# press-conference-revision-evidence-bound

Adopt this skill: https://proskills.md/skills/other/lensback940701-evidence-bound-press-conference-revision-skill

Page: https://proskills.md/skills/other/lensback940701-evidence-bound-press-conference-revision-skill
This document is the public skill bundle for that page. It contains SKILL.md and every helper file the catalog carries. No login is required.

## File: SKILL.md
size: 8987
sha: b31dc068d99ac1bc0897cef32ec4e9970c5f21bb
upstream: https://github.com/lensback940701/evidence-bound-press-conference-revision-skill/blob/HEAD/SKILL.md

```text
---
name: press-conference-revision-evidence-bound
description: Diagnose and revise defensive academic writing while preserving claim ceilings, evidence status, scope conditions, rival explanations, and conceptual hierarchy. Use for a whole-manuscript batch audit or a tightly authorized, tracked-change revision when an academic paper hides its contribution beneath self-defence, repeated caveats, process narration, or unnecessary comparisons; especially when a qualitative, historical, or mixed-method manuscript must become contribution-forward without becoming promotional or overclaimed. Do not use for first drafting, new research, citation repair, evidence admission, grammar-only polishing, or making claims sound stronger than the evidence permits.
---

# Evidence-Bound Press-Conference Revision

Treat a paper as a clear academic presentation of its best supported contribution, not as a work log or a self-audit. Reframe defensiveness; never remove the information that lets readers judge what the evidence supports.

## Non-negotiable distinction

Do not treat a cue word as a verdict. The target is **rhetorical defensiveness**, not scholarly caution.

- Cut, compress, or reframe self-deprecation, anticipatory answers to objections the manuscript has not invited, redundant hedge stacks, process-log narration, and negative judgments broader than the evidence.
- Keep or relocate a real scope condition, source-status distinction, rival explanation, contradiction, negative finding, ethical caveat, or method limitation whenever it changes how the claim must be read.
- When uncertain, retain the wording and register a query. Do not convert possible to certain, report to observation, authorization to delivery, sequence to mechanism, or a bounded case result to a general law.

Read `references/triage-and-regression.md` before classifying candidates. For the Dalian *Urban Geography* manuscript, also read `references/urban-geography-guardrails.md` before any diagnosis or revision.

## Required inputs and authority order

Collect the smallest current set of materials needed for the task:

1. the latest clean manuscript and, for a revision pass, its tracked baseline;
2. the current manuscript lock and any later author-authorized departure from it;
3. current author decisions and unresolved queries;
4. target-journal guidance and any supplied corpus/style calibration;
5. the source principles and prior audit/return records, as context rather than evidence.

Apply this order: manuscript evidence and active author decisions → lock and claim ceiling → verified journal guidance → corpus-level rhetorical calibration → this skill. Do not use a corpus as a phrase bank or let a style preference override a source status.

## Phase 0 — create a contribution-and-scope contract

Before scanning, write a compact contract. Do not edit until it is complete.

| field | required content |
| --- | --- |
| core contribution | One evidence-bound sentence naming the primary analytical move and payoff. |
| contribution hierarchy | Primary concept; supporting lens; bounded empirical contribution; portable payoff, if any. |
| claim ceiling | The strongest permissible formulation already supported by the manuscript and its evidence. |
| load-bearing cautions | Scope, source-status, method, rival, contradiction, and ethics statements that must survive. |
| edit authority | Sections/paragraphs authorized, locked items, tracking requirements, and prohibited changes. |
| live queries | Meaning-dependent choices that must not be silently resolved. |

If the hierarchy, evidence ceiling, or editing authority is unclear, stop after the diagnostic register and request a decision.

## Phase 1 — batch diagnostic (read-only)

Read the whole manuscript once for section roles and the contribution contract. Then make a candidate register; do not revise prose in this phase.

1. Optionally run `scripts/scan_defensive_cues.py` on the clean DOCX to create a **candidate list only**. Inspect every result in context; the script makes no rhetorical or evidentiary decision.
2. Review every prose-bearing paragraph, including title, abstract, introductions, discussion, conclusion, captions, notes, and references only where they affect visible prose.
3. Classify each candidate using the taxonomy in `references/triage-and-regression.md`.
4. Record the candidate in a copy of `assets/defensive-writing-audit-template.csv` or an equivalent table. Include its evidentiary function, source-status tag, contribution relation, and provisional disposition.
5. Mark non-defensive but load-bearing language as `KEEP`; this makes the integrity decision auditable rather than invisible.

Allowed dispositions are `KEEP`, `TIGHTEN`, `REFRAME`, `RELOCATE`, `CUT`, and `QUERY`. A `CUT` must remove rhetoric only, never a factual proposition, citation function, evidence status, rival, or scope limit.

## Phase 2 — concentrated revision (only with authority)

Revise by pattern group and section role, not by a blind word hunt. Use tracked changes when the manuscript is under review.

1. Start with the abstract, introduction, discussion, and conclusion: make the problem, gap, distinctive analytical move, evidence-bound finding, and payoff legible before qualifications appear.
2. In theory and methods, replace self-defence with precise positioning, but keep definitions, scope, evidence-status distinctions, and analytical rivals.
3. In empirical sections, retain contradictions, delay, non-adoption, reversal, and negative findings when they are the data or explain the mechanism. Remove only authorial apology or unnecessary rehearsal of a limitation.
4. Keep a contribution sentence standing in each major section. A later qualifier may narrow it; it may not erase it or turn the paragraph into a response to a hypothetical reviewer.
5. For every changed sentence, run the three-part preservation test: (a) same or narrower claim, (b) same source-status distinction, and (c) same citation/quotation role.
6. Route a change to `QUERY` if it would alter research questions, terminology, hierarchy, evidence, citation role, table/figure text, paragraph order, or a real methodological limitation.

Never solve a mismatch by adding evidence, strengthening a causal verb, or extending a case-specific statement into a general claim.

## Phase 3 — whole-manuscript regression

Read the revised manuscript independently of the change list. Use the regression protocol in `references/triage-and-regression.md`.

Verify all of the following:

- contribution hierarchy is stable from title and abstract through conclusion;
- a new general-first opening is aligned with the body, or the mismatch is explicitly left as a query;
- every changed claim stays within the original evidence ceiling;
- plan, authorization, reported work, observation, outcome, interpretation, and causal inference remain distinct;
- terminology is consistent without flattening meaningful distinctions;
- citations, quotations, dates, numbers, captions, footnotes, figure/table cross-references, and paragraph boundaries remain protected;
- no automatic global replacement has elevated a secondary lens or introduced a prohibited keyword;
- the clean file has accepted changes only and the tracked file exposes every text change.

Use a global terminology scan only to locate candidates. Never replace a concept merely because it is frequent.

## Dalian Urban Geography special gate

When the target is the Dalian heritage-metabolism manuscript, apply the project guardrails before Phase 1 and again before delivery. In particular:

- keep heritage metabolism primary and operation-centred;
- keep rescaling supporting, rival-open, and conditional on a documented capacity-to-operation relation;
- do not promote governance/governing into a master concept or weaken source-specific terms by global substitution;
- preserve overlapping configurations, contradiction, lag, decoupling, reversal, and recombination as analytic findings rather than defensive residue;
- treat the recent general-first abstract/introduction reframe as a cross-section consistency question, not permission to rewrite the body without explicit authority.

## Deliverables and stop rules

For a diagnostic pass, return:

- the contribution-and-scope contract;
- the candidate register with `KEEP` decisions visible;
- a section-level concentration map and an exception/query register;
- a bounded revision recommendation, including which changes require author authority.

For an authorized revision pass, additionally return:

- tracked and clean manuscript copies;
- a change summary organized by pattern group;
- a claim/evidence regression report;
- a list of deliberate non-edits and unresolved author decisions.

Stop after the requested diagnostic or authorized revision pass. Do not submit, conduct new research, change citations, silently resolve a meaning-dependent query, or turn the paper into promotional prose.
```

