# Language AI Bench: interim balanced polyglot report

## Bottom line

**All 20 valid attempts in the balanced four-language brownfield cohort passed.** Correctness tied, while efficiency and workflow measurements still varied.

The full public history now contains 32 valid completions, including the earlier JavaScript/TypeScript greenfield study. The new comparative cohort uses two matched existing-project task families with five runs per language: two optimistic-concurrency attempts and three schedule-data-model attempts.

**Comparability warning:** this cohort predates the task-source integrity gate. Some TypeScript, Go, and Python starters were unusually compressed, and schedule-variants used different project topology across languages. The correctness results remain valid for the tested contracts; do not treat the recorded step, token, cost, or time differences as clean language effects.

## Balanced results

| Language | Optimistic concurrency | Schedule variants | Total | Mean cost | Mean output | Mean steps | Mean agent time |
|---|---:|---:|---:|---:|---:|---:|---:|
| JavaScript | 2/2 | 3/3 | 5/5 | $0.005514 | 2302 | 7.00 | 33.66s |
| TypeScript | 2/2 | 3/3 | 5/5 | $0.005496 | 2132 | 6.60 | 49.13s |
| Python | 2/2 | 3/3 | 5/5 | $0.005698 | 2536 | 6.80 | 31.32s |
| Go | 2/2 | 3/3 | 5/5 | $0.006450 | 2695 | 8.20 | 45.96s |

## How the agent got there

| Language | First verifier pass | Mean verifier invocations | Passing verification before submit | Malformed actions |
|---|---:|---:|---:|---:|
| JavaScript | 4/5 | 2.60 | 4/5 | 0 |
| TypeScript | 4/5 | 1.80 | 4/5 | 0 |
| Python | 5/5 | 2.60 | 5/5 | 0 |
| Go | 1/4 | 2.80 | 3/5 | 0 |

These historical telemetry values describe the recorded runs, but fixture formatting and topology confound cross-language efficiency comparisons. Agent steps count model actions; agent time excludes container setup. Explicit verifier counts come from trajectories and do not split out checks performed inside scripts/verify-local.

## Infrastructure exclusion and spend

One additional Go trial was excluded before submission because Pier's egress proxy temporarily failed DNS resolution. It consumed $0.00068230; no completed patch was graded. Total measured provider spend, including that excluded event and the earlier study, was **$0.17974737**.

## Scope

The 20-run comparison is balanced retrospectively across languages but was assembled in stages, before source-format and topology checks existed. It covers only two related backend task families, one model, one effort level, and one bash-only agent scaffold. The earlier 12 JavaScript/TypeScript greenfield runs remain published separately. A clean prospective rerun will supersede these efficiency estimates.

See `data/polyglot-results.json` for machine-readable aggregates and `DECISION_REPORT.md` for the earlier report.