psyclaw 0.27.23 → 0.28.1

package/vendor/ars/academic-pipeline/examples/mid_entry_example.md

@@ -0,0 +1,414 @@
+# Example: Mid-Entry into Pipeline
+
+## Scenario Description
+
+**User type**: University professor, experienced in paper writing
+**Scenario**: Already completed an English paper draft, wants a simulated review before submission
+**Goal**: Review --> Revision --> Format conversion (LaTeX)
+**Language**: English (paper is in English; conversation with orchestrator in English)
+
+---
+
+## User Input
+
+```
+I've already written an English paper titled "Machine Learning Approaches for
+Predicting University Attrition Risk in Taiwan," approximately 7,000 words.
+I'd like a simulated review before submitting.
+```
+
+---
+
+## Pipeline Execution Log
+
+### Orchestrator: Mid-Entry Detection
+
+```
+Detection results:
+- Materials: Complete English paper draft (7,000 words)
+- Goal: Simulated review + revision + submission preparation
+- Entry point: Stage 3 (REVIEW)
+- User type: Experienced researcher
+
+Detected that you have a complete paper — you can proceed directly to Stage 3 (REVIEW).
+
+Before starting, let me confirm a few things:
+1. What's the paper structure? (IMRaD / other)
+2. What review mode would you like?
+   - full: Complete 5-person review (Journal-Fit Reviewer + R1/R2/R3 + Devil's Advocate)
+   - quick: Journal-Fit Reviewer quick assessment (15-minute version)
+   - guided: Socratic guided review (interactive dialogue)
+3. Which journal (or what tier) are you targeting?
+4. Will you need revision and format conversion after the review?
+
+If you later want to revise based on review comments, I'll automatically guide you to Stage 4.
+```
+
+**User**: IMRaD structure, I'd like a quick review first to see if there are any big-picture issues. Targeting Computers & Education (SSCI Q1). Will decide about revision after the review.
+
+### Orchestrator: Confirm Configuration
+
+```
+Got it. Pipeline configuration:
+
+  Stage 1 RESEARCH    [--] Skipped (you already have a research foundation)
+  Stage 2 WRITE       [--] Skipped (you already have a complete paper)
+  Stage 3 REVIEW:     quick -- Journal-Fit Reviewer quick assessment first, checking the big picture
+  Stage 4 REVISE:     TBD -- based on Stage 3 results
+  Stage 5 FINALIZE:   TBD -- when needed
+
+Start Stage 3 (REVIEW) quick mode?
+
+Progress: [--]Research -> [--]Writing -> [..]Review -> [ ]Revision -> [ ]Finalization
+```
+
+**User**: Start.
+
+---
+
+### Stage 3: REVIEW (quick mode)
+
+```
+Entering Stage 3 (REVIEW) -- using academic-paper-reviewer quick mode
+
+Loading academic-paper-reviewer SKILL.md...
+Receiving user-provided paper draft...
+field_analyst_agent analyzing paper domain...
+eic_agent executing quick assessment...
+```
+
+**Field Analysis (quick version):**
+```
+Paper domain: Educational Data Mining / Higher Education
+Method type: Machine Learning (Supervised Classification)
+Target journal: Computers & Education (SSCI Q1, IF: 12.0)
+
+Journal-Fit Reviewer Persona:
+  Computers & Education Associate Editor
+  Specialization: Learning Analytics, Educational AI
+```
+
+**Journal-Fit Reviewer Quick Assessment:**
+
+```
+Overall: Promising but needs work (Minor-to-Major Revision range)
+
+Key Strengths (top 3):
+1. Timely topic -- university attrition prediction highly relevant
+   in Taiwan's demographic context
+2. Good feature engineering -- combining MOE open data with
+   institutional records is novel
+3. Clear methodology -- model comparison (RF, XGBoost, LSTM)
+   is well-structured
+
+Key Concerns (top 3):
+C1: [CRITICAL] Ethical considerations section is missing.
+    C&E requires explicit discussion of algorithmic fairness,
+    student privacy, and potential discriminatory outcomes.
+C2: [MAJOR] Model interpretability not discussed.
+    C&E reviewers expect SHAP values or similar explainability
+    analysis, not just accuracy metrics.
+C3: [MAJOR] Literature gap -- no mention of Dropout Factory
+    literature (Balfanz & Legters, 2004) or recent European
+    attrition studies. Too Taiwan-centric for an international
+    journal.
