Applied research
Research here is a discipline across the three products and a publishing programme — not a staffed laboratory, a foundation model, or a portfolio of papers. We publish engineering notes when the work is engineering, and label an experiment as an experiment.
Publications
Nothing is published yet
Rather than fill this index with titles that have no method behind them, it stays empty until the first note clears review. Below are the questions we are working on, what a published note has to carry, and the rules any benchmark number must meet before it appears anywhere on this site.
We would rather start with two strong technical notes and three honest product evidence assets than a long list of unsupported research titles.
What we are investigating
A theme is a question, not a finding. Each one lists the evidence that would have to be published for an answer to count.
Bahulam
Reliable agent execution
Which orchestration, context, and recovery choices actually improve the rate of accepted outcomes?
Evidence we would publish
- Versioned evaluation tasks and the selection rule used
- A baseline configuration to compare against
- Run manifests — model, provider, harness commit, allowed tools, retry budget
- Success and failure analysis, including the failures
- Cost and latency, with the methodology for both
CoDash
Semantic modernisation
Where do translated BI calculations preserve meaning, and where do they diverge?
Evidence we would publish
- Permitted fixtures with their source platform and version
- Source and target outputs, side by side
- Parity checks and the tolerance applied
- Constructs that are not supported, named individually
- Reviewer effort to repair the exceptions
OneAdmin
Authority and operational evidence
Can an exact proposal, the policy that governs it, and the resulting state change be validated before the action is taken?
Evidence we would publish
- An explicit state model for proposal, authority, policy, and receipt
- Negative cases — unauthorised attempts and policy denials
- The scope that is implemented
- The backend gaps that remain
Bahulam
Interactive work artifacts
How do persistent workspaces and plugin views change review and handoff between people?
Evidence we would publish
- A working implementation, not a mockup
- A usability protocol stated before the observations
- Observations, including the ones that contradicted the hypothesis
- Limitations and reproducible examples
What every published note carries
For a confidential study we publish an approved summary and say plainly what cannot be reproduced publicly.
- Title, named authors, date, and publication type
- The research question and why it matters to enterprise work
- Method, baseline, and the dataset or fixture scope
- Implementation and model versions
- Results, the failures, and the limits
- Reproduction instructions where the work can be reproduced publicly
- The product behaviour the work relates to
Before a benchmark number appears
An attractive number without this record does not reach the homepage, a product page, or the co-pilot.
- Dataset version and split, sample count, and how tasks were selected
- Model, provider, and version; agent or harness commit; configuration
- Allowed tools, retry budget, and environment
- Timestamps, failures, and any human intervention
- Cost methodology
- Whether this was a completed official evaluation or an internal sample
- Denominator and uncertainty, comparing like configurations only
What is being written
- 1
A reproducible agent-task evaluation
Accepted and failed cases both reported, with the run manifest.
- 2
A calculation-conversion walkthrough
One exception, its repair, and the validation that confirmed the repair.
- 3
The administration authority model
Where the website simulation ends and production enforcement would begin.
- 4
From customer workflow to scoped internal product
How a workflow becomes an evaluation plan and an acceptance decision.