## File: README.md
size: 24682
sha: dea42d0bbea40c6e446413233f6dd5cd37942ae9
upstream: https://github.com/lensback940701/evidence-bound-press-conference-revision-skill/blob/HEAD/README.md

````text
# Evidence-Bound Press-Conference Revision

[中文说明](README.zh-CN.md)

**Contribution-forward academic revision without overclaiming.**

Co-created on RedNote by [@Lensback's Lab](https://www.xiaohongshu.com/user/profile/64507c470000000012036d99) and [@Naruto AI](https://www.xiaohongshu.com/user/profile/666940c40000000007005b9a).

<p align="center">
  <img src="assets/readme/hero.png" alt="Evidence-Bound Press-Conference Revision" width="900">
</p>

`press-conference-revision-evidence-bound` is an academic revision skill for diagnosing and reducing **defensive writing** while preserving the evidentiary limits that make scholarly claims credible.

It is designed for manuscripts whose contribution is already present and supported, but is obscured by repeated caveats, self-defence, work-log narration, excessive qualification, unnecessary comparisons, or contribution statements buried beneath anticipatory responses to hypothetical reviewers.

The skill does **not** make a paper more assertive by inflating its claims. Instead, it helps a manuscript present its strongest supported contribution clearly while retaining the scope conditions, source-status distinctions, rival explanations, contradictions, methodological limits, and citation functions that determine what the evidence actually permits.

> **Core principle:** remove rhetorical defensiveness, not scholarly caution.

---

## Why “press-conference revision”?

Academic manuscripts often accumulate defensive prose during drafting, co-author revision, peer review, and repeated attempts to anticipate criticism.

The result can be a paper that reads less like an argument and more like a record of everything the authors worried about while producing it:

* the contribution appears only after several qualifications;
* limitations are repeated far beyond the point at which they constrain interpretation;
* paragraphs begin by explaining what the paper does *not* claim;
* methodological caution becomes self-deprecation;
* failed analytical routes remain visible as work-log narration;
* comparisons with stronger cases or literatures are volunteered even when they do not advance the argument;
* the paper answers objections that its own claims never created.

The **press-conference** metaphor asks a simple question:

> If the evidence is fixed, what should the paper say first about what it has actually established?

A good revision therefore does not conceal weaknesses or manufacture certainty. It reorganizes the manuscript so that the supported proposition leads, while the necessary qualification appears where it performs real analytical work.

The goal is:

**contribution → evidence → qualification**

rather than:

**qualification → self-defence → more qualification → eventual contribution**

---

## What this skill protects

The method is intentionally evidence-bound.

Every revision must preserve, or narrow, the original manuscript's epistemic commitments. In particular, the skill protects:

### Claim ceiling

A revision cannot increase the certainty, causality, novelty, generality, representativeness, or portability of a claim beyond what the existing evidence supports.

`may shape` cannot silently become `determines`.

`the interviews suggest` cannot become `the evidence demonstrates`.

`in this case` cannot become `cities generally`.

### Evidence status

The workflow keeps analytically different kinds of evidence separate, including:

* proposal;
* authorization;
* reported activity;
* direct observation;
* implementation;
* outcome;
* interpretation;
* causal inference.

Fluency is never a sufficient reason to collapse these distinctions.

### Scope conditions

Case, temporal, population, institutional, geographical, and comparative boundaries remain visible wherever they are needed to interpret a claim correctly.

### Rival explanations and negative evidence

Contradiction, delay, non-adoption, decoupling, reversal, failed implementation, and competing explanations are not treated as embarrassing residue when they are part of the empirical record.

### Citation and quotation roles

A stylistic revision must not silently change what a citation, quotation, date, number, source-specific term, table, figure, or footnote is being asked to support.

<p align="center">
  <img src="assets/readme/principle.png" alt="Evidence-Bound Press-Conference Revision" width="900">
</p>

---

## What the skill does

The skill supports two distinct modes.

### 1. Read-only diagnostic

The manuscript is audited without changing its prose.

The diagnostic produces:

* a **contribution-and-scope contract**;
* a manuscript-wide candidate register;
* explicit `KEEP` decisions for necessary caution;
* a section-level concentration map;
* an exception and query register;
* a bounded recommendation for any later revision.

This mode is useful when the author wants to understand whether defensive writing is actually a manuscript-level problem before authorizing changes.

### 2. Authorized revision

Once revision authority is explicit, the skill can perform a concentrated revision while preserving the evidence ceiling.

Typical outputs include:

* a tracked manuscript;
* a clean manuscript;
* a change summary organized by rhetorical pattern;
* a claim/evidence regression report;
* a record of deliberate non-edits;
* unresolved author decisions requiring human judgment.

The distinction is important: **diagnosis does not automatically authorize rewriting.**

---

## Workflow