+
+Quick Verdict: Major Revision
+  Before full submission, address C1-C3.
+  Estimated revision effort: 2-3 weeks.
+```
+
+**Stage 3 Deliverables:**
+- Journal-Fit Reviewer Quick Assessment Report
+- 3 key issues list (priority sorted)
+
+---
+
+### Orchestrator: Transition Recommendation
+
+```
+Stage 3 (REVIEW) quick mode Complete!
+
+Review result: Major Revision direction
+  C1: [CRITICAL] Missing ethics discussion (hard requirement for C&E)
+  C2: [MAJOR] Missing model interpretability analysis
+  C3: [MAJOR] Literature review lacks international perspective
+
+This is the quick review result. You have several options:
+
+  a) Proceed directly to Stage 4 (REVISE) to address these 3 issues
+  b) Upgrade to full review (complete five-seat review),
+     to see if there are other issues
+  c) Revise on your own first, come back when done
+
+My recommendation: Since you're targeting C&E (Q1 journal),
+fix C1-C3 first, then do a full review to ensure quality.
+
+Which do you choose?
+```
+
+**User**: Fix C1-C3 first, then do full review.
+
+### Orchestrator: Update Configuration
+
+```
+OK, updating pipeline:
+
+  Stage 3 REVIEW (quick):  [v] Completed
+  Stage 4 REVISE:          [..] Next -- fix C1-C3
+  Stage 3' RE-REVIEW:      full mode -- full review after revision
+  Stage 5 FINALIZE:        TBD
+
+Start Stage 4 (REVISE)?
+
+Progress: [--]Research -> [--]Writing -> [v]Review -> [..]Revision -> [ ]Re-review -> [ ]Finalization
+```
+
+**User**: Start revision.
+
+---
+
+### Stage 4: REVISE (revision mode, Round 1)
+
+```
+Entering Stage 4 (REVISE) -- using academic-paper revision mode
+
+Loading academic-paper SKILL.md (revision mode)...
+Passing Paper Draft + Quick Assessment Report...
+```
+
+**Revision Progress Tracking:**
+```
+C1: [v] Added Ethics & Fairness section (Section 5.3)
+    - Algorithmic fairness analysis (demographic parity, equalized odds)
+    - Student privacy framework (FERPA-equivalent for Taiwan)
+    - Potential discriminatory outcomes discussion
+    - Added 5 citations on AI ethics in education
+
+C2: [v] Added Model Interpretability analysis (Section 4.3)
+    - SHAP value analysis for top 20 features
+    - Feature importance comparison across 3 models
+    - Case study: 3 individual prediction explanations
+    - Added SHAP visualization figures
+
+C3: [v] Expanded Literature Review (Section 2)
+    - Added Dropout Factory framework (Balfanz & Legters)
+    - Added 6 European attrition studies (2020-2026)
+    - Added cross-cultural comparison discussion
+    - Restructured into 3 sub-sections: global -> Asia -> Taiwan
+```
+
+**Stage 4 Deliverables:**
+- Revised Draft (9,200 words, 2,200 words more than original)
+- Response to Reviewers (R-A-C format, 3 items all addressed)
+
+---
+
+### Orchestrator: Stage 4 --> 3' Transition
+
+```
+Stage 4 (REVISE) Complete!
+
+Revision results:
+- 3/3 issues all addressed
+- Paper expanded from 7,000 to 9,200 words
+- Added Ethics section, SHAP analysis, international literature
+- Added 11 new citations
+
+Next step is Stage 3' (RE-REVIEW) -- this time using a user-requested fresh full review,
+not the contract-governed verification re-review. Five role-separated seats will
+review the revised version (four configured identities plus the fixed Devil's Advocate).
+This is revision cycle 1 (max 2 cycles).
+
+Submit for full review?
+
+Progress: [--]Research -> [--]Writing -> [v]Review -> [v]Revision -> [..]Re-review -> [ ]Finalization
+```
+
+**User**: Submit.
+
+---
+
+### Stage 3': RE-REVIEW (full mode, Loop 1)
+
+```
+Entering Stage 3' (RE-REVIEW) -- full mode, Loop 1/2
+
+Loading academic-paper-reviewer SKILL.md (full mode)...
