zrwangems/dsh-zrwang-plugin / ars/skills/deep-research
deep-research
Flagged学术研究前置:把模糊想法收敛成可研究的问题,并做文献综述、事实核查、系统综述(PRISMA)与三向文献扫描。Triggers: research, deep research, literature review, systematic review, meta-analysis, PRISMA, fact-check, guide my research | 研究, 文献综述, 系统性文献回顾, 事实核查, 引导我研究, 帮我确定研究问题。输出 RQ Brief、方法论蓝图、带核验的注释书目。
_shared/references/review_criteria_consumer_protocol.md
View on GitHub> 输出语言(本 DSH 集成强制):所有面向用户的文字、稿件、摘要、评审与报告一律使用**简体中文**;英文术语与引文保持英文。文中的 `zh-TW` / `zh-tw-en` 只是上游语言代码与脚本参数,不代表要用繁体输出。用户明确要求其他语言时以用户为准。
>
> 多角色执行(评审团、十阶段流水线、角色扮演)必须使用 DSH 的 Agent Team 工具(`spawn_teammate`、`team_task_*`、`wait_agent`、`send_message`);当前会话看不到这些工具,说明用户没开 Agent Team,必须先提示用户开启,不得退化成单会话顺序扮演。
# Review criteria consumer protocol (#684)
This protocol binds the #683 `ReviewTargetContext` across formative planning,
the academic-paper internal evaluator, and the academic-paper-reviewer external
panel. The binding is workflow conformance only. It never determines severity,
an editorial verdict, a checkpoint state, or author triage.
Normative design:
`docs/design/2026-08-11-684-review-criteria-consumer-binding-spec.md`.
Schemas:
`_shared/contracts/review_target/review_criteria_binding_manifest.schema.json`
and
`_shared/contracts/review_target/constructive_review_findings.schema.json`.
Tool: `scripts/review_criteria_binding.py`.
## 1. Inputs and authority
The orchestrating layer receives four explicit inputs:
1. the author-confirmed `ReviewTargetDeclaration` already embedded in the
resolved context;
2. the exact `ReviewTargetContext` JSON;
3. the exact criteria registry named by that context; and
4. a caller-supplied `target_review_id`.
Initialize one manifest before any of the three consumers run:
```bash
python scripts/review_criteria_binding.py init \
--context phase0/review_target_context.json \
--context-ref phase0/review_target_context.json \
--registry _shared/review_criteria_registry.json \
--registry-ref _shared/review_criteria_registry.json \
--target-review-id review-001 \
--output phase0/review_criteria_binding.json
```
If an active manifest exists, pass it as `--prior-manifest`. A substantive
context change under the same `target_review_id` is a conflict. Starting a new
target review requires a new id and makes the predecessor non-comparable.
The manifest is the exact source of:
- `context_ref` and its raw SHA-256;
- `resolved_digest`;
- ordered `selected_criterion_ids`;
- exact criterion id/version/content-digest pointers;
- unresolved target dimensions; and
- parallel interdisciplinary conflict groups.
Consumers never reconstruct these fields from the Markdown brief and never
copy registry statements into their own instruction files.
## 2. Marker and receipt lifecycle
Before dispatch, render the role-specific marker:
```bash
python scripts/review_criteria_binding.py marker \
--manifest phase0/review_criteria_binding.json \
--consumer formative_planning --role FORMATIVE
```
The marker is appended verbatim to the consumer's output. After a successful
consumer call, the output also carries exactly one
`criteria_parallel_conflicts: <canonical compact JSON array>` line rendered
from the manifest. Record only explicitly named artifacts:
```bash
python scripts/review_criteria_binding.py record \
--manifest phase0/review_criteria_binding.json \
--consumer formative_planning \
--artifact FORMATIVE=phase2/paper_outline.md
```
The internal receipt similarly names the paper-blind Phase 6a artifact, not
the later paper-visible decision:
```bash
python scripts/review_criteria_binding.py record \
--manifest phase0/review_criteria_binding.json \
--consumer internal_evaluator \
--artifact INTERNAL=phase6/phase6a_precommitment.md
```
For the external panel, supply all five Phase 1 artifacts in one call:
```bash
python scripts/review_criteria_binding.py record \
--manifest phase0/review_criteria_binding.json \
--consumer external_panel \
--artifact EIC=phase1/eic.md --artifact R1=phase1/methodology.md \
--artifact R2=phase1/domain.md --artifact R3=phase1/perspective.md \
--artifact DA=phase1/da.md
```
The recorder checks the complete marker and exact conflict line inside every
artifact, computes raw hashes itself, and appends one receipt atomically. Exact
retry is a no-write. Same-consumer drift is a conflict.
Before a review crosses its handoff boundary, run:
```bash
python scripts/review_criteria_binding.py validate \
--manifest phase0/review_criteria_binding.json \
--context phase0/review_target_context.json \
--registry _shared/review_criteria_registry.json \
--require-complete
```
## 3. Formative planning and writing
Owner: `academic-paper` Phase 2 `structure_architect_agent`.
- Phase 0 resolves the author-confirmed target before Phase 2.
- The Structure Architect receives the pointer-only context and Target Criteria
Brief. Its outline names relevant criterion ids beside planned sections and
evidence needs, while keeping scientific validity, venue fit, and submission
readiness separate.
- `argument_builder_agent` and `draft_writer_agent` consume the same context
pointer and the Structure Architect's receipt. They do not create a second
receipt, re-resolve the target, copy criterion prose into their prompts, or
silently change the author's contribution claim.
- A criterion can motivate a planned check; it never licenses invented evidence,
results, methods, or a new research question.
The formative output records `FORMATIVE` with the exact marker.
## 4. Internal evaluator
Owner: `academic-paper` Phase 6 `peer_reviewer_agent`.
