Production-scaleempiricalresearch.

Technical reports based on real data from Vokrix. Single-organization studies — we report what the data shows.

6 papers published|Weekly publication schedule|Open access

Telemetry Asymmetry·Oct 25, 2026·31 min read

Telemetry Density Asymmetry Between Verification and Market Sensing in Autonomous Product Pipelines: A Production Observational Study

This report tests whether a production autonomous software development system records internal verification telemetry at a materially higher density than external market-sensing telemetry. The pre-specified hypothesis was that verification records would outnumber market-sensing records by at least 1.5:1 within the observation corpus. Anonymized production records released by a single study organization were examined: 1,000 recorded verification (QA) check results and 500 recorded external intelligence signals, accompanied by contextual records covering 500 build attempts distributed across 46 builds, including 27 QA failures, 12 escalations, 6 pending outcomes, and 8 verified builds. The observed record-count ratio was 2.0 verification records per market-sensing record, exceeding the 1.5 threshold, and the hypothesis was therefore supported for this corpus. Contextual build telemetry showed strong right skew in iteration counts (mean 71.3, median 29, range 4–300), and 39 of 46 builds required at least one fix attempt. The design is observational, single-organization, and without a control or comparison condition; the two record streams are not measured in identical units. The finding is therefore descriptive of this corpus only and is not a benchmark. AI assistance in manuscript preparation is disclosed. ---

Read paper →

Repair Observability·Oct 18, 2026·26 min read

Repair Provenance Sparsity in Autonomous Software Pipelines: A Production Observational Study

Autonomous software development pipelines generate verification artifacts at high volume, yet the extent to which they preserve explicit records of repair activity is largely unknown. This study tests the hypothesis that, in such pipelines, recorded repair events occur less than once per completed build while QA results occur more than fifty times per completed build. A black-box observational design was used, drawing on anonymized production records from an autonomous software development system operated by a single organization. The study tabulated build iteration profiles across 500 build attempts, QA outcomes for 46 completed builds, per-check failure distributions from 1,000 executed QA checks, and the sizes of the repair-history and QA-result record stores. Repair activity was pervasive in outcome terms: 39 builds (84%) required at least one fix attempt, with a mean of 2.3 attempts where required and a maximum of 7 on a single build, while only 8 builds (17%) reached verified status. Verification volume was high, with 1,000 QA checks yielding a 94% pass rate and 17 unique products recording failures. Recorded repair provenance was sparse relative to both build count and verification volume. The hypothesis was supported. Generalization is limited by the single-organization, uncontrolled, observational design. ---

Read paper →

Market Signal Conversion·Oct 17, 2026·25 min read

Signal-to-Opportunity Yield in Autonomous Market Discovery: A Production Observational Study

This report presents an observational study of signal-to-opportunity yield in the upstream stage of a production autonomous market-discovery pipeline operated by a single organization. The hypothesis under test was that qualifying intelligence signals outnumber generated company opportunities by more than 50% within such a pipeline. Data were drawn from anonymized production records comprising an aggregate tally of 1,287 intelligence signals and 832 generated company opportunities, a record-level extract of 500 scored signals, 500 build attempts, and quality-assurance (QA) telemetry covering 46 builds in the current system population. The observed signal-to-opportunity ratio was approximately 1.55, corresponding to a signal surplus of approximately 54.7% over opportunity volume; the hypothesis was therefore supported. The most frequent signal category was *AI Tool* (255 of 500 scored records), and the mean signal score was 7.6 on a scale that was not documented in the extract. Downstream build and QA telemetry is reported as contextual description only. The study is observational, single-organization, and uncontrolled, and the manuscript was prepared with AI assistance; the findings describe one pipeline and should not be generalized.

Read paper →

Autonomous Product Pipeline·Oct 16, 2026·22 min read

Opportunity-to-Build Conversion Efficiency in Autonomous Product Pipelines: A Production Observational Study

This study tests the hypothesis that fewer than 10% of generated business opportunities in a production autonomous software development pipeline result in completed builds. A single-organization observational design was used, analysing an anonymized production log extract comprising 824 generated opportunities, 46 completed builds, 500 build attempts, 500 collected intelligence signals, and 1,000 quality-assurance checks. No intervention was applied and no control condition exists. Observed opportunity-to-build conversion was 5.6% (46 of 824), consistent with the stated hypothesis. Build effort was strongly right-skewed: 500 attempts averaged 71.3 iterations (median 29; range 4–300), with 314 attempts (62%) below 50 iterations and 62 attempts (12%) above 200. Of the 46 builds, 39 (84%) required at least one fix attempt (mean 2.3 where required; maximum 7 on a single build), and 8 builds were verified (17%). Check-level quality assurance recorded a 94% pass rate across 1,000 checks. The conversion estimate rests on 46 events in one organization, cannot establish causation, and should be read as descriptive rather than generalizable. **Keywords:** opportunity conversion, autonomous pipelines, funnel efficiency, quality assurance, production observation, black-box measurement ---

Read paper →

Autonomous QA·Oct 4, 2026·28 min read

Quality Signal Density in Autonomous Verification Pipelines: A Production Observational Study of QA-to-Build Ratios

This study examines whether verification density at the quality assurance (QA) stage substitutes for downstream corrective work in an autonomous software build pipeline. The hypothesis under test states that pipelines exhibiting more than 50 QA events per completed build also exhibit proportionally fewer fix-history entries per build. Anonymized production records from a single autonomous software development system operated by one organization were analyzed: 500 build attempts with iteration telemetry, 1,000 executed QA checks, QA disposition records for 46 builds, and fix-history records for those same builds. The aggregate QA-to-build ratio derived from the reported totals is approximately 21.7 QA checks per build, which falls below the threshold specified by the hypothesis; the hypothesis was therefore not tested in the density regime it describes. Within the observed low-density regime, 39 of 46 builds (84%) required at least one fix attempt, with a mean of 2.3 fix attempts when required and a maximum of 7 on a single build. The per-check pass rate was 94% and the per-check fail rate was 5%. The study is observational, single-organization, and uncontrolled; it reports descriptive associations only and makes no causal claim. ---

Read paper →

Build Iteration·Sep 28, 2026·25 min read

Iteration Concentration in Autonomous Software Generation: A Production Observational Study

This study examines whether iteration effort in autonomous software generation is distributed uniformly across builds or concentrated in a minority of them. The hypothesis under test was that the maximum single-build iteration count would substantially exceed the mean, implying disproportionate consumption of iteration budget by a small subset of builds. Anonymized production telemetry from a single organization operating an autonomous software development system was analysed. The extract comprised 500 build attempts and 1000 QA checks, alongside build-level records for 41 builds in the system. Iteration counts per build attempt ranged from 4 to 300, with a mean of 68.3 and a median of 29. The maximum was approximately 4.4 times the mean and approximately 2.4 times the median. Of the build attempts, 65% completed under 50 iterations and 11% exceeded 200. The distribution was strongly right-skewed, consistent with the stated hypothesis. At build level, 34 of 41 builds (82%) required at least one fix attempt, with a mean of 2.4 fix attempts when needed and a maximum of 7; 8 builds (19%) were verified. The design is observational, single-organization, and uncontrolled. No causal claims are supported. **Keywords:** autonomous software generation, iteration budget, agentic software engineering, production telemetry, effort concentration ---

Read paper →