+Passing Revised Draft + Response to Reviewers + Editorial Decision Letter (#539 Judge Record input)...
+field_analyst_agent analyzing (legitimate here: the user requested a NEW full review
+of the revised version — a fresh review, not a verification re-review — and full mode
+runs field_analyst by definition; Round 1 was an ARS quick assessment. The yardstick
+freeze in re_review_mode_protocol.md § Yardstick Continuity governs re-review mode only)...
+```
+
+**Reviewer Configuration:**
+```
+Journal-Fit Reviewer (serialized source ID EIC):
+            Computers & Education Associate Editor
+            Educational AI, Learning Analytics
+Reviewer 1: Machine Learning expert
+            Predictive modeling, Model evaluation
+Reviewer 2: Higher Education researcher
+            Student success, Retention strategies
+Reviewer 3: Research Ethics / Data Science ethicist
+            Algorithmic fairness, Educational data mining ethics
+Fixed seat: Devil's Advocate
+            Core-argument stress test, strongest counter-explanations
+```
+
+**5 Review Reports Compiled:**
+
+```
+Journal-Fit Reviewer: Minor Revision
+  - C1 response adequate, Ethics section well-structured
+  - Recommend updating Abstract to reflect new ethics + interpretability contributions
+  - Recommend adding a Practical Implications section for university administrators
+
+Reviewer 1: Accept
+  - SHAP analysis thorough and insightful
+  - Model comparison fair and well-documented
+  - Minor: Table 3 precision/recall should include confidence intervals
+
+Reviewer 2: Minor Revision
+  - International literature coverage greatly improved
+  - Recommend Discussion add dialogue with Tinto's Student Integration Model
+  - Minor: Section 2.1 citation Chen (2023) should be Chen (2024)
+
+Reviewer 3: Accept
+  - Ethics section quality is high, exceeding typical C&E standards
+  - Algorithmic fairness metrics well-chosen
+  - Recommend adding a "limitations of fairness metrics" reflection paragraph
+
+Devil's Advocate: No unresolved Critical challenge
+  - Fairness metrics still rely on contested group definitions
+  - Preserve this limitation explicitly rather than presenting the metrics as complete
+```
+
+**Editorial Decision: Minor Revision (Accept direction)**
+
+```
+Required Revisions:
+R1: Update Abstract (reflect new contributions)
+R2: Add Practical Implications section
+
+Suggested Revisions:
+S1: Table 3 add confidence intervals
+S2: Discussion add Tinto's model dialogue
+S3: Fix Chen (2023) -> Chen (2024) citation year
+S4: Add fairness metrics limitations paragraph
+```
+
+---
+
+### Orchestrator: Quick Revision + FINALIZE
+
+```
+Stage 3' (RE-REVIEW) Complete!
+
+Review result: Minor Revision (Accept direction)
+  - Among the 4 non-DA scoring reviewers: 2 Accept + 2 Minor Revision
+  - The fixed Devil's Advocate raised no unresolved Critical challenge
+  - 2 required items (Abstract update + Practical Implications)
+  - 4 suggested modifications
+
+These are all minor changes that can be handled during Stage 5 (FINALIZE).
+Or would you prefer a formal Stage 4 revision before another review?
+
+My recommendation: Since all items are minor, proceed directly to Stage 5 and handle them together.
+
+What format do you need?
+  a) LaTeX (.tex + .bib) -- suitable for C&E submission
+  b) Other formats
+
+Progress: [--]Research -> [--]Writing -> [v]Review -> [v]Revision -> [v]Re-review -> [..]Finalization
+```
+
+**User**: Go straight to finalize, LaTeX format.
+
+---
+
+### Stage 5: FINALIZE (format-convert mode)
+
+```
+Entering Stage 5 (FINALIZE) -- format-convert mode
+
+Simultaneously handling minor revision items:
+R1: [v] Abstract updated
+R2: [v] Added Practical Implications section (200 words)
+S1: [v] Table 3 confidence intervals added
+S2: [v] Discussion Tinto's model paragraph added
+S3: [v] Chen citation year corrected
+S4: [v] Fairness metrics limitations paragraph added
+
+formatter_agent executing LaTeX conversion...
