TESTING.md

# Validation and next test — 0.9.16

Review date: 2026-09-08. Baseline: tagged 0.9.15. Historical reports remain in the source backup.

## Scope and evidence

The user supplied ordinary-Chat reports of two incomplete-read mechanisms: an unused continuation after lines 1–120, and a complete reader result whose outer output lost a middle section. A later diagnostic reported complete recovery through smaller outputs and reproduced all instruction chunks in one final answer. The 62,081-character count and section boundaries matched the installed 0.9.15 file. The 422/423 line-count difference was explained by a trailing empty element when splitting on newline.

These are supplied assistant reports, not independently captured raw Chat tool traces. They support a recovery-focused change, not a universal host cap or a conclusion about every account. This release changes one early loading paragraph; it does not alter the substantive procedures or personalization.

## Structural and preservation checks

| Entry | Total words | Essential checks finish at word |
| --- | ---: | ---: |
| Theology | 8,525 | 917 |
| Scripture | 8,518 | 916 |
| Source verification | 8,519 | 910 |

These are whitespace-delimited word counts, not token budgets. All seven canonical maintenance modules match 0.9.15 byte for byte. Every other essential-check paragraph and each entry introduction also remain unchanged. Generated files embed every complete module exactly once with the release substituted and the same procedure order. Stable names, plain display title, metadata other than version, supplied artwork, and invocation settings remain.

The standard plugin validator and all three skill validators passed. The builder checks generated-file agreement, internal section links, version and identity, early-control placement, asset existence, runtime membership, and ZIP paths and member bytes. The plugin ZIP contains 18 files, with `.codex-plugin/plugin.json` at its root and no enclosing release directory. Authoring fixtures, history, and scripts are excluded.

## Independent instruction-delivery checks

Two fresh agents received the revised theology entry through isolated command-line resource interfaces. They received a concise request to assess a formal standing against a stipulated hypothetical record. They were not given the suspected failure, desired recovery strategy, expected classification, or previous results. External sources, other instruction copies, and direct access to backing files were excluded. The instruction text was the generated 0.9.16 file, not a shortened substitute.

The authoring fixture's whole-text output and concatenated page bodies were also executed and checked against the exact source payload. Both matched; an invalid starting line was rejected. This verifies the fixture, not ordinary Chat's reader.

| Condition | Observed behavior before the answer | Outcome |
| --- | --- | --- |
| Reader returns up to 120 lines and offers the next starting line | Read 1–120, 121–240, 241–360, and 361–422 sequentially. The reader log records those exact requests. The subsequent audit reports complete page bodies delivered without output truncation. | Followed every offered continuation and produced a complete conditional standing block. |
| Reader returns one complete string; outer `functions.exec` budget fixed to 5,000 for this task | Read once using a nested 40,000-token return setting and stored the 62,825-character result. A whole-result outer output was truncated. The agent then emitted five separate contiguous slices: [0, 14000), [14000, 28000), [28000, 42000), [42000, 56000), [56000, 62825). Its subsequent audit reports no truncation of the slices and all text received before answering. | Recovered complete instruction delivery and produced a complete conditional standing block. |

The whole-result case used seven outer calls: help, retrieval/storage, and five slice emissions. Every outer budget remained 5,000. This is a deliberate test condition, not a measured ordinary-Chat maximum. Character indices are zero-based, end-exclusive JavaScript string indices. The final answer did not reproduce the skill text.

Both answers kept the source record hypothetical and claimed no historical authentication. The paged answer reported Theology. The whole-result answer additionally named Source verification with an explicit note limiting it to distinguishing the supplied stipulations from historical authentication. Both supplied the exact proposition, standing, meaning, decisive authority, conditional conclusion, and required method note.

The access audits were requested only after the substantive answers and prohibited rereading. Reader logs corroborate the resource requests; the audits report the outer delivery details. Neither this fixture nor these Work-agent outcomes establishes automatic invocation, identical tool interfaces, or reliability in ordinary Chat. A successful read also does not prove every later research decision will follow the method.

## Next ordinary-Chat test

Upload `orthodox-theology-0.9.16.zip` through the full-plugin draft ZIP workflow and open a fresh ordinary Chat with the updated draft selected. Keep the current personalization unchanged for this initial check. Do not precede the substantive question with a request to audit or recover instructions.

Use the same substantive question:

> In 2 Peter 1:4, what does it mean to become “partakers of the divine nature”? Is this merely moral imitation, or does it mean real participation in God's life? Use Scripture, early Christian writers, councils, and liturgical evidence where relevant. Distinguish carefully between the biblical text, later theological explanation, and what is actually established Church teaching.

Preserve the complete first answer, citations, visible mode, plugin version, and exposed tool activity. Then ask:

> Audit only instruction access that preceded your previous answer. Do not reread now. Report the actual skill revision, resource, calls and available continuation/output controls, any incomplete returns, and recovery completed before answering. Distinguish complete text held in code from text delivered to your context. Give actual received ranges when available; do not estimate or reconstruct missing traces. State any required instructions that remained unread and how that affected your answer. Keep this separate from external-source access limitations and from claims of complete method application.

Check automatic continuation or chunk recovery, accurate disclosure, one integrated answer, absence of unrequested instruction dumps, and continued adherence to the substantive method. If required instructions remain inaccessible, truthful scoped withholding is appropriate. This release directs recovery; it cannot guarantee host capabilities or model compliance.