```mermaid
flowchart LR
    A["Phase 0<br/>Contribution & Scope Contract"] --> B["Phase 1<br/>Batch Diagnostic"]
    B --> C{"Author authority<br/>to revise?"}
    C -- No --> D["Diagnostic Register<br/>+ Queries + Recommendation"]
    C -- Yes --> E["Phase 2<br/>Concentrated Revision"]
    E --> F["Phase 3<br/>Whole-Manuscript Regression"]
    F --> G["Tracked + Clean Manuscript<br/>+ Integrity Report"]
```

### Phase 0 — Contribution-and-scope contract

Before editing begins, the skill establishes a compact contract containing:

| Field                      | Purpose                                                                              |
| -------------------------- | ------------------------------------------------------------------------------------ |
| **Core contribution**      | The strongest evidence-bound statement of the manuscript's primary analytical payoff |
| **Contribution hierarchy** | Primary concept, supporting lenses, empirical contribution, and portable payoff      |
| **Claim ceiling**          | The strongest formulation already permitted by the manuscript and its evidence       |
| **Load-bearing cautions**  | Qualifications that must survive revision                                            |
| **Edit authority**         | What may and may not be changed                                                      |
| **Live queries**           | Meaning-dependent decisions that must remain with the author                         |

If the contribution hierarchy, evidence ceiling, or editing authority cannot be established, the workflow stops at diagnosis rather than improvising a solution.

---

### Phase 1 — Batch diagnostic

The manuscript is first read as a whole.

The skill does **not** perform a blind search-and-replace operation. Lexical cues are treated only as signals for contextual inspection.

Candidates are classified using the following taxonomy:

| Code   | Pattern                                                    | Typical response                               |
| ------ | ---------------------------------------------------------- | ---------------------------------------------- |
| **D1** | Self-deprecation or apology                                | Cut or reframe                                 |
| **D2** | Anticipatory defence                                       | Cut, tighten, or relocate                      |
| **D3** | Stacked hedge                                              | Retain the epistemically necessary hedge       |
| **D4** | Work-log narration                                         | Reframe around problem, solution, and evidence |
| **D5** | Buried contribution                                        | Reframe or relocate                            |
| **D6** | Unnecessary comparison / volunteered loss                  | Narrow, reframe, or cut                        |
| **D7** | Broad negative self-judgment                               | Tighten to demonstrated scope                  |
| **D8** | Necessary caveat in the wrong rhetorical position          | Relocate                                       |
| **K1** | Scope condition                                            | Keep                                           |
| **K2** | Source-status distinction                                  | Keep                                           |
| **K3** | Rival, contradiction, delay, reversal, or non-adoption     | Keep                                           |
| **K4** | Ethical, provenance, reproducibility, or method limitation | Keep                                           |

The `K` categories are deliberately visible in the audit.

A high number of `KEEP` decisions is **not a failed revision**. It may indicate that the manuscript contains disciplined evidence-bound qualification rather than rhetorical defensiveness.

The unit of judgment is the **function of a statement in context**, not the presence of a particular word.

---

## The three-question test

Every candidate must pass three questions before it is changed:

1. **What function does this wording serve?**
   Is it rhetorical self-defence, evidence status, scope, method, rival explanation, an empirical result, or several of these simultaneously?

2. **What would disappear if it were removed?**
   The answer must identify the proposition or analytical boundary that would be lost—not simply describe a change in tone.

3. **Does the replacement preserve the ceiling?**
   The new wording must not increase certainty, causality, novelty, generality, representativeness, or portability.

If these questions cannot be answered confidently, the candidate becomes `QUERY`.

---

## Phase 2 — Concentrated revision

Authorized revisions are performed by **pattern group and section role**, rather than through global cue-word replacement.

Priority normally goes to:

1. abstract;
2. introduction;
3. discussion;
4. conclusion;
5. theoretical positioning;
6. methods;
7. empirical sections where defensive rhetoric materially obstructs interpretation.

A major section should normally contain a visible contribution sentence before qualification begins to dominate the paragraph.

Necessary caution remains, but is moved to the location where it does real analytical work.

For example:

**Defensive structure**

> Although this study cannot establish a universally applicable causal relationship and has several limitations arising from its case-specific design, the findings may nevertheless tentatively suggest that...

**Contribution-forward structure**

> The case shows that X operated through Y under these institutional conditions. This inference is limited to the documented case and does not establish a general causal relationship.

The second version is not stronger in evidentiary terms. It is simply clearer about **what is supported first** and **where the boundary lies**.

---

## Phase 3 — Whole-manuscript regression

Revision is followed by an independent regression pass.

This is not a proofreading step. Its purpose is to detect whether rhetorical improvement has accidentally altered the manuscript's epistemic structure.

<p align="center">
  <img src="assets/readme/audit-loop.png"
       alt="Evidence-bound audit and regression loop"
       width="900">
</p>

The regression checks:

* **claim delta** — no evidentiary inflation;
* **evidence status** — plan, report, observation, outcome, interpretation, and causation remain distinct;
* **citation role** — sources still support the propositions for which they were originally used;
* **scope** — relevant boundaries remain visible;
* **conceptual hierarchy** — supporting concepts do not accidentally become master concepts;
* **section role** — different sections perform complementary analytical work;
* **voice** — prose is direct without becoming promotional.

---

## Candidate dispositions

The audit uses six possible editorial dispositions:

| Disposition | Meaning                                                                                            |
| ----------- | -------------------------------------------------------------------------------------------------- |
| `KEEP`      | The wording performs necessary evidentiary or analytical work                                      |
| `TIGHTEN`   | The function is necessary but expressed redundantly                                                |
| `REFRAME`   | The proposition should remain but its rhetorical organization is defensive                         |
| `RELOCATE`  | Necessary information is currently placed where it obscures the contribution                       |
| `CUT`       | The rhetoric can disappear without removing a proposition, evidence boundary, or citation function |
| `QUERY`     | The decision depends on authorial meaning or evidentiary judgment                                  |

`CUT` therefore has a deliberately high threshold.

It can remove rhetoric. It cannot remove evidence.

---

## Automated cue scanner

The repository includes a lightweight DOCX cue scanner:

```text
scripts/scan_defensive_cues.py
```

It can create a preliminary candidate list from a clean `.docx` manuscript.

Example:

```bash
python scripts/scan_defensive_cues.py manuscript.docx \
  --out defensive-writing-candidates.csv
```

The script currently detects selected lexical patterns associated with:

* self-deprecation;
* anticipatory defence;
* hedge clusters;
* work-log narration;
* unnecessary comparison or concession.

It uses the Python standard library and reads paragraph text directly from the DOCX package.

### Important

**The scanner is not an automated editor and not a rhetorical classifier.**

A matched cue is only a candidate for human or agent inspection.

The scanner intentionally outputs:

```text
REVIEW_REQUIRED
```

rather than deciding whether a sentence should be kept, tightened, reframed, relocated, cut, or queried.

This distinction is central to the design of the skill:

> lexical detection may be automated; epistemic judgment should remain contextual.

---

## Audit register

The repository includes:

```text
assets/defensive-writing-audit-template.csv
```

The register records, among other fields:

* manuscript location;
* pattern code;
* candidate text;
* rhetorical function;
* evidentiary function;
* evidence status;
* relationship to the manuscript's contribution;
* provisional disposition;
* claim-delta check;
* citation/source-status check;
* whether an author decision is required;
* reviewer notes.

This makes the revision process inspectable rather than leaving integrity judgments implicit.

---

## Repository structure

```text
press-conference-revision-evidence-bound/
├── README.md
├── README.zh-CN.md
├── SKILL.md
├── agents/
│   └── openai.yaml
├── assets/
│   ├── defensive-writing-audit-template.csv
│   └── readme/
├── references/
│   ├── triage-and-regression.md
│   └── urban-geography-guardrails.md
└── scripts/
    └── scan_defensive_cues.py
```

### `SKILL.md`

Defines the core workflow, authority order, revision phases, preservation tests, deliverables, and stop rules.

### `references/triage-and-regression.md`

Contains the D1–D8 / K1–K4 candidate taxonomy, three-question test, caution-placement rules, and post-revision regression protocol.

### `scripts/scan_defensive_cues.py`

Provides lightweight lexical pre-screening of DOCX manuscripts. It generates candidate locations but deliberately makes no editorial decision.

### `assets/defensive-writing-audit-template.csv`

Provides a structured audit surface for candidate classification and integrity checking.

### `references/urban-geography-guardrails.md`

Demonstrates how the generic method can be supplemented with **project-specific guardrails**.

The bundled example protects a manuscript-specific conceptual hierarchy and evidence firewall. It is not required for generic use of the skill.

This illustrates a broader design principle:

```text
generic revision protocol
        +
project-specific evidence/concept guardrails
        =
bounded manuscript revision
```

Users working with manuscripts that have tightly defined concepts, source hierarchies, or previously agreed author decisions can create comparable project-specific guardrail files.

---

## Installation

Clone the repository into your Codex skills directory:

```bash
git clone https://github.com/lensback940701/Evidence-Bound-Press-Conference-Revision-Skill.git \
  ~/.codex/skills/press-conference-revision-evidence-bound
```

On Windows, the usual destination is:

```text
%USERPROFILE%\.codex\skills\press-conference-revision-evidence-bound
```

Restart or refresh Codex after installation. Invoke the skill as `$press-conference-revision-evidence-bound`.

---

## When to use this skill

This workflow is particularly useful when:

* a manuscript has already undergone several rounds of revision;
* peer-review responses have made the prose increasingly defensive;
* the contribution exists but is difficult to locate;
* the introduction spends too much space anticipating criticism;
* caveats recur across multiple sections;
* the manuscript repeatedly explains what it does not claim;
* qualitative or historical evidence requires careful source-status distinctions;
* competing explanations must remain visible;
* a co-author wants stronger prose but the evidentiary ceiling must not move;
* an AI-assisted revision needs an explicit integrity harness;
* tracked and auditable revision is preferable to unrestricted rewriting.

It is especially suited to qualitative, historical, interpretive, case-based, and mixed-method research where a small rhetorical change can materially alter what the evidence appears to establish.

---

## When *not* to use it

This skill is not designed for:

* first drafting;
* literature searches;
* new empirical research;
* citation discovery or citation repair;
* fact checking;
* evidence admission;
* generic grammar correction;
* copyediting alone;
* restructuring a paper without author authority;
* inventing stronger contribution claims;
* making results sound more causal, general, novel, or representative than they are.

If the underlying problem is weak evidence rather than defensive presentation, this skill should not be used to disguise that weakness.

---

## Example prompts

### Whole-manuscript diagnostic

```text
Use $press-conference-revision-evidence-bound to conduct a read-only,
whole-manuscript audit for defensive academic writing.