+- Using Elsevier article template (C&E format)
+- Producing .tex + .bib + figures/
+- Producing cover letter draft
+```
+
+**Stage 5 Deliverables:**
+- `ml_university_attrition_taiwan.tex` (LaTeX main text)
+- `references.bib` (BibTeX file)
+- `figures/` (all figures and tables)
+- `cover_letter.tex` (submission cover letter draft)
+
+---
+
+### Orchestrator: Pipeline Complete
+
+```
++=========================================+
+|   Academic Pipeline Status              |
++=========================================+
+| Topic: ML Approaches for Predicting     |
+|        University Attrition Risk        |
++-----------------------------------------+
+
+  Stage 1 RESEARCH    [--] Skipped
+  Stage 2 WRITE       [--] Skipped
+  Stage 3 REVIEW      [v] Completed
+    Mode: quick
+    Decision: Major Revision (3 items)
+  Stage 4 REVISE      [v] Completed
+    Round: 1, Addressed: 3/3
+  Stage 3' RE-REVIEW  [v] Completed
+    Mode: full, Loop: 1/2
+    Decision: Minor Revision (Accept)
+  Stage 5 FINALIZE    [v] Completed
+    Format: LaTeX (Elsevier)
+
++-----------------------------------------+
+| Materials:                              |
+|   [--] RQ Brief (skipped)               |
+|   [--] Methodology (skipped)            |
+|   [--] Bibliography (skipped)           |
+|   [--] Synthesis (skipped)              |
+|   [v] Paper Draft (user-provided)       |
+|   [v] Review Reports (quick + full)     |
+|   [v] Revision Roadmap                  |
+|   [v] Revised Draft (9,500 words)       |
+|   [v] Response to Reviewers             |
+|   [v] Final Paper (LaTeX)              |
++-----------------------------------------+
+| Revision History:                       |
+|   Round 1: Major -> 3/3 addressed       |
+|   RE-REVIEW: Minor (Accept direction)   |
+|   Final minor fixes in FINALIZE         |
++-----------------------------------------+
+| Pipeline Complete!                      |
++=========================================+
+
+Your paper is ready for submission to Computers & Education.
+
+Final deliverables:
+1. ml_university_attrition_taiwan.tex (main text)
+2. references.bib (citations)
+3. figures/ (figures and tables)
+4. cover_letter.tex (submission cover letter draft)
+
+Pre-submission recommendations:
+- Read through the cover letter one more time, adjust to your tone
+- Confirm C&E Author Guidelines for any additional formatting requirements
+- Prepare Highlights (3-5 bullet points) and Graphical Abstract
+```

package/vendor/ars/academic-pipeline/references/adapters/overview.md

@@ -0,0 +1,153 @@
+# Adapter Contract — Producing `literature_corpus[]` Entries
+
+**Status**: Stable as of ARS v3.6.4
+**Applies to**: any adapter that produces a Material Passport `literature_corpus[]` field
+**Authoritative schemas**:
+- [`shared/contracts/passport/literature_corpus_entry.schema.json`](../../../shared/contracts/passport/literature_corpus_entry.schema.json)
+- [`shared/contracts/passport/rejection_log.schema.json`](../../../shared/contracts/passport/rejection_log.schema.json)
+
+## 1. What an adapter is
+
+An **adapter** is a program (in any language) that reads a user-owned corpus source — a Zotero export, an Obsidian vault, a folder of PDFs, a Notion database, a custom SQLite library — and produces:
+
+1. **`passport.yaml`** with a top-level `literature_corpus` array conforming to `literature_corpus_entry.schema.json`.
+2. **`rejection_log.yaml`** conforming to `rejection_log.schema.json`, always emitted (empty when no rejections).
+
+ARS provides three Python reference adapters in [`scripts/adapters/`](../../../scripts/adapters/). Users are expected to write their own adapters for non-reference sources. The three reference adapters are starting points, not production tools.
+
+## 2. Why a contract, not a plugin API
+
+ARS deliberately does NOT run adapter code itself. The adapter runs in the user's environment, reads the user's data, and emits YAML files. ARS reads those YAML files. Rationale:
+
+- Each corpus source (Zotero, Obsidian, custom) has its own versioning, auth, and data quirks. Pinning one plugin API would force ARS to track all of them.
+- Keeping the boundary at "YAML files on disk" means a user can use any language, any scheduling mechanism, and any privacy posture without touching ARS internals.
+- ARS remains a writing/review-layer tool; corpus integration stays in user-owned code.
