Follow the claim to its source.
Start with the material you can read now. Research, reference implementations and fictional walkthroughs have different evidential value; each is labeled below.
Watch the maintenance walkthroughWhat the published research found.
Taiga’s July 2026 experiment gave one frozen invoice-service specification to Claude Code and GitHub Copilot CLI, three times each. All 71 named self-authored tests passed, but the later source-assisted security review found none safe to deploy as shipped. Four implementations allowed unauthenticated approver creation; five had object-level authorization failures.
On 11 October, a bounded rerun rechecked the six existing outputs without new model calls or generated-code changes. All 71 named tests passed; the runners counted 72 passing items including one helper module. The 60 automated acceptance checks recorded 56 pass and 4 partial. The July manual regrade is preserved separately.
Limits of this evidenceOne specification, two tool/model combinations and three draws each, published by Taiga. There is no independent replication, no Taiga factory comparison arm and no customer outcome. These results do not establish a general failure rate or a vendor ranking.
One specification. Six implementations.
The published experiment compares six systems produced from one frozen specification. Inspect the method, acceptance checklist, runner and per-run results.
Limits of this evidenceA small experiment by Taiga, which sells a related product. It is not a customer outcome, a general failure rate or a benchmark of the Taiga platform.
Open the public artifact bundleHÄLYTIN
Inspect the public civil-preparedness reference project and the scope described on its page.
Limits of this evidenceA reference implementation is not a government procurement, endorsement or proof of a customer deployment.
Explore HÄLYTINA requirement, a failed check, a reviewed change.
The North Pole walkthroughs expose a permission rule and a capacity rule, the corresponding failed-and-corrected checks, and the handoff to a configured deployment pipeline.
Limits of this evidenceThese are authored explanations. Their project records, test results and browser applications are illustrative, not exported production execution evidence.
Inspect the existing-software flowThe complete claims register
The receipts register keeps the source and date behind the site’s numerical and operational claims. Confidential detail remains requestable; an inquiry is never required to read these public materials.
The complete claims register