Do not revise the manuscript yet.

First establish the contribution-and-scope contract, then classify candidates
using the D1–D8 / K1–K4 taxonomy. Keep necessary scholarly caution visible as
KEEP decisions and identify all changes that would require author authority.
```

### Authorized revision

```text
Use $press-conference-revision-evidence-bound to revise the authorized sections
of this manuscript.

Make the supported contribution more visible and reduce rhetorical defensiveness,
but preserve the existing claim ceiling, evidence-status distinctions, scope
conditions, rival explanations, citations, quotations, and conceptual hierarchy.

Run the full regression protocol after revision and report deliberate non-edits
and unresolved author decisions.
```

### Diagnostic with cue scanner

```text
Use the bundled defensive-writing cue scanner only as a candidate-generation
step. Inspect every matched paragraph in context before assigning a disposition.

Do not treat lexical frequency as evidence of poor writing and do not perform
automatic global replacement.
```

---

## Design philosophy

### Contribution-forward, not promotional

The purpose is not to make academic writing sound confident for its own sake.

A sentence becomes stronger only when its **rhetorical position** becomes clearer—not when its evidentiary status is upgraded.

### Evidence-bound, not maximally cautious

More hedging is not automatically more rigorous.

Redundant caution can obscure exactly what the evidence *does* support.

The objective is to preserve the **minimum sufficient epistemic qualification**, not the maximum possible amount of defensive prose.

### Contextual, not lexical

Words such as `may`, `might`, `cannot`, `limitation`, or `however` are not intrinsically problematic.

The relevant question is what the wording is doing in that sentence, paragraph, section, and evidence chain.

### Auditable, not invisible

A revision system should record why something was changed—and why something conspicuously cautious was deliberately left unchanged.

For that reason, `KEEP` is a first-class audit outcome.

### Human-authority preserving

Meaning-dependent decisions remain with the author.

When a revision would alter the research question, conceptual hierarchy, evidence interpretation, citation role, substantive limitation, or other authorial commitment, the correct output is `QUERY`, not improvisation.

---

## A compact formulation

The skill can be summarized as:

```text
Diagnose rhetoric
        ↓
Identify evidentiary function
        ↓
Surface the supported contribution
        ↓
Preserve the claim ceiling
        ↓
Preserve evidence status and scope
        ↓
Regression-test the manuscript
```

Or, more simply:

> **Say clearly what the evidence supports. Keep exactly the caution needed to show what it does not.**

---

## Responsible use

This project supports experimental, human-controlled academic revision. It does not guarantee manuscript quality, factual accuracy, publishability, or compliance with journal or institutional policy.

Users remain responsible for:

* verifying facts, data, citations, quotations, interpretations, and substantive claims;
* reviewing and approving every AI-assisted revision;
* ensuring that no data, sources, procedures, or findings are fabricated or misrepresented;
* following current journal, institutional, funder, and professional rules for generative AI use and disclosure; and
* protecting unpublished research materials, interview data, personal information, and confidential documents.

The skill is designed to assist bounded, auditable revision—not to replace scholarly judgment, authorship, or accountability.

---

## Limitations

This skill cannot determine whether the underlying research is empirically correct.

It does not independently verify citations, reconstruct missing evidence, resolve contradictory source material, or establish whether a substantive claim is true.

Its integrity mechanisms depend on the manuscript, source-status information, author decisions, and other materials supplied to the revision environment.

The bundled cue scanner is deliberately conservative and incomplete. It should be treated as a retrieval aid, not as a measure of manuscript quality.

No raw cue count should be interpreted as an editing target or a before/after quality score.

---

## Contributing

Contributions are welcome, especially those that improve:

* defensive-writing candidate taxonomies;
* false-positive handling;
* multilingual cue detection;
* evidence-status classification;
* claim-delta regression;
* auditability;
* project-specific guardrail patterns;
* integration with tracked-change manuscript workflows.

New automated rules should follow one principle:

> **Automation may expand candidate discovery, but it must not silently replace contextual evidentiary judgment.**

Changes that encourage blind hedge removal, automatic claim strengthening, global terminology substitution, or evidence-blind rewriting are inconsistent with the purpose of this project.

---

## Acknowledgment

This project is built around a simple premise: academic revision should improve the visibility of a contribution without changing the epistemic contract between the manuscript and its evidence.

The strongest version of a paper is not necessarily the version with the strongest language.

It is the version in which readers can see, as directly as possible, **what the evidence establishes, why it matters, and where its limits actually begin**.
````

