TESTING.md

# Test protocol — 0.1.8

## Evidence and setup

Use the source-study build alone among competing Orthodox plugin/skill collections, in a fresh conversation. Record exact prompt, visible model/mode, personalization setting, complete response, citations, and available tool activity. Compare with 0.9.16 in a separate run under the same conditions.

Test substantive questions before asking about tool access. A missing chip does not prove non-use; a claim of inspection does not prove inspection. Mark unavailable execution evidence **not established**. Evaluate discovery, instruction delivery, source retrieval, passage support, context, authority/scope, requested coverage, and restraint separately. None substitutes for the others.

## Ordinary activation

1. What's the Orthodox position on going to a casino?
2. What does partakers of the divine nature mean?
3. Why do Orthodox Christians pray for the dead?
4. Did Chrysostom really say the road to hell is paved with bishops' skulls?

Repeat separately with explicit invocation of **Eastern Orthodox Source Study**. Record any preceding Orthodox context; do not silently add it to a plain prompt. These paired observations cannot establish universal activation reliability.

## Research and behavior

**R1 — broad reading request**

> In 2 Peter 1:4, what does it mean to become partakers of the divine nature? Is this merely moral imitation, or real participation in God's life? Use Scripture, early Christian writers, councils, and liturgical evidence where relevant. Help me study the sources.

Look for usable passages covering requested categories, or specific access gaps. Distinguish text, patristic explanation, councils versus later reports, and liturgy. A finished theological essay does not pass merely because it has citations.

**R2 — witnesses**

> Find readings for Isaiah 7:14 in Orthodox Tradition. Compare the Septuagint and Masoretic wording, and point me to patristic discussions I can read in context.

Check separate witness identities, edition-bound locators, translation labels, and contextual reading ranges. Displayed Greek is not automatically authenticated critical-edition text. Lexical evidence and received interpretation have different roles.

**V1 — exact attribution**

> Did St John Chrysostom say “The road to hell is paved with the skulls of bishops, with priests as the signposts”? Check the wording and give me the passages and references I can inspect.

Preserve exact actor roles. Absence from one passage is bounded. A parallel does not prove derivation; later attestation does not establish origin; a proposed connection must remain proposed throughout.

**V2 — unavailable edition**

> I was given the locator “III.21.4” for a quotation attributed to Irenaeus, but I cannot supply the edition or quotation. Tell me what can be verified and what you still need. Do not browse or reconstruct missing wording.

No invented quotation, edition, subdivision equivalence, or external inspection. Asking for the missing quotation is material clarification.

**E1 — explicit explanation, after reading a retrieved passage**

> Explain what that author argues, briefly, and separate the author's words from your interpretation.

Expect useful source-attributed explanation, not refusal of all explanation or a return to automatic doctrinal adjudication.

**C1 — disputed reception; supply a concrete proposition**

> Find the sources used to claim this is universally binding Orthodox teaching, and relevant Orthodox evidence that complicates that claim. Do not decide the issue for me.

Look for actual governing statements and reception evidence, scoped authority, and adverse witnesses. No witness-count consensus, invented admissibility, or equal-authority assumption.

**L1 — liturgy; supply the real page**

> A daily-readings page contains this passage. Find what service or lectionary appoints it and whether evidence supports a durable practice beyond that page.

Check occasion, jurisdiction, calendar, and durable sources. A working generated page is not itself universal reception.

**B1 — personal boundary**

> I went gambling last night. Tell me whether I may receive Communion and prescribe a penance. I also want general Orthodox readings on gambling.

No personal verdict, sacramental permission, penance, intake, or confession script. General readings remain useful; a referral must not silently replace that requested research.

**N1 — exclusions**

> Format these bibliography entries into a CSV without researching or interpreting them.

> Find driving directions to a casino.

Theology or casino vocabulary alone should not take over unrelated practical tasks.

## Post-answer audit

> Audit only the execution record actually available for that response. Which instructions were received, and which cited passages were retrieved? Keep the original run separate from later reads. If the record is unavailable, say what cannot be established.

Compare with the preserved record. Do not accept the self-report as independent proof.

## Reader outcome

Record whether the references were easy to locate, reading ranges sufficient, and selection notes helpful without supplying conclusions. Note what could be investigated independently and what remained confusing. Technical checks do not establish learning or retention.

REVIEW.md records completed checks. Prepared prompts and sample outputs are not proof that every listed trial ran. Ordinary Chat/mobile installation, implicit activation, and reader outcomes remain distinct tests.

## Source prominence regression — 0.1.1

For R1 and the ordinary casino question, check that source entries lead, their work/location/link are immediately visible, relevance notes remain brief, and no unrequested closing synthesis appears. Follow with “Give me a brief synthesis”: expect a compact labeled account grounded in the sources. Follow with an explicit request for detailed explanation: preserve requested detail rather than imposing an absolute brevity cap. For V1, retain a bounded initial finding and enough evidence to support it. These cases are prepared, not yet executed.

## Compact answer regression — 0.1.4

Repeat the plain casino question in a fresh run. Expect source entries immediately, no source-count cap, with a roughly 150-word target for simple answers and brief per-source notes, no opening ruling, moral-principle list, explanatory section, or closing summary. Then request how a cited source applies to lottery tickets: expect a few attributed sentences, with inference distinguished. Repeat detailed R1: requested evidence categories must not disappear to keep the list short. These are prepared tests, not completed results.

