Trained candidateCavumo Context 1
Working today: permission-scoped hybrid context resolution. Research: learned selection under a token budget. The candidate remains inactive pending its research gate.
Not yet validatedIntelligence, built
with evidence.
Systems research, specialist models, and the engineering of private intelligence. Explore what we’ve built, what we’ve measured, and what we’re testing next.
Local MVP runningOriginal engineering notesPre-release research
01 / FROM THE LAB
Ideas meet implementation.
CAVUMO / COMPILER
SCOPED STUDY / p95 LATENCY
Fixed4,983 ms
Compiled2,669 ms
Same model. Identical generated outputs.
01Profile + identity
02Permission check
03Bounded plan
04Local execution
Systems study
Less repeated work. More useful execution.
Inside a paired Gemma study of resource-bounded planning and guarded reuse on an 8 GiB Mac.
CAVUMO / TRAINING
GEMMA + MLXLOCAL SYSTEM
01Reviewed data
02Training grant
03Local LoRA
04Held-out evaluation
Engineering note
Learning locally, with permission.
Real Gemma LoRA fitting, bounded training grants, and why lower loss is only the beginning.
CAVUMO / CONTEXT
CONTEXT 1RESEARCH CANDIDATE
01Permitted sources
02Candidate selection
03Budget check
04Evidence bundle
Research notebook
The context we keep. The evidence we need.
Learned selectors, strong clause baselines, and the decision to keep a candidate inactive.
CAVUMO / MEMORY
ORG BRAINLOCAL SYSTEM
01Agent proposes
02Owner reviews
03Scoped knowledge
04Authorized agent
Architecture note
Shared understanding needs boundaries.
How reviewed memory and Org Brain connect agents without making every source available to everyone.
02 / SPECIALIST PROGRAM
Our models. A focused purpose.
Complementing compatible base models with context selection, evidence assessment, and improvement recommendations.
Trained candidate
Baseline implementedCavumo Verify 1
Working today: exact cited-quote checks against fresh permitted evidence. Research: a trained semantic-support model with calibrated uncertainty; not trained yet.
Not yet validated
Baseline implementedCavumo Adapt 1
Working today: explicit-rule guidance for reviewed failures. Research: a learned experiment-outcome predictor; not trained yet. Advice cannot authorize training.
Not yet validatedPart of a wider body of work.
External papers by their respective authors. Cavumo’s notes above are internal engineering reports, not peer-reviewed publications.