What we measure
Questions we test in the laboratory—small in team size, serious about controls—so peers can disagree productively.
General-purpose language models are strong at dialogue and weak at accountable fact in regulated work. Our laboratory compares many evaluation approaches—benchmarked runs, reviewer-approved knowledge, and controlled assistance patterns—so we can see what holds up in complex fields that cannot rely on undifferentiated “chat” alone. We evaluate versions, keep what measurably helps, and discard what does not.
How we evaluate
After a controlled run through the Explico® environment, we can apply the same reproducible outcome to labeled experimental builds (informally versioned as v1, v2, and so on). That is lab machinery, not a product SKU. Comparisons feed internal and partner reports unless we publish results with permission. We contribute mainly by sharing evaluation discipline and evolving designs with collaborators—not by claiming a large standalone research institute.
Software and the machine it runs on
Software-only teams often rediscover throughput, isolation, and the cost of being wrong. We try to shape releases to the envelope they run in—security and repeatability from the first constraint. Topology details stay in engineering agreements; the principle here is simple: controlled outcomes need honest limits, not slogans.
Topics on the bench (not all active at once)
We may explore hybrid assistance patterns, efficiency-oriented models, and routing that spends depth only where it pays. These are under evaluation; we name them so partners know what we are trying, not what we have already shipped.
1. Reviewable truth
Problem. Undifferentiated model memory is not a substitute for material that has passed human review.
Hypothesis. High-stakes answers should anchor in curator-governed knowledge domains—versioned, attributable corpora with explicit validity—not in weights alone.
Limit. No method is infallible; auditable answers must remain inspectable under your policies.
2. Instruction following without unbounded drift
Problem. Rigid rule systems break at the edges; end-to-end neural systems drift from procedure.
Hypothesis. Dialogue and clarification can stay flexible while commitments about what may be retrieved or asserted remain in governed artifacts reviewers can change without retraining massive stacks.
3. Energy and context scaling
Problem. Long context and dense models are expensive in memory and joules.
Hypothesis. Alternative architectures and selective depth are worth measuring alongside moving factual work into curated stores when the answer already exists. We report results under stated workloads—not as universal speedup marketing.
4. Isolation and responsiveness
Problem. Shared training and shared indexes create subtle cross-tenant risk; long contexts amplify serving cost.
Hypothesis. Policy partitions at the knowledge and rules layer are the right place to enforce who may see what. Fast wrong answers are as serious as slow right ones.