+
+## 3. Entry field reference
+
+Refer to the [`literature_corpus_entry` schema](../../../shared/contracts/passport/literature_corpus_entry.schema.json) for the authoritative contract. The field tables below are generated from that schema and must not drift.
+
+<!-- GENERATED:LITERATURE_CORPUS_REQUIRED:START -->
+| Field | Type | Description (first sentence) |
+|---|---|---|
+| `authors` | array | CSL-JSON name format. |
+| `citation_key` | string | Stable unique identifier for this entry within the passport. |
+| `source_pointer` | string | Stable URI locating this work in the user's own KB. |
+| `title` | string | — |
+| `year` | integer | Publication year. |
+<!-- GENERATED:LITERATURE_CORPUS_REQUIRED:END -->
+
+<!-- GENERATED:LITERATURE_CORPUS_OPTIONAL:START -->
+| Field | Type | Description (first sentence) |
+|---|---|---|
+| `abstract` | string | PRIVATE FIELD. |
+| `adapter_name` | string | Optional. |
+| `adapter_version` | string | — |
+| `arxiv_id` | string | Optional arXiv identifier for arXiv-hosted works. |
+| `bibliographic_integrity_signals` | array | #678 additive v1 carrier. |
+| `contamination_signal_omissions` | object | #511 Part A reason-provenance for degraded-lookup omissions. |
+| `contamination_signals` | object | v3.7.3 + v3.9.0 + v3.11 contaminated-source advisory field (spec v3.7.3 §3.2 + v3.9.0 §3.4–§3.5 + v3.11 #182 Delta 1). |
+| `contamination_signals_backfilled_at` | string | v3.7.3 backfill provenance (issue #105). |
+| `description_last_audit` | null \| string | v3.7.1 trust-chain field. |
+| `description_source` | string | v3.7.1 trust-chain field. |
+| `doi` | string | DOI without leading 'doi:' or URL prefix. |
+| `obtained_at` | string | Strongly recommended. |
+| `obtained_via` | string | Strongly recommended. |
+| `source_acquired` | boolean | v3.7.1 trust-chain field (spec § 3.1, D1). |
+| `source_acquisition_date` | string | v3.7.1 trust-chain field. |
+| `source_acquisition_path` | string | v3.7.1 trust-chain field. |
+| `source_verification_method` | string | v3.7.1 trust-chain field. |
+| `source_verified_against_original` | boolean | v3.7.1 trust-chain field. |
+| `tags` | array | User-assigned tags from the source KB. |
+| `user_notes` | string | PRIVATE FIELD. |
+| `venue` | string | — |
+| `venue_type` | string | v3.10 (spec §3 PR-B item 2). |
+| `venue_type_provenance` | string | v3.10 (spec §3 PR-B item 3). |
+| `venue_type_source` | string | v3.10 (spec §3 PR-B item 4). |
+<!-- GENERATED:LITERATURE_CORPUS_OPTIONAL:END -->
+
+### 3.1 `authors` format (CSL-JSON names)
+
+Each entry in `authors[]` is one of:
+
+- **Personal name**: `{family: "Chen", given: "Cindy"}`. `family` is required; `given` and CSL particles (`suffix`, `dropping-particle`, `non-dropping-particle`, `comma-suffix`, `static-ordering`, `parse-names`) are optional.
+- **Institution / corporate name**: `{literal: "World Health Organization"}`. `literal` is required; no other fields.
+
+Adapters SHOULD preserve the upstream distinction (personal vs. institution). See [CSL-JSON name spec](https://docs.citationstyles.org/en/stable/specification.html#names) for edge cases.
+
+### 3.2 Privacy caveat for `abstract` and `user_notes`
+
+Publishers typically retain rights to abstracts; user notes often quote copyrighted material. The schema marks these fields as PRIVATE and does NOT enforce anything at the CI level. **If you publish a passport (e.g., commit it to a public repo or share it on the web), you are responsible for removing or clearing these fields.** ARS consumers treat both fields as optional; omitting them never causes failure.
+
+## 4. Rejection log
+
+Whenever an adapter cannot produce a valid `literature_corpus_entry` for an input item, it MUST push a rejection into `rejection_log.yaml.rejected[]` and MUST NOT silently drop the item.