## File: scripts/scan_defensive_cues.py
size: 5468
sha: 960d9869dccda7b077c46addd7e22d959c631bd1
upstream: https://github.com/lensback940701/evidence-bound-press-conference-revision-skill/blob/HEAD/scripts/scan_defensive_cues.py

```text
#!/usr/bin/env python3
"""Create a candidate register from a DOCX without making editorial decisions.

The output is a cue list, not a diagnosis. Review every row in paragraph and
source context before assigning KEEP, TIGHTEN, REFRAME, RELOCATE, CUT, or QUERY.
"""

from __future__ import annotations

import argparse
import csv
import re
import zipfile
from pathlib import Path
from xml.etree import ElementTree as ET

W_NS = "{http://schemas.openxmlformats.org/wordprocessingml/2006/main}"

CUE_PATTERNS = {
    "D1_self_deprecation": re.compile(
        r"\b(?:regrettably|unfortunately|sadly|merely|falls?\s+short|fails?\s+to|"
        r"lags?\s+behind|unable\s+to|cannot\s+(?:match|support|show|demonstrate))\b",
        re.IGNORECASE,
    ),
    "D2_preemptive_defence": re.compile(
        r"\b(?:we\s+do\s+not\s+claim|this\s+(?:article|paper|study)\s+does\s+not|"
        r"it\s+should\s+be\s+noted|this\s+is\s+not\s+to\s+say|one\s+might\s+object|"
        r"not\s+intended\s+to|not\s+designed\s+to)\b",
        re.IGNORECASE,
    ),
    "D3_hedge_cluster": re.compile(
        r"\b(?:may|might|could|perhaps|possibly|potentially|appears?\s+to|tends?\s+to|"
        r"suggest(?:s|ed)?)\b(?:[\w,;:\-\s]{0,72})\b(?:may|might|could|perhaps|possibly|"
        r"potentially|appears?\s+to|tends?\s+to|suggest(?:s|ed)?)\b",
        re.IGNORECASE,
    ),
    "D4_work_log": re.compile(
        r"\b(?:we\s+first|initially|then\s+tried|later\s+tried|eventually\s+found|"
        r"after\s+several\s+attempts|经过多次尝试|我们先|后来|最终)\b",
        re.IGNORECASE,
    ),
    "D6_comparison_concession": re.compile(
        r"\b(?:does\s+not\s+outperform|weaker\s+than|outperformed\s+by|"
        r"cannot\s+compete|performs?\s+worse|明显落后|效果有限|未能超过)\b",
        re.IGNORECASE,
    ),
}

def paragraph_text(paragraph: ET.Element) -> str:
    """Return visible Word text, excluding deletion text in tracked documents."""
    return "".join(node.text or "" for node in paragraph.iter(W_NS + "t")).strip()

def paragraph_style(paragraph: ET.Element) -> str:
    ppr = paragraph.find(W_NS + "pPr")
    if ppr is None:
        return ""
    style = ppr.find(W_NS + "pStyle")
    return "" if style is None else style.get(W_NS + "val", "")

def reference_like(text: str) -> str:
    """Flag a likely bibliography entry without suppressing it automatically."""
    compact = " ".join(text.split())
    looks_like_entry = bool(
        re.match(r"^[A-Z][A-Za-z'’\-]+, .{0,180}\b(?:18|19|20)\d{2}\. ", compact)
    )
    return "yes" if looks_like_entry else "no"

def iter_paragraphs(docx_path: Path):
    with zipfile.ZipFile(docx_path) as archive:
        xml = archive.read("word/document.xml")
    root = ET.fromstring(xml)
    for index, paragraph in enumerate(root.iter(W_NS + "p"), start=1):
        text = paragraph_text(paragraph)
        if text:
            yield index, paragraph_style(paragraph), text

def snippet(text: str, start: int, end: int, width: int = 280) -> str:
    left = max(0, start - width // 2)
    right = min(len(text), end + width // 2)
    prefix = "…" if left else ""
    suffix = "…" if right < len(text) else ""
    return prefix + text[left:right].strip() + suffix

def main() -> int:
    parser = argparse.ArgumentParser(
        description="Export defensive-writing cue candidates from a DOCX; no candidate is an edit decision."
    )
    parser.add_argument("docx", type=Path, help="Clean DOCX preferred; tracked DOCX may contain both baseline and insertion text.")
    parser.add_argument("--out", required=True, type=Path, help="Destination CSV path.")
    args = parser.parse_args()

    if args.docx.suffix.lower() != ".docx" or not args.docx.is_file():
        parser.error("docx must name an existing .docx file")
    args.out.parent.mkdir(parents=True, exist_ok=True)

    fields = [
        "issue_id",
        "paragraph_id",
        "word_style",
        "reference_like",
        "pattern_code",
        "matched_cue",
        "candidate_text",
        "provisional_disposition",
        "integrity_note",
        "reviewer_decision",
    ]
    rows = []
    for index, style, text in iter_paragraphs(args.docx):
        for pattern_code, pattern in CUE_PATTERNS.items():
            for match in pattern.finditer(text):
                rows.append(
                    {
                        "issue_id": f"C{len(rows) + 1:03d}",
                        "paragraph_id": f"P{index:03d}",
                        "word_style": style,
                        "reference_like": reference_like(text),
                        "pattern_code": pattern_code,
                        "matched_cue": match.group(0),
                        "candidate_text": snippet(text, match.start(), match.end()),
                        "provisional_disposition": "REVIEW_REQUIRED",
                        "integrity_note": "Cue only; test evidence status, scope, rival role, and claim ceiling.",
                        "reviewer_decision": "",
                    }
                )

    with args.out.open("w", newline="", encoding="utf-8-sig") as handle:
        writer = csv.DictWriter(handle, fieldnames=fields)
        writer.writeheader()
        writer.writerows(rows)

    print(f"Wrote {len(rows)} cue candidates to {args.out}")
    print("No editorial decision was made. Review every candidate in context.")
    return 0

if __name__ == "__main__":
    raise SystemExit(main())
```

## File: references/triage-and-regression.md
size: 4575
sha: 05c4137be5f4cd2008d2305eeaea654db2bba730
upstream: https://github.com/lensback940701/evidence-bound-press-conference-revision-skill/blob/HEAD/references/triage-and-regression.md

```text
# Triage and regression reference

Use this reference before applying the evidence-bound press-conference revision skill. A lexical cue is a prompt for contextual reading, never a decision.

## Candidate taxonomy

| code | candidate pattern | usual action | do not remove when |
| --- | --- | --- | --- |
| D1 | Self-deprecation or apology | CUT or REFRAME | it states a specific material limitation readers need to interpret a result. |
| D2 | Anticipatory defence against an uninvited objection | CUT, TIGHTEN, or RELOCATE | it is the only sentence preventing a demonstrable over-reading. |
| D3 | Stacked hedge | TIGHTEN to the one hedge that carries the real epistemic limit | multiple hedges mark distinct source, scope, and causal limits. |
| D4 | Work-log narration or failed-attempt chronology | REFRAME as problem, gap, solution, and evidence | the sequence itself is causal evidence or part of the empirical account. |
| D5 | Buried contribution | REFRAME or RELOCATE | the surrounding qualification is a required claim boundary. |
| D6 | Unnecessary comparison or volunteered loss | NARROW, REFRAME, or CUT | comparison names the relevant rival explanation or a substantive trade-off. |
| D7 | Broad negative self-judgment | TIGHTEN to its demonstrated scope | the limitation is a documented finding or method boundary. |
| D8 | Caveat in the wrong rhetorical position | RELOCATE | moving it detaches it from the evidence it qualifies. |
| K1 | Scope condition | KEEP or state once precisely | deleting it broadens the claim. |
| K2 | Source-status distinction | KEEP | it separates plan, authorization, report, observation, outcome, or inference. |
| K3 | Rival, contradiction, delay, reversal, or non-adoption | KEEP | it is data, mechanism, or a live alternative explanation. |
| K4 | Ethical, provenance, reproducibility, or method limitation | KEEP | it is needed for honest interpretation. |

## Three question test

For every candidate, answer in writing:

1. **What function does this wording serve?** Identify rhetorical self-defence, evidence status, scope, method, rival explanation, empirical finding, or more than one.
2. **What would disappear if it were removed?** Name the lost proposition or boundary, not merely a change in tone.
3. **Does the replacement preserve the ceiling?** Confirm that it does not increase certainty, causality, novelty, generality, representativeness, or portability.

If Questions 1–3 cannot be answered confidently, use `QUERY`.

## Correct placement of necessary caution

Keep the caution, but place it where it does its analytical work.

- Keep a source-status distinction next to the source-dependent claim.
- State a general method or scope boundary once in the relevant methods, framework, or discussion location; do not repeat it as an apology in every empirical paragraph.
- Keep a rival explanation next to the inference it qualifies.
- Lead a paragraph with the supported proposition when possible, then place the limit immediately after it rather than replacing the proposition with the limit.

## Regression protocol

Run all checks after each revision batch and again after the full pass.

| check | test | failure response |
| --- | --- | --- |
| Claim delta | The new verb and modifier do not exceed the original supported claim. | Revert or query. |
| Evidence status | Plan, authorization, report, observation, outcome, interpretation, and causation remain separate. | Restore the distinction. |
| Citation role | A citation, quotation, date, number, and source-specific wording keep their original support role. | Revert or route to evidence review. |
| Scope | The case, time, population, and comparison boundary remain visible where needed. | Restore or relocate the boundary. |
| Hierarchy | The primary contribution remains primary and secondary lenses remain conditional. | Recast without adding concepts. |
| Section role | Abstract, introduction, framework, evidence, discussion, and conclusion make complementary rather than duplicative moves. | Flag for authorized consistency revision. |
| Voice | The prose is direct and readable without sounding promotional, formulaic, or copied from a comparator. | Tighten locally. |

## Candidate-register interpretation

Use `KEEP` as a positive audit result. A high number of retained cautions is not a failed press-conference revision; it may show that the manuscript has disciplined evidence writing. Assess the concentration of D1–D8 patterns, not the raw number of cue words.

Do not use raw counts as an edit quota or a proof of improvement.
```

