research
every post and study we have published, newest first. negative and incomplete results included.
november 2026
- a 25 ms commit window for t3 code: 22% fewer sqlite commits on real agent turnst3 code commits once per streamed event. grouping plain provider updates for up to 25 ms cut commits per stored event by 22% on 156 real co…
- the free stack every contributor already has17 free llm lanes, used all at once: 149 small tasks in 1.4 s. a case study in speeding up open-source work.
- a chart is not a web pagesame model, same data: an openui board took 7.2× fewer output tokens and finished 2.5× sooner than an html page.
october 2026
- 100 Android chats: a send is not a captured replyAn operational run of the reusable uiautomator2 workflow, with dispatch, result capture, and verification kept separate.
- what we have sent to cua, and what is still only a fixturea living record of our trycua/cua work: merged fixes, open rfcs, and experiments that have not earned a claim yet.
- the guard passed. the gpu did not help.bend agreed with the oracle. the best cpu algorithm made the large-batch chart beside the point.
- the memory study, october updatethe real-source gate stayed closed, the handoff froze, and the replay caught a real error.
- offline fan-out is not a phone fleetthe git rollback was real. the screens and preview failures were synthetic.
- the proof passed. the work was still stranded.three statecharts, then a replay of the things our agents actually left behind.
- same model, four harnessesthe fast profile matters. so does the difference between an engine and its cli.
- perfect recall, stale firstthree native retrieval surfaces found the fixture. none put the new fact first.
- we predicted the next action. nobody committed it.the frozen baseline is a negative result. the pull-first replacement is still a proposal.
- an agent api that does not steal the foregroundthe bounded phone run worked. that says nothing about the quality of health advice.
- voice parity failed. the fixes still mattered.we scored the audio, kept the failures, and did not call a partial rerun a pass.
- the tap was fast. proving it landed cost more.we kept the lexical matcher and did not ship the on-phone model tier.
- studydoes optchat stay small after fifty tool calls?the transcript kept growing and the derived view stayed smaller. context bytes only, no answer-quality claim.
- studycan four agents remember the same thing?optmem foundations landed and optchat split off; the four-harness recall cohort is still partial.
- claude code savings, re-measured496 pre-registered runs against the current z0 integration: one lever holds everywhere, one depends on the task, one barely shows up.
september 2026
- the boundary worked. the model study did not run.deterministic integration evidence is useful. it is not an answer-quality result.
- where do claude code tokens actually go?one night of dogfooding z0intelligence inside claude code, measured with reported usage and checked outcomes.
- three small decisions, three pre-registered barsno small model beats the 8b at forming z0's own calls, and 'auto-merge on green' fails on our own history.
- four pre-registered bars, four missesa keyless gpu tier below its bar, a learned router that loses to a generic one, and an authority check that belongs on the tool call.
- a format check is not a faithfulness checka local worker passed our offload bar, then failed on faithfulness. the learning loop does not have the data yet.
- studyhow do you learn the frontier fast?the research trail behind z0intelligence: classifiers, calibration, routing, conditional compute and tiny specialists.
- studywhat actually deserves a model?we kept trying to make routing smarter. the useful part was learning which decisions should not be model problems.
- How z0evals Fits Into Zer0a federated evaluation layer built from the fixtures and contracts the rest of the zer0 ecosystem already owns.
- studyhow much of the llm do we actually need?phase 1b: which local model routes best, and why the compiler turned out to be the intelligence.