- Phase 6a receives the sprint/evaluator contract, paper metadata, the
pointer-only context, and Target Criteria Brief. It remains paper-blind.
- Phase 6a commits the selected criterion ids and keeps each
`parallel_conflicts[]` group separate. It does not declare applicability.
- The Phase 6a artifact records the `INTERNAL` receipt before any draft is
visible. Phase 6b receives that unchanged artifact and the draft, assesses
applicability, repeats the continuity marker, and emits its Critical/Major
constructive sidecar when applicable.
The internal evaluator may identify a scientific-validity problem independent
of venue fit. It may not turn missing or advisory venue guidance into a blocking
finding.
## 5. External panel, including DA
Owner: the academic-paper-reviewer orchestrating layer.
- Phase 0 uses author-confirmed target metadata. `field_analyst_agent` may
suggest panel expertise but cannot infer or overwrite venue/track/type.
- Each of `EIC`, `R1`, `R2`, `R3`, and `DA` receives the same context authority
and brief during the existing paper-content-blind Phase 1 call.
- Each Phase 1 output repeats its role-specific marker and commits the selected
criterion ids. Conflicting criteria stay parallel.
- Phase 2 receives the unchanged Phase 1 output plus manuscript data. Findings
cite exact criterion pointers and manuscript evidence/absence anchors.
- `editorial_synthesizer_agent` checks all five marker bindings before
synthesis. Missing or mismatched bindings abort the criteria-aware run; the
synthesizer never silently falls back to a self-recalled target.
DA adversarial register does not relax the finding contract. Every DA Critical
or Major item enters the same constructive sidecar and carries the same remedy,
cost, trade-off, author-intent, and no-honest-remedy fields as another seat.
## 6. Constructive finding sidecar
The sidecar covers Critical/Major weaknesses only. Minor and copyedit channels
retain their existing formats. Validate it with:
```bash
python scripts/review_criteria_binding.py validate-findings \
--manifest phase0/review_criteria_binding.json \
--context phase0/review_target_context.json \
--registry _shared/review_criteria_registry.json \
--findings phase2/constructive_review_findings.json
```
Every row separates:
- criterion pointers from manuscript evidence;
- scholarly relevance from confirmed-target relevance;
- the minimum viable remedy from a stronger costlier option;
- effort scope from trade-offs; and
- reviewer advice from author-owned research intent.
If no honest remedy exists, `remedy_status=none`, both option fields are null,
and `no_honest_remedy_reason` explains why. If a stronger option is not
meaningful, it is null and `stronger_option_reason` explains why. An available
option always records `proposed_result_values=false`; new data or intent change
is an author choice, not an assertion that work or results exist.
A stale/unverified or otherwise non-blocking criterion may appear in ordinary
advisory prose, but the validator rejects a Critical/Major sidecar row whose
every criterion pointer has `blocking_eligible=false`.
## 7. Interdisciplinary applicability
Phase 1 commits each selected pointer without manuscript-derived applicability.
After the paper-visible boundary, a consumer may state:
- `applicable`, with a manuscript evidence/absence anchor;
- `not_applicable`, with a manuscript-grounded rationale; or
- `unresolved`, requiring human resolution.
No consumer averages parallel criteria, assigns an adaptive numeric weight, or
selects the criterion whose outcome is easiest to satisfy. Human resolution
belongs in a separate author/reviewer record and starts a new target review if
it changes the selected context.
## 8. Failures and degradation
Binding failures are visible workflow-conformance failures:
- malformed/oversized/symlinked input;
- context or registry hash drift;
- context recomputation mismatch;
- missing/duplicate/wrong-role receipt;
- missing or altered marker;
- selected-pointer mismatch;
- same-id profile change; or
- invalid constructive sidecar.
They stop only the criteria-aware handoff. They do not create an editorial
decision. The user may correct the artifacts, explicitly start a new target
review, or run the pre-#684 field-general workflow with an explicit
`criteria_binding_unavailable` disclosure. That degraded path makes no venue-
alignment claim and is non-comparable to a bound treatment run.
## 9. Measurement boundary
`evals/heldout/review_criteria_constructive_value/measurement_plan.md` owns the
paired behavioral design. CI validates schemas, fixtures, scorer behavior, and
prompt wiring only. It never dispatches a model, judge, or network request.
Until a valid `heldout-measurement/1.1` report is published, describe the
mechanism as implemented but its effect on unsupported-finding rate, severity,
venue alignment, and usefulness as **unmeasured**.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| add-product-e2e-evalpaperclipai/paperclipAdd or extend a Paperclip full-stack runner E2E workflow, fixture, matcher, or report evidence path for local or Daytona execution. | Stars: 115k | Updated | Claude CodeCodex | Verified | Install |
| add-runner-evalpaperclipai/paperclipAdd or extend a Paperclip Runner protocol evaluation definition, roster, assertion, or report fixture with provenance and narrow validation. | Stars: 115k | Updated | Claude CodeCodex | Verified | Install |
| diagnose-why-work-stoppedpaperclipai/paperclipCompatibility entry point for failed, stalled, looping, or over-recovered Paperclip work. Use fix-failed-paperclip-run for the full diagnosis, regression, repair, real verification, and green-PR workflow; honor an explicitly requested diagnosis-only scope. | Stars: 115k | Updated | Claude CodeCodex | Verified | Install |
| fix-failed-paperclip-runpaperclipai/paperclipDiagnose and fix failed, stuck, slow, or incorrectly blocked Paperclip tasks and runs. Use with a task/run URL or failure report to trace the root cause, prove a regression red then green, check second-order effects, verify the real workflow, and deliver a PR with passing checks. | Stars: 115k | Updated | Claude CodeCodex | Verified | Install |