## File: references/urban-geography-guardrails.md
size: 2619
sha: 888174ca5951e2cb7601524d6cf335a7a3a27fb0
upstream: https://github.com/lensback940701/evidence-bound-press-conference-revision-skill/blob/HEAD/references/urban-geography-guardrails.md

```text
# Dalian Urban Geography guardrails

Use this reference only for the returned Dalian manuscript package. It turns the generic press-conference method into a safe, project-specific audit.

## Fixed hierarchy

1. **Heritage metabolism** is the primary middle-range, operation-centred analytic.
2. **Heritage operations** connect inherited places, institutional conditions, interventions, and differentiated effects.
3. **Rescaling** is a supporting, rival-open capacity lens. It matters only where a change in authority, resources, legitimacy, instruments, or intervention arena bears on the capacity to carry out a specific operation.
4. **Capacity → operation → outcome** is a test, not a presumption of delivery.
5. Configuration, operation, outcome, contradiction, and feedback remain distinct. Coupling, lag, decoupling, reversal, and recombination are empirical possibilities, not stylistic clutter.

## Evidence firewall

Do not collapse the following distinctions for fluency or rhetorical force:

- proposal versus implementation;
- authorization versus delivery;
- reported work versus direct observation;
- material persistence versus stable meaning or use;
- sequence versus mechanism;
- a documented episode versus citywide prevalence;
- a portable protocol versus a universal model.

Do not add, remove, move, or reinterpret citations, quotations, dates, numbers, proper names, footnotes, tables, figures, captions, source notes, or cross-references. Do not turn the three overlapping configurations into stages or a universal chronology.

## Terminology and rescaling gate

Treat `governance` and `governing` as terminology candidates, not as defensive-writing cues. They may be source-specific or analytically required in a specific sentence, but must not be elevated into a master concept or inserted by global replacement.

Do not describe rescaling as something the paper set out to find, evaluate, prove, or make primary. Preserve the supporting, rival-open framing and active alternatives.

## Current integration risk

The returned clean and tracked manuscripts contain an author-authorized general-first reframe in the abstract and introduction. The framework, evidence, empirical, discussion, and conclusion sections remain substantially Dalian-specific. Diagnose cross-section alignment before changing it. A body-wide consistency pass is a substantive, author-authorized task; it is not a by-product of language polishing.

Retain the unresolved terminology decision concerning `colonial remainders` versus the manuscript's defined terms unless the author supplies the intended distinction.
```

## File: assets/defensive-writing-audit-template.csv
size: 263
sha: d095d6c0cdb93b62b58b24d3c7f6b09de1097750
upstream: https://github.com/lensback940701/evidence-bound-press-conference-revision-skill/blob/HEAD/assets/defensive-writing-audit-template.csv

```text
issue_id,paragraph_id,section,pattern_code,cue_or_trigger,candidate_text,rhetorical_function,evidentiary_function,evidence_status,contribution_relation,provisional_disposition,claim_delta_check,citation_or_source_status_check,author_decision_needed,reviewer_note
```

## File: assets/readme/audit-loop.png
size: 777443
sha: 869141366e9e6e939a2c340c9de2f557b79e8435
upstream: https://github.com/lensback940701/evidence-bound-press-conference-revision-skill/blob/HEAD/assets/readme/audit-loop.png

Content is not inlined in the catalog. Read the upstream file: https://github.com/lensback940701/evidence-bound-press-conference-revision-skill/blob/HEAD/assets/readme/audit-loop.png

## File: assets/readme/hero.png
size: 1843900
sha: 74db7e76a9247984fc3fc09830062cddca47ad80
upstream: https://github.com/lensback940701/evidence-bound-press-conference-revision-skill/blob/HEAD/assets/readme/hero.png

Content is not inlined in the catalog. Read the upstream file: https://github.com/lensback940701/evidence-bound-press-conference-revision-skill/blob/HEAD/assets/readme/hero.png

## File: assets/readme/principle.png
size: 1071790
sha: f94200dbda3ed4381085259ecc599c63a29e4cc1
upstream: https://github.com/lensback940701/evidence-bound-press-conference-revision-skill/blob/HEAD/assets/readme/principle.png

Content is not inlined in the catalog. Read the upstream file: https://github.com/lensback940701/evidence-bound-press-conference-revision-skill/blob/HEAD/assets/readme/principle.png

## File: agents/openai.yaml
size: 293
sha: 059b92b78ffe3629be2aed8e2d14832909615f32
upstream: https://github.com/lensback940701/evidence-bound-press-conference-revision-skill/blob/HEAD/agents/openai.yaml

```text
interface:
  display_name: "Evidence-Bound Press Revision"
  short_description: "Contribution-forward editing without overclaiming"
  default_prompt: "Use $press-conference-revision-evidence-bound to batch-audit this manuscript for defensive writing while preserving all evidence limits."
```

## File: LICENSE
size: 1071
sha: 15856b49f3d81bd54757bef784e27a71abfe73a5
upstream: https://github.com/lensback940701/evidence-bound-press-conference-revision-skill/blob/HEAD/LICENSE

```text
MIT License

Copyright (c) 2026 lensback940701

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
```

## File: README.zh-CN.md
size: 18289
sha: 0f37aa8e35d704ada7a8d2ca5ccbffd1feb4b5f2
upstream: https://github.com/lensback940701/evidence-bound-press-conference-revision-skill/blob/HEAD/README.zh-CN.md

````text
# 证据约束型“发布会式”论文修订

[English](README.md)

**让有证据支撑的贡献先被看见，同时坚决避免过度主张。**

本项目由小红书 [@Lensback's Lab](https://www.xiaohongshu.com/user/profile/64507c470000000012036d99) 与 [@鸣人学AI](https://www.xiaohongshu.com/user/profile/666940c40000000007005b9a) 共同创作。

<p align="center">
  <img src="assets/readme/hero.png" alt="证据约束型发布会式论文修订" width="900">
</p>

`press-conference-revision-evidence-bound` 是一个用于识别和减少学术论文中**防御性写作**的修订 Skill。它帮助作者把已经存在、且已经获得证据支持的贡献表达得更直接，同时保留决定学术主张边界的证据状态、适用范围、竞争性解释、矛盾材料、方法限制与引文功能。

它并不通过夸大结论让论文显得更强势。它要改变的是贡献与限定的**修辞位置**，而不是论文与证据之间的认识论契约。

> **核心原则：减少修辞性防御，不减少必要的学术审慎。**

---

## 为什么叫“发布会式修订”？

论文在起草、合作者修改、同行评议和反复预判批评的过程中，常会不断累积防御性表达。最终的文本可能不再像一篇清晰提出论点的论文，而更像作者所有担忧和修改过程的工作日志：

* 核心贡献要经过多层限定才出现；
* 同一局限在不同章节反复说明；
* 段落先花大量篇幅解释论文“不主张什么”；
* 方法上的审慎逐渐变成自我贬低；
* 已经放弃的分析路径仍以过程叙述留在正文；
* 不必要地拿个案与更强案例或更大文献比较；
* 回答了论文自身主张并未引出的假想质疑。

“发布会式”这个比喻只提出一个问题：

> 当证据保持不变时，论文应该先说清楚自己究竟建立了什么？

好的修订不会遮掩弱点，也不会制造确定性。它让得到支持的命题先出现，再把必要限定放到真正承担分析功能的位置。

目标是：

**贡献 → 证据 → 限定**

而不是：

**限定 → 自我辩护 → 更多限定 → 最后才出现贡献**

---

## 这个 Skill 保护什么？

所有修改都必须维持或缩窄原稿的认识论承诺。

### 主张上限

修改不能让主张在确定性、因果性、创新性、普遍性、代表性或可迁移性上超过现有证据允许的程度。

`可能影响` 不能悄悄变成 `决定`。

`访谈材料表明` 不能变成 `证据证明`。

`在本案例中` 不能变成 `城市通常都会`。

### 证据状态

工作流明确区分：

* 提议；
* 授权；
* 被报告的活动；
* 直接观察；
* 实施；
* 结果；
* 解释；
* 因果推断。

语言更加流畅，不能成为合并这些状态的理由。

### 适用范围

个案、时间、人群、制度、地理与比较边界必须在影响主张解释的位置保持可见。

### 竞争性解释与负面证据

如果矛盾、延误、未采纳、脱耦、逆转、实施失败或竞争性解释属于实证记录，它们就不是需要隐藏的“尴尬残余”。

### 引文与引用功能

风格修改不能悄悄改变一条引文、一段直接引语、日期、数字、来源专用术语、表格、图片或脚注原本被要求支持的命题。

---

## 两种工作模式

### 1. 只读诊断

只审计稿件，不修改正文。典型产出包括：

* **贡献与范围契约**；
* 全稿候选项登记表；
* 对必要审慎表达作出的显式 `KEEP` 判断；
* 分章节的防御性表达分布图；
* 例外与待作者判断事项；
* 后续修订的有限建议。

这一模式适合作者先判断：防御性写作是否真的构成全稿层面的问题。

### 2. 经授权的集中修订

只有在修改权限明确后，Skill 才能在不移动证据上限的前提下开展集中修订。典型产出包括：

* 带修订痕迹的稿件；
* 清洁稿；
* 按修辞模式组织的修改摘要；
* 主张—证据回归检查报告；
* 有意不修改的内容记录；
* 必须由作者处理的未决事项。

这一区分很重要：**完成诊断并不自动获得改写授权。**

---

## 工作流

