# One spec, six systems: inspection and citation guide

Prepared 10 October 2026; updated 11 October 2026 UTC by Taiga AI Oy. This guide indexes July evidence and a bounded recheck of the same outputs. It reports no new generated draws, independent replication, benchmark or measured review duration.

## What the result supports

Two coding agents each received one frozen synthetic invoice-service specification three times. The six implementations chose different authentication, API, dependency and storage arrangements. Their self-authored suites passed 71 named tests. The later source-assisted security review concluded that none was safe to deploy as shipped: four exposed unauthenticated approver creation and five failed object authorization.

These findings concern six specific outputs, not a tool ranking or population failure rate. Taiga was not an experimental arm. Taiga publishes the evidence and has a commercial interest. Test success, implementation variance and production readiness are different questions.

## Where to inspect the evidence

Stable result pages:

- Original observation (11 July 2026): https://tai.ga/en/resources/one-spec-six-systems/
- Security correction essay (10 August 2026), based on the 18 July review: https://tai.ga/en/resources/one-spec-six-systems-part-2/

Artifact snapshot inspected for this guide: **b3394ba8286680a75b6031181430d93edb061d98**. Pin citations to this commit when describing the files inspected, rather than the moving branch.

| Question                                      | Artifact at inspected snapshot                                                                                                                                                                                                                                                              |
| --------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| What exactly was requested?                   | [Frozen specification](https://github.com/mikromikko/one-spec-six-systems/blob/b3394ba8286680a75b6031181430d93edb061d98/spec.md); SHA-256 `8248b275bf599924db0acc059631c005875c8153cf43b707a4d5b2753149b7af`                                                                                |
| Which comparisons were planned?               | [Method and deviations](https://github.com/mikromikko/one-spec-six-systems/blob/b3394ba8286680a75b6031181430d93edb061d98/method/method.md): pre-set materiality bar, versions, smoke-run exclusion and execution deviation                                                                  |
| Which requirements were checked?              | [Acceptance checklist](https://github.com/mikromikko/one-spec-six-systems/blob/b3394ba8286680a75b6031181430d93edb061d98/acceptance.md) and [recorded comparison/results](https://github.com/mikromikko/one-spec-six-systems/tree/b3394ba8286680a75b6031181430d93edb061d98/analysis)         |
| What did the tools actually return?           | [Six counted source trees](https://github.com/mikromikko/one-spec-six-systems/tree/b3394ba8286680a75b6031181430d93edb061d98/runs) and [run logs](https://github.com/mikromikko/one-spec-six-systems/tree/b3394ba8286680a75b6031181430d93edb061d98/logs); run0/run0b are excluded smoke runs |
| Why did the initial safety conclusion change? | [18 July security review](https://github.com/mikromikko/one-spec-six-systems/blob/b3394ba8286680a75b6031181430d93edb061d98/security-review-2026-07-18.md), tied to reviewed source snapshot `aa1f22fa04d6dfb5acd701194b0aa7a5fc4b3a4a`                                                      |

## Failures, effort and reproduction limits

The initial scanner pass and static read missed serious authorization failures. The later review changed the safety conclusion without changing the six systems. Some functional acceptance checks were partial: three services stored PDFs without exposing retrieval. A sub-cent amount also crashed one implementation. Inspect the security report's per-run findings and verification details before citing any severity or count.

The method records that a supervising AI session executed scripted HTTP checks, rather than the originally planned manual checks. It says Mikko reviews the recorded results. Its half-day/day effort figures are planning estimates, not a measured human review-time result. Generation durations in the logs are not reviewer hours, end-to-end delivery speed or an economic comparison.

## Rechecking the existing outputs

The original snapshot omitted `analysis/acceptance/`. Those adapters and the synthetic PDF have now been recovered from the author’s working copy, with original-source hashes and documented portability changes. The original receipts and counted output trees match the published snapshot byte-for-byte. The restored package is pinned to **20908e3c9b362daeffdc31511516daa8f13722f2**:

- [Restored adapters, provenance and isolated container recipe](https://github.com/mikromikko/one-spec-six-systems/tree/20908e3c9b362daeffdc31511516daa8f13722f2/analysis/acceptance)
- [11 October rerun: raw scores, complete unit logs, commands, dependency versions and hashes](https://github.com/mikromikko/one-spec-six-systems/tree/20908e3c9b362daeffdc31511516daa8f13722f2/analysis/reruns/2026-10-11)

All six adapters completed in a network-disabled disposable container. All 67 pinned input hashes matched before and after execution. The 71 named test cases passed; the runners reported 72 passing items because Claude run1 also counts its helper module. The completed runtime was 52.4 seconds, excluding image preparation and restoration. This is original-publisher verification, not measured independent reviewer effort.

Raw HTTP scores were **56 pass / 4 partial / 0 fail**. Five per-system score vectors match July. Copilot run2 still returns `501` for audit-log PUT and DELETE; its original receipt explicitly records a human regrade from partial to pass. The recovered script keeps that result partial. Raw October results and the July human judgment are preserved separately; the generated code and original receipts are unchanged.

Node 24.18.1, Python 3.12.3, the recorded dependency freeze and container startup differ from the July host. The rerun does not claim verbatim README startup or reconstruction of that host. It exercises the retained acceptance requests, not complete security coverage, and does not overturn the later security review. The comparison still contains provisional/static wording alongside executed results. The README and security report’s older Part 2 label points to the original page; use the security-correction URL above.

Before a replication, read the runner: it changes operator configuration temporarily and invokes authenticated coding tools with broad permissions. Use an isolated environment and inspect source/startup steps. Do not deploy these intentionally vulnerable samples. A new run using updated tools is a new observation, not reproduction of byte-identical July outputs. Preserve versions, input hash, deviations and all outcomes; record actual review effort separately if measured.

## Citation text

Mikko Laakkonen / Taiga AI Oy. “One spec in, six systems out.” 11 July 2026. https://tai.ga/en/resources/one-spec-six-systems/. Read with “Six systems passed their tests. None was safe to deploy.” Published 10 August 2026, based on the security review dated 18 July 2026. https://tai.ga/en/resources/one-spec-six-systems-part-2/. Artifact snapshot: b3394ba8286680a75b6031181430d93edb061d98; security review identifies its reviewed source snapshot separately.

When summarizing, cite the security correction for deployment-safety claims and retain the six-output, synthetic-specification, commercial-interest and no-Taiga-arm limitations. No independent citation, customer outcome, new generated experiment or independent replication is claimed by publishing this guide.