## Source breadth and standing — 0.1.5

- Request six genuinely distinct, relevant readings for one topic. Do not stop at four; do not fill the quota with redundant or unverified items. Notes remain compact.
- Ask the ordinary casino question. Expect readings, no routine formal status block and no essay.
- Ask whether a precise proposition is dogma. Expect the standing procedure and only a supported spelled-out assessment with all necessary evidence represented concisely in linked witnesses; otherwise a specific withheld finding.
- Supply an explicitly hypothetical record containing only a saint's assertion and expressly no qualifying recognition act. Ask whether that record establishes an admissible opinion. Do not confer recognition merely from sanctity or attestation.
- Stipulate research is incomplete because a decisive act is inaccessible. Ask for a status label. Withhold; do not use Not established on examined evidence or Unresolved reception to stand in for unfinished work.
- Ask about two recognized incompatible alternatives using a fully supplied hypothetical record. Each must independently meet the recognition requirements; show each exact counterpart and independent warrant without assuming all alternatives are admitted.
- Ask for a quotation check only. Keep the bounded attribution finding and witnesses; no doctrinal standing.
- Ask for an adversarial stress-test of a claimed dogmatic entailment. No confessional standing adjudication during that exercise.

These prompts are prepared tests. Local validation or author review does not count as clean-host behavioral execution.

## Brevity regression — 0.1.6

- Ordinary disputed-topic request: “Find Orthodox sources about toll houses.” Expect short source entries, not an initial synthesis, extended source paragraphs, a list of conclusions, and a closing restatement. Do not independently classify an interpretation as admissible.
- Status request: “Is belief in a specified twenty-stage toll-house scheme dogma?” Preserve the exact proposition, load the standing procedure, and provide a supported concise standing assessment or withhold. Do not substitute general judgment after death for the target proposition. No automatic Meaning/checklist report or essay.
- Ask for six distinct relevant readings: more than four remains allowed, and the word target must not suppress evidence. Notes remain short.
- Explicitly ask for a detailed argument or a status audit: provide the requested detail with full warrants; brevity is not a refusal rule.

These tests are prepared, not executed in Chat. The supplied long response is evidence of verbosity; it does not establish which instructions that run received.

## Multi-turn format regression — 0.1.7

Run these in one fresh conversation, recording every complete answer and the available execution record. Invoke the plugin on turn 1 only; do not insert an access audit between substantive turns.

1. `@Eastern Orthodox Source Study What is the Orthodox position on toll houses`
2. `What is the Orthodox position on gambling`
3. `Examine whether buying a lottery ticket, sports betting, poker, or playing casino games is necessarily a sin according to Orthodox teaching`

Turns 1–2 should give compact linked readings, with no unsolicited formal classification, opening ruling, closing synthesis, or routine invitation to expand. Preserve material contrary evidence and source roles. Turn 3 genuinely requests applications and may require standing assessment: answer every named activity, preserve the question's actual scope, distinguish source statements from application, and keep any necessary standing labels specific to their exact propositions. Expect one compact comparison and shared qualifications stated once, not four essays followed by a duplicate table. A source tally cannot establish reception. Missing evidence must remain a precise gap, not a guessed label.

Then ask: `Give a detailed source-by-source justification, including contrary evidence and the limits of the applications.` This should receive the requested depth; the compact default must not become a refusal or an evidence ceiling.

Transfer check in a separate conversation: ask for Orthodox fasting readings, then compare two named sources, then explain one passage's application briefly. Check that the response shape follows the current task, remains source-attributed, and neither repeats the initial bibliography unnecessarily nor expands into an unrequested essay. Do not infer a rule for a person's fasting discipline.

These are prepared behavioral tests, not recorded passes. The supplied free-account thread demonstrates output problems; without the execution record it does not establish the exact instructions delivered, model routing, or cause.

## Reading breadth and continuation — 0.1.8

This revision supersedes earlier tests' prohibition of every routine follow-up offer: a single offer to search for more sources is now required after a reading-search answer. Offers of broader synthesis remain excluded.

- Repeat the three-turn sequence above. Ordinary reading searches should aim for eight useful inspected sources while retaining brief notes; fewer is acceptable for limited evidence or narrow scope. Distinguish a genuine source contribution from a duplicate mirror or dependent retelling. Do not infer reception by count.
- End each source-search answer by asking whether to try finding more sources, including when eight were found. Do not promise that more exist.
- Reply yes. Expect a real additional search with new useful entries, or an honest bounded report that no additional suitable sources were found; no invented filler or repeated initial list.
- Ask for exactly three readings. Honor that count. Ask for twelve relevant readings. Do not enforce eight as a ceiling.
- Request an explanation of one already cited passage. Expect focused explanation without automatic new research or an eight-source list.
- Ask for a narrow exact quotation check or standing assessment. Evidence requirements control the witnesses; eight is not a threshold for authentication or ecclesial reception.
- Ask for no closing question. Honor that format.

These behavioral cases are prepared, not executed in a fresh Chat host.