```mermaid
flowchart LR
    A["阶段 0<br/>贡献与范围契约"] --> B["阶段 1<br/>全稿批量诊断"]
    B --> C{"作者是否授权<br/>修改正文？"}
    C -- 否 --> D["诊断登记表<br/>待决事项 + 有限建议"]
    C -- 是 --> E["阶段 2<br/>集中修订"]
    E --> F["阶段 3<br/>全稿回归检查"]
    F --> G["修订稿 + 清洁稿<br/>完整性报告"]
```

### 阶段 0——建立贡献与范围契约

正式修改前，先建立一份紧凑的契约：

| 字段 | 作用 |
| --- | --- |
| **核心贡献** | 用一句受证据约束的话说明论文最主要的分析动作与学术收益 |
| **贡献层级** | 主概念、辅助视角、有限的实证贡献，以及可能存在的可迁移收益 |
| **主张上限** | 当前稿件与证据已经允许的最强表述 |
| **承重型限定** | 修改后仍必须保留的范围、证据状态、方法、竞争性解释、矛盾与伦理限定 |
| **修改权限** | 可修改与锁定的章节、追踪要求及禁止事项 |
| **未决问题** | 不能由代理擅自决定的意义依赖型问题 |

如果贡献层级、证据上限或修改权限无法明确，工作流应停在诊断阶段，而不是自行补造答案。

---

### 阶段 1——全稿批量诊断

首先把论文作为一个整体阅读。Skill 不进行盲目的关键词查找替换；词汇线索只用于定位需要结合上下文判断的位置。

候选项使用以下分类：

| 代码 | 模式 | 常见处理 |
| --- | --- | --- |
| **D1** | 自我贬低或道歉 | 删除或重构 |
| **D2** | 预防性辩护 | 删除、压缩或移动 |
| **D3** | 叠加式模糊限定 | 只保留认识论上必要的限定 |
| **D4** | 工作日志式叙述 | 围绕问题、解决方式和证据重构 |
| **D5** | 被掩埋的贡献 | 重构或前移 |
| **D6** | 不必要比较或主动示弱 | 缩窄、重构或删除 |
| **D7** | 超出证据范围的负面自我判断 | 缩窄到已经展示的范围 |
| **D8** | 放错修辞位置的必要限定 | 移动 |
| **K1** | 适用范围条件 | 保留 |
| **K2** | 来源或证据状态区分 | 保留 |
| **K3** | 竞争解释、矛盾、延误、逆转或未采纳 | 保留 |
| **K4** | 伦理、来源、可复现性或方法限制 | 保留 |

`K` 类别必须在审计中明确可见。

大量 `KEEP` 并不表示修订失败。它可能说明稿件原本就包含纪律严明、受证据约束的必要限定。

判断单位是**一句话在具体上下文中的功能**，不是某个单词是否出现。

---

## 修改前的三个问题

每个候选项在被修改前，都必须回答：

1. **这段表达承担什么功能？**
   它是修辞性自我防御、证据状态、适用范围、方法说明、竞争性解释、实证结果，还是同时承担多种功能？

2. **如果删掉它，什么会消失？**
   答案必须指出会丢失的命题或分析边界，而不能只说语气变得不同。

3. **替代表达是否保持主张上限？**
   新表述不能提高确定性、因果性、创新性、普遍性、代表性或可迁移性。

无法可靠回答时，正确处理是 `QUERY`。

---

### 阶段 2——集中修订

经授权的修订应按照**修辞模式与章节功能**成组处理，而不是全局删除某些词。

通常优先处理：

1. 摘要；
2. 引言；
3. 讨论；
4. 结论；
5. 理论定位；
6. 方法；
7. 防御性表达确实妨碍理解的实证章节。

每个主要章节应当先让贡献句可见，再出现必要限定。限定不能消失，但应被放在真正承担分析功能的位置。

例如：

**防御性结构**

> 尽管本研究由于个案设计存在若干局限，也无法建立具有普遍适用性的因果关系，但研究发现或许仍可初步提示……

**贡献前置结构**

> 本案例表明，在这些制度条件下，X 通过 Y 发挥作用。这一判断仅适用于现有记录所覆盖的案例，并不建立一般性的因果关系。

第二种表述在证据意义上没有变强，只是先说清楚**什么得到了支持**，再说明**边界从哪里开始**。

---

### 阶段 3——全稿回归检查

修订后必须独立进行全稿检查。这不是普通校对，而是为了发现修辞改善是否意外改变了论文的认识论结构。

重点检查：

* **主张变化**——没有证据膨胀；
* **证据状态**——计划、报告、观察、结果、解释与因果仍然可区分；
* **引文功能**——来源仍支持其原本承担的命题；
* **适用范围**——必要边界仍然可见；
* **概念层级**——辅助概念没有意外变成主概念；
* **章节功能**——不同章节继续承担互补的分析任务；
* **写作声音**——表达更直接，但没有变成宣传性语言。

---

## 六种处理结果

| 处理结果 | 含义 |
| --- | --- |
| `KEEP` | 原表述承担必要的证据或分析功能 |
| `TIGHTEN` | 功能必要，但表达重复 |
| `REFRAME` | 命题应保留，但当前修辞结构过度防御 |
| `RELOCATE` | 必要信息所在位置遮蔽了贡献 |
| `CUT` | 可以删除修辞，而不删除命题、证据边界或引文功能 |
| `QUERY` | 处理方式依赖作者意图或证据判断 |

因此，`CUT` 的门槛被有意设置得很高：

> 可以删除修辞，不能删除证据。

---

## 自动线索扫描器

仓库提供一个轻量级 DOCX 扫描器：

```text
scripts/scan_defensive_cues.py
```

它可以从清洁版 `.docx` 中生成初步候选列表：

```bash
python scripts/scan_defensive_cues.py manuscript.docx \
  --out defensive-writing-candidates.csv
```

扫描器可定位与自我贬低、预防性辩护、模糊限定叠加、工作日志叙述以及不必要比较相关的部分词汇模式。

**它不是自动编辑器，也不是修辞分类器。** 命中结果只表示需要结合上下文检查。脚本有意输出：

