Methodology
How a Finexia run is set up so that its result can be argued with: scope, as-of date, admissible sources, missing data, provenance, checks, and what counts as evidence.
This is the protocol the product enforces in code. Where this page and the code disagree, the code is right and this page is a bug.
#The eleven steps of a run
- The question is written down as asked, not as reformulated.
- The scope is fixed: which entity, which segment, which currency.
- The as-of date is fixed before any data is fetched. Everything downstream is evaluated at that date.
- Admissible sources are declared. A role that may not reach the open web cannot widen this list at run time.
- Missing data is recorded as missing. It is never filled by interpolation, and never silently dropped.
- Provenance is attached to each figure: the source, the period it covers, the metric, the unit, and the date we read it.
- Calculations are expressed as formulas in the deliverable, not as values copied into a cell.
- Checks compare what was expected against what was obtained, with a tolerance, and report pass or gap.
- If the data cannot answer at the as-of date, the card returns BLOCKED with the reason named.
- The run states which of its parts are tested, which are packaged, and which were observed.
- The run is replayable from the files it left behind.
#The only date we are sure of
A source carries a publication date that we did not observe and cannot verify. It also carries the date we read it, which we did observe. The core stores the second one under the name that says what it is, and the tool that reports a source to the journal redacts any secret embedded in the URL before writing it down 2.
#Tested, packaged, observed
These three words are not synonyms, and most disagreements about what a system can do come from treating them as one.
| Word | What it means | What it does not mean |
|---|---|---|
| Tested | Code exists and a test exercises it, including its failure path. | That anyone has seen it work in the product. |
| Packaged | It is inside a build that was produced and signed. | That it was opened, clicked, or observed on screen. |
| Observed | Someone ran the application and watched it happen. | That it is covered by a test, or that it will happen again. |
EvidenceApplied to the current version: the workbook engine is tested; the interface fixes from 0.0.20 are packaged and only one of the four was observed; the preflight refusal of a missing deck engine was observed in a live run 2.
#What BLOCKED means
BLOCKED is not a failure to be retried. It is a result: the question, at that date, with those admissible sources, cannot be answered honestly. A card that returns BLOCKED names the reason — a missing engine, an unavailable window, a source that does not cover the period.
The cost of not having this state is measurable. Version 0.0.20 spent 9.4 million tokens on a request whose format had no engine behind it. Version 0.0.22 refuses that request before starting a single process, and the refusal was observed in a live run 2.
#Reproducing a run
- A run is replayed from its own recorded files, not from a description of it.
- The version and the commit are named. A capture from an older version is shown as that version.
- The trace is the event stream, not a summary written afterwards.
- A deliverable is verified by reopening it with a different library than the one that wrote it.
LimitThe public evidence bank on this site was sealed against Finexia OS 0.0.15 on 2026-09-08. The product is now at 0.0.23. Every screenshot carries its own version; none of them is presented as the current build.
#Sources
- Finexia OS — product state at 0.0.23FinexiaInternal document, not published
- Finexia OS — handover report 0.0.22FinexiaInternal document, not published
- Finexia OS — handover report 0.0.19FinexiaInternal document, not published
- The canonical role registry, mirrored on this siteFinexiaInternal document, not published
Read next
- Finance intelligence researchWhat makes a financial document hard for a language model, and what a desk has to add before an answer is worth acting on.
- Evaluations and benchmarksFour published benchmarks for financial reasoning, documented from their primary sources — and a plain statement that Finexia has not run any of them.
- Case studiesRuns that actually happened, with what was asked, which roles were convened, what was produced, and where each one stopped.
More in Research