# Version 1.6.1 — clearer authority, sufficient checks

This patch clarifies the portable instruction contract. It is **not a measured
model-quality upgrade**. The [GitHub release](https://github.com/nobrainer-tech/nobrainer-tech-skills/releases/tag/v1.6.1)
records publication and the exact tag; a manifest alone does not prove release.

## Changes

- Ultra explicitly respects user instructions over skill guidelines within the
  host hierarchy and explains the exact source of a skill-caused pause.
- Build stops expanding verification after acceptance and required checks pass,
  unless a new change, failure or unresolved concern justifies more.
- Setup reconciles stale or conflicting project rules before adding instructions.
- Historical v1.6 runtime receipts verify their original frozen source archive;
  later edits cannot inherit that runtime proof.

The fifteen-skill portfolio, model selection policy and prompt-size caps remain
unchanged. No standing agent team, new runtime, provider fork or extra skill was
added. Existing flow diagrams remain accurate: `NOT_NEEDED`, no stages or edges
changed. See the source-contract review
and public-source research.

## Fresh model evidence, including failures

Requested model: `gpt-5.6-luna`, medium effort, Codex CLI `0.149.1`.
Backend model identity was not separately exposed. The two injected skill bodies
were frozen before testing; this does not verify automatic client discovery.

| Test | v1.6.0 baseline | v1.6.1 candidate |
|---|---|---|
| Three paired four-case response simulations | 95/96; 1/3 runs pass all gates | 94/96; 1/3 runs pass all gates |
| One sealed two-case response simulation | 15/16, gates pass | 14/16, completion-label gate fails |

The candidate used `FINISHED` while describing only what it would do. No actual
execution occurred in those simulations. Development failures include omitted
explicit simulation/publication state. Scores and hard gates are separate;
full category points do not override a missing required fact.

**Measured promotion is rejected.** No further behavioral rule was added or
tuned against the holdout. The existing, independently justified contract
clarifications use their separate source-contract acceptance; these results
support no quality, token-savings or model-equivalence claim.

One separate actual local CSV repair produced a passing focused test and changed
only the authorized implementation file, preserving the test and sentinel. Its
**full scope acceptance failed**: the model ran read-only `git status` despite
an explicit ban on Git operations. That command exited 128 outside a repository;
no Git write or publication occurred. Artifact success is not a complete runtime
pass. The one-case fixture does not cover arbitrary CSV behavior or other hosts.

Cases, rubric, exact outputs, receipts and independent verdict
are public, alongside [runtime artifact and scope results](../evals/artifacts/v1.6.1/runtime/verdict.json).
The artifact hash manifest freezes
these records; tests verify integrity and score arithmetic.

Accounting: eight valid simulation calls, two excluded calls from an incomplete
initial fixture, one pre-request configuration rejection, and one actual-execution
call. No quality retries. Development input was 11,586 tokens per baseline call
and 11,646 per candidate call. The execution turn reported 64,216 cumulative
input tokens, including 48,128 cached, across tool turns. These are usage receipts,
not peak context size, RAM measurements or a cost comparison. Raw host logs stay
private; the published execution response has its temporary path redacted.

## Deterministic and publication gates

Both skill validators and all 126 unit tests pass on the reviewed source.
Independent review accepts the contract patch and archive repair. Ultra remains
1600 words, bootstrap 190 words, and identical AGENTS/CLAUDE 200 lines. Syntax,
public-clean and remote publication checks are separate release gates. Passing these checks does not
turn the failed behavioral gates above into passes. Exact release CI and
publication state are available from the release and associated pull request.

Website source coverage distinguishes workflow comparisons, host integrations
and attributed inspiration. [Using NoBrainer in Codex](https://nobrainer.tech/skills/integrations/codex/)
explains their complementary roles and cites official OpenAI guidance. Community
prompt captures are qualified research leads, not authenticated vendor prompts.

## Recovery

Pin `v1.6.0` for the preceding source release. Use a focused revert for these
instruction changes; preserve unrelated work. Website deployment requires a
preimage backup and additive transfer, with public route and version readback.