+
+The `reason` field uses a closed enum of categorical values:
+
+| `reason` | When |
+|----------|------|
+| `missing_required_field` | One or more required fields (citation_key / title / authors / year / source_pointer) cannot be derived. Adapter SHOULD also populate `missing_fields`. |
+| `invalid_field_format` | A field is present but fails schema format (e.g., DOI pattern mismatch). |
+| `duplicate_citation_key` | Another entry already used this citekey and the adapter cannot disambiguate. |
+| `unresolvable_source_pointer` | The source URI points to something that does not exist. Adapters typically do not check this; use only if the adapter actually tries to resolve. |
+| `year_unparseable` | Year cannot be extracted from the source field ("n.d.", "forthcoming", "Spring 2024"). |
+| `authors_unparseable` | Authors field is empty, contains only non-author creators, or has unparseable content. |
+| `adapter_error` | Adapter-internal bug, not an input data problem. Use sparingly. |
+| `other` | Anything else. When `reason=other`, `detail` is REQUIRED. |
+
+The `raw` field, when present, MUST be either an object (structured input) or a string (filename or text line). Arrays, numbers, booleans, and null are disallowed so downstream viewers can assume printable shapes.
+
+## 5. Error handling (fail-soft)
+
+**Entry-level problems** → push to rejection log, continue.
+**Adapter-level problems** (input file missing, unreadable, malformed at the root) → write nothing, print a clear error to stderr, exit with code 1.
+
+Do NOT emit a partial passport. Either the passport represents a complete scan of the user's input, or the adapter failed loudly.
+
+## 6. Determinism
+
+Running an adapter twice on identical input MUST produce byte-identical output except for:
+
+- `generated_at` in the rejection log
+- `obtained_at` on each entry
+
+Achieve this by:
+
+- Sorting `literature_corpus[]` by `citation_key`
+- Sorting `rejection_log.rejected[]` by `source`
+- Serializing YAML with sorted keys
+
+## 7. Provenance
+
+Each `literature_corpus` entry SHOULD carry `obtained_via` and `obtained_at`. The `rejection_log` MUST carry `adapter_name`, `adapter_version`, and `generated_at`. These fields let downstream users trace which adapter version produced which entry — useful when a library's schema or a source KB layout changes.
+
+## 8. Extension points for user-written adapters
+
+Custom adapters are welcome and expected. Recommended conventions:
+
+- Set `obtained_via: "other"` on each entry, and set `adapter_name` to a clear string (e.g., `"notion-adapter-v1"`, `"my-custom-sqlite-reader"`).
+- Set `adapter_version` to a semver-ish string so downstream tools can diagnose output changes.
+- Follow the same CLI shape as the three reference adapters (`--input`, `--passport`, `--rejection-log`) if you want to slot into similar tooling.
+
+Common user-written adapter families include:
+
+- **Zotero Web API**: one that calls `api.zotero.org` with a user API token. Not shipped with ARS by design (auth and rate-limit management would pull ARS into data-layer territory).
+- **Notion / Readwise / Airtable**: each has its own SDK; the adapter's job is only to map the source's book/article objects into `literature_corpus_entry` shape.
+- **Cross-source merger**: reads multiple sources and emits a combined passport. Handle citekey collisions explicitly.
+
+## 9. Testing your adapter
+
+Before committing passport output to any workflow:
+
+1. Validate against the schema: `python scripts/check_literature_corpus_schema.py --passport <your_passport.yaml> --rejection-log <your_rejection_log.yaml>`.
+2. Run your adapter twice on the same input and diff the outputs (after stripping `generated_at` / `obtained_at`). Any non-timestamp difference is a determinism bug.
+3. Feed a known-bad entry to confirm it lands in the rejection log rather than silently vanishing or crashing.
+
+The three reference adapters have pytest coverage under `scripts/adapters/tests/` — copying that pattern for your own adapter is a good starting point.
+
+## 10. Relationship to other ARS artifacts
+
+- [`shared/handoff_schemas.md`](../../../shared/handoff_schemas.md) Schema 9: the `literature_corpus[]` field lives inside the Material Passport.
+- [`academic-pipeline/references/passport_as_reset_boundary.md`](../passport_as_reset_boundary.md): `literature_corpus[]` is consumed across reset boundaries like any other passport field.
+- ARS agents that consume `literature_corpus[]` are **deferred** to v3.6.5+. As of v3.6.4, the field is a defined input port with no runtime consumer; adapters produce it, future ARS versions read it.