```text
REVIEW_REQUIRED
```

而不会自动决定 `KEEP`、`TIGHTEN`、`REFRAME`、`RELOCATE`、`CUT` 或 `QUERY`。

> 词汇检测可以自动化，认识论判断必须保留上下文。

---

## 审计登记表

仓库中的：

```text
assets/defensive-writing-audit-template.csv
```

用于记录候选项位置、模式代码、原文、修辞功能、证据功能、证据状态、与核心贡献的关系、初步处理、主张变化检查、引文或来源状态检查、是否需要作者决定以及审阅说明。

这使修改决策可以被检查，而不是让完整性判断停留在代理内部。

---

## 仓库结构

```text
press-conference-revision-evidence-bound/
├── README.md
├── README.zh-CN.md
├── SKILL.md
├── agents/
│   └── openai.yaml
├── assets/
│   ├── defensive-writing-audit-template.csv
│   └── readme/
├── references/
│   ├── triage-and-regression.md
│   └── urban-geography-guardrails.md
└── scripts/
    └── scan_defensive_cues.py
```

* `SKILL.md`：核心工作流、权限顺序、保留测试、产出与停止规则。
* `references/triage-and-regression.md`：D1–D8、K1–K4 分类、三个问题测试与回归协议。
* `scripts/scan_defensive_cues.py`：只生成候选位置，不作编辑判断。
* `assets/defensive-writing-audit-template.csv`：结构化候选项审计表。
* `references/urban-geography-guardrails.md`：展示如何为通用方法补充项目专用的概念与证据边界；它不是一般使用的必需文件。

核心设计可以概括为：

```text
通用修订协议
        +
项目专用证据与概念护栏
        =
有边界的论文修订
```

---

## 安装

将仓库克隆到 Codex 的 skills 目录：

```bash
git clone https://github.com/lensback940701/Evidence-Bound-Press-Conference-Revision-Skill.git \
  ~/.codex/skills/press-conference-revision-evidence-bound
```

Windows 的常用目录为：

```text
%USERPROFILE%\.codex\skills\press-conference-revision-evidence-bound
```

安装后重启或刷新 Codex。调用名称为 `$press-conference-revision-evidence-bound`。

---

## 适合何时使用？

特别适合以下情况：

* 稿件已经经历多轮修订；
* 回复审稿意见后，正文越来越像防御性答辩；
* 贡献确实存在，但读者很难找到；
* 引言用太多篇幅预判批评；
* 同一限定在多个章节反复出现；
* 论文不断解释自己“不主张什么”；
* 质性或历史材料要求严格区分来源状态；
* 竞争性解释必须继续可见；
* 合作者希望语言更有力，但证据上限不能移动；
* AI 辅助修改需要显式的完整性护栏；
* 与自由改写相比，更需要可追踪、可审计的修改。

它尤其适合质性、历史、解释性、案例型与混合方法研究，因为这些论文中的细小修辞变化也可能改变证据看起来能够建立什么。

---

## 不适合哪些任务？

本 Skill 不用于：

* 从零起草论文；
* 文献检索；
* 新的实证研究；
* 引文发现或引文修复；
* 事实核查；
* 接纳新证据；
* 一般性语法修改；
* 单纯复制编辑；
* 未经作者授权重构论文；
* 发明更强的贡献主张；
* 让结果显得比证据允许的更因果、更普遍、更新颖或更具代表性。

如果根本问题是证据不足，就不应使用本 Skill 掩饰这一缺陷。

---

## 调用示例

### 全稿只读诊断

```text
请使用 $press-conference-revision-evidence-bound，
对这篇完整稿件开展防御性学术写作的只读审计。

暂时不要修改正文。

先建立贡献与范围契约，再使用 D1–D8 / K1–K4 分类候选项。
请把必要的学术审慎明确标为 KEEP，并指出所有必须获得作者授权的修改。
```

### 经授权的修订

```text
请使用 $press-conference-revision-evidence-bound，
修改这篇稿件中已经获得授权的部分。

让得到支持的贡献更容易被看见，并减少修辞性防御，
但必须保留现有主张上限、证据状态、适用范围、竞争性解释、
引文、直接引语和概念层级。

修改后执行完整回归检查，并报告有意不修改的内容与未决作者事项。
```

### 配合自动扫描器

```text
只把仓库中的防御性写作线索扫描器用于生成候选项。
为每一个命中段落读取上下文后再决定处理方式。

不得把词汇频次视为写作质量证据，也不得执行全局自动替换。
```

---

## 设计理念

### 贡献前置，而非宣传性写作

目标不是让论文为了显得自信而使用更强语言。句子之所以“更强”，应当是因为它的修辞位置更清晰，而不是因为证据状态被升级。

### 证据约束，而非最大限度谨慎

限定词越多并不自动意味着越严谨。重复限定可能反而掩盖证据真正支持的内容。目标是保留**最小充分的认识论限定**，而不是堆叠最大数量的防御性表达。

### 结合上下文，而非按词判断

`may`、`might`、`cannot`、`limitation` 或 `however` 本身都不是问题。真正要判断的是它们在句子、段落、章节和证据链中承担什么功能。

### 可审计，而非不可见

修订系统既应记录为什么修改，也应记录为什么有意保留某句显得谨慎的表达。因此，`KEEP` 是正式的审计结果。

### 保留作者权限

涉及研究问题、概念层级、证据解释、引文功能、实质性局限或其他作者承诺的判断，必须交回作者。正确结果是 `QUERY`，不是代理擅自补全。

---

## 一句话概括

```text
诊断修辞
   ↓
识别证据功能
   ↓
让得到支持的贡献可见
   ↓
保留主张上限
   ↓
保留证据状态与适用范围
   ↓
对全稿进行回归检查
```

更简单地说：

> **清楚说出证据支持什么，并保留恰好足够的审慎，说明证据不支持什么。**

---

## 负责任使用

本项目用于实验性的、由人控制的学术修订。它不保证稿件质量、事实准确性、可发表性，也不保证符合期刊或机构政策。

使用者仍须负责：

* 核验事实、数据、引文、直接引语、解释与实质主张；
* 阅读并批准每一项 AI 辅助修改；
* 确保不虚构或歪曲数据、来源、研究程序与发现；
* 遵守期刊、机构、资助者和专业组织关于生成式 AI 使用与披露的现行规定；
* 保护未公开研究材料、访谈数据、个人信息与机密文件。

本 Skill 用于辅助有边界、可审计的修订，不能替代学术判断、作者身份与作者责任。

---

## 局限

本 Skill 无法判断底层研究在实证上是否正确，也不会独立核实引文、重建缺失证据、解决来源之间的矛盾，或证明某项实质主张为真。

其完整性机制依赖于修订环境中提供的稿件、证据状态信息、作者决定与其他材料。

线索扫描器有意保持保守且并不完备。它只能作为检索辅助，不能作为稿件质量指标。任何原始命中数量都不应被当作编辑目标或修改前后的质量分数。

---

## 参与贡献

欢迎改进：

* 防御性写作候选项分类；
* 假阳性处理；
* 多语言线索识别；
* 证据状态分类；
* 主张变化回归检查；
* 审计能力；
* 项目专用护栏模式；
* 与修订痕迹工作流的整合。

新增自动规则必须遵守：

> **自动化可以扩大候选项发现范围，但不能悄悄取代结合上下文的证据判断。**

鼓励盲目删除限定、自动强化主张、全局替换术语或脱离证据改写的修改，与本项目目的不符。

---

## 致谢

本项目建立在一个简单前提上：学术修订应当提高贡献的可见性，而不能改变论文与证据之间的认识论契约。

一篇论文最好的版本，不一定是语言最强势的版本。

真正重要的是让读者尽可能直接地看见：**证据建立了什么、为什么重要，以及它的边界究竟从哪里开始。**
````
