KAPI coding cost case study — offline reproduction bundle
Prepared October 29, 2026

Purpose
Recompute descriptive costs from one retained matched v2 coding cohort: 12 tasks, three repetitions, three models, 108 artifacts. This is historical accounting on October 5 UTC / October 4 Vancouver records. No current model quality, cheapest-model or retry-policy claim is established.

Run
Unzip the archive into a new folder. From that folder run:

    python3 verify_bundle.py

Python 3.9 or later and its standard library are sufficient. There are no package installations, API keys, model downloads, network calls or provider charges. The script reads only analysis.json and retained-coding-evidence.json beside it. Model prompts and answers remain inert JSON strings; the script never executes them or calls a benchmark runner.

Optional file integrity check
On macOS/Linux with shasum installed:

    shasum -a 256 -c SHA256SUMS

SHA256SUMS covers every other file in this folder. It detects byte changes against the included manifest, not an external signature or proof of provider identity.

What is checked
Complete population; unique trial and response-hash identifiers; per-attempt historical token cost using ceiling to one microdollar; total and rejected-artifact cost; input/output and recorded reasoning token totals; retained strict-pass counts; cost per pass; rejected-spend share; denominator uplift and passing-only baseline loading.

Files
verify_bundle.py — offline reader checker
retained-coding-evidence.json — selected coding outputs, usage, settings, retained grades and prompt records
analysis.json — reported descriptive totals and historical rates
source-provenance.json — original local source paths and byte hashes retained during the audit
SHA256SUMS — byte checksums of bundle files

Limits
The original complete raw provider response payloads and settled invoices are not included. Stored response hashes identify retained payloads, but this bundle cannot independently verify absent payload bytes. Grades are retained labels, not regraded results. A strict pass is not human acceptance of production code. The coarse execution_rejected label does not reveal a specific runtime cause. Independent repetitions are not iterative repair attempts. Provider-specific medium reasoning settings are not equal compute. Hardware, human repair and review, taxes and tool fees are outside these estimated provider charges. Historical prices are not current quotes.

Reproduction starts from the supplied exports and rates. Matching arithmetic is not independent validation of task representativeness, source metering or provider billing. The source-provenance paths describe the original working tree; no access to that tree is required.

Source context
https://kingy.ai/best/coding/
https://kingy.ai/best/methodology/
https://kingy.ai/tools/ai-cost-calculator/

Model outputs are retained as source material. No additional rights to third-party model output or material are asserted by this bundle.
