# nobrainer-tech-flow 2.2.0 verification record

The communication reference now chooses the simplest useful presentation and
closes non-trivial work with inspected proof and visible unfinished requirements.
Text remains the default; diagrams explain relationships; HTML needs useful
interaction. Explicit formats and artifact-only output take precedence. Video
requires a bounded purpose and available tools, with a text/storyboard fallback.

## Scope and weight

One existing reference owns these rules. The eighteen skills, bootstrap and skill
bodies are unchanged. No new renderer, provider, dependency, helper, data schema
or result store is shipped. The reference also makes two observed failure classes
explicit: machine-only output must not gain Markdown fences, and exact source
text must stay separate from process metadata. Other execution stages and their
workflow diagrams are unchanged; no diagram refresh is needed for this prose policy.

## Actual checks and limits

The [frozen inputs and raw answers](../evals/clear-results-v2.2.0.json) record
fresh local CLI processes on 3 October 2026, using an explicitly supplied
communication reference. Codex CLI 0.157.1 requested `gpt-6-astra` at `xhigh`;
its event stream does not independently identify the served model. Claude Code
2.1.286 requested `opus` at `xhigh`; returned usage identifies first-party
`claude-opus-5-5` for the text trials. The HTML artifact receipt reports both
`claude-opus-5-5` and `claude-opus-4-8`; the reason is unproven. Tools and external
actions were disabled for these probes.
No provider or client configuration was changed.

Each baseline answered three development cases; each first candidate answered
those three plus three fresh holdout cases. All per-case coarse assertions passed,
but independent review found two failures in the first candidate output: Codex
combined an exact error with the exit status; Claude fenced the outer JSON map.
Those failures remain in the record. After the narrow instruction clarification,
both clients passed a fresh single-task case: strict JSON, exact stdout, separate
integer exit status and unverified public work. This is a focused correction
check, not a re-score of the earlier run or proof of a quality gain.

Two earlier candidate artifacts were inspected and exercised by MAIN: a Python
stable-deduplication function passed six checks, including unhashable values,
order, input preservation and an iterator; a self-contained retry explanation
worked in the Codex browser at default, zero and four failures. Keyboard increment,
390-pixel layout without overflow and inert markup-like label text were checked.
These artifact checks bind the first candidate reference, before the format/literal
clarification. They do not establish autonomous client tool execution.

This is explicit-source behavior evidence, not automatic installed-skill discovery,
a full-task client benchmark, a matched competitor comparison or proof of lower
cost, faster work or better completion rates. Single runs do not estimate reliability.
The current models can still violate instructions; MAIN must inspect real outputs.
Video rendering, full client lifecycle and general compatibility remain unverified.

Required release gates are the structure/suite validators, full unit tests, staged
secret scan, independent source/output review and separate release/archive/site
readback. CI results belong to the published commit, not the initial failed local
run: that run encountered an installed `tests` namespace conflict, an intermittent
short-timeout case and a release-note link before its target was written.

## Rollback

Return to v2.1.1 through the documented installer preview/backup flow. No persisted
schema or migration is introduced. Keep the source checkout while client links
refer to it. A requested simpler format or explicit language overrides the defaults
without requiring rollback.