marks reached 16.4%. The result supports structured visual context, but Set-of-Marks is not a Noru feature.