v0.1.66 to git:20261117.5d6cc46
8 added, 10 removed. Audit A to A.
What this update changes in the checks
No check changed: every fixed check gives the same answer before and after this update.
A tripped check is a label, not a verdict: read the matched text and the lines below. This week's updates that newly trip a check.
---
name: evidence-backed-analytical-writing
description: >-
Write an evidence-backed analytical report — Chinese policy briefs, capability assessments, or
technical arguments that combine supplied claims with retrieved sources — by classifying every
load-bearing number, building an evidence map, and drafting a uniform claim → evidence → mechanism
→ limitation spine. Use when the deliverable is prose whose every figure must survive an audit.
- Not for citation mechanics (that is grounded-citations), social posts, changelogs, or any piece
+ Not for citation-ledger mechanics, social posts, changelogs, or any piece
where the claims are already verified and only the wording needs work.
metadata:
- version: "0.1.66"
+ revision: "1"
areas: "work-management"
targets: "runtime-agnostic"
summary: "evidence map plus a claim → evidence → mechanism → limitation spine for Chinese analytical reports; recency window, status ladder, negative findings"
summary_zh: "中文分析报告:证据地图与「主张 → 证据 → 机制 → 局限」骨架;时效边界、状态阶梯、否定性结论"
kind: "entry"
---
# Evidence-backed analytical writing
Use this skill for Chinese analytical reports, policy briefs, capability assessments,
and technical arguments that combine user-supplied claims with retrieved evidence.
- It complements `grounded-citations`: this skill governs claim hygiene, structure,
- and user-facing prose; `grounded-citations` governs the citation ledger and source
- rendering.
+ This skill governs claim hygiene, structure, and user-facing prose; it leaves the
+ citation-ledger tooling to whatever the operator uses.
## Procedure
1. **Normalize the brief.** Extract the requested language, approximate length,
section count, section symmetry, audience, required claims, and requested
source recency. Preserve the requested point order and visible form: when the
input uses “一是/二是/三是” analytical paragraphs, keep that numbered
structure unless the user asks for a different format. Do not preserve an
unsupported number as fact. Locate any project style contract (`style/*.md`) before
drafting and copy its bans into the QA checklist — contracts may forbid direct “我军”
address (use 军队/相关力量) or require “要/应” imperatives in recommendation sections;
contract wording overrides a prior draft's convention.
2. **Set the recency boundary before searching.** When the user asks for
“最新/近期” data, establish a cutoff date, search current official releases
first, and label older figures as trend baselines rather than current status.
Prefer current summary-level statistics over a pile of newly published case
stories; a newer anecdote is not automatically better evidence for a fleet-
or system-level conclusion.
A relevance-ranked search tool does **not** honour a recency window on its own: it
will return pre-window material for a query that merely reads as “recent”. Put the
window in the query text, forbid the fallback explicitly (“if nothing exists after
<date>, say exactly that — do not fill the gap with older material”), then audit the
returned source list and drop every item below the window floor before it reaches the
evidence map. When a round leaks, keep it as a trace and re-run those routes — only
the enforced round is admissible.
A model-backed search tool may also answer from its own knowledge **without searching
at all**; such an answer reads confident and well-structured. An answer with an empty
source list is a lead, not evidence — never register or cite it.
3. **Classify every load-bearing claim before searching.** Mark each number as
inventory, procurement, project target, type count, prototype/test result,
or analytical estimate. Record scope, date, unit, and whether it describes a
fielded capability or a plan. Use `references/claim-audit.md` for the wording
decision.
4. **Build an evidence map before drafting.** For each important claim, record
`claim → source → exact figure/date/scope → safe interpretation`. Prefer
official government, audit, program, and primary technical sources. Use
secondary reporting only for context or when the primary source is unavailable;
label the limitation instead of laundering it into certainty.
5. **Resolve incompatible denominators.** Do not add or compare procurement
quantities, fielded quantities, model counts, prototypes, and cumulative
historical purchases unless the source defines a common denominator. When a
supplied statistic cannot be independently supported, say so briefly and
replace it with a sourced trend or metric.
6. **Draft each section with the same spine:** claim → measured evidence → mechanism or operational meaning → limitation → implication. Make the causal link explicit: unmanned-equipment application → change in personnel, process, visibility, or reach → resulting logistics characteristic. For a balanced multi-point brief, preserve the user's visible numbering and give each section comparable length and depth, normally within roughly 10–15%. Prefer a small set of fleet-, program-, or system-level summary metrics over many anecdotes; when the user asks for summary data and few examples, avoid platform-specific narratives unless the measured result is load-bearing, and do not let the final section become a case-study or data dump.
7. **Use precise status language.** Say “指标/计划/演示/试验/原型/已列装”
exactly as supported. Separate deployed capability from a research demonstration
and from a future aspiration. Convert absolutes such as “无死角” or “完全自主”
into bounded claims about reach, exposure, endurance, or resilience. Rank every claim on the
evidence-status ladder — 需求/概念 → 合同 → 试验/演示 → 交付/列装 → 有限使用 → 独立测量结果 —
and never let one tier's evidence carry the next tier's conclusion (see
`references/claim-audit.md`).
8. **Cite while drafting, not after.** Put inline citations immediately after the
sentence carrying the figure or externally sourced claim. Register every source
in the citation ledger before using its number, then render the Sources block
- mechanically — `grounded-citations/scripts/sources.py` owns both.
+ mechanically from that ledger, never by hand.
9. **Run a source-to-claim audit after edits.** After adding, removing, or
reordering sources, inspect each citation number's title/topic against the
sentence it supports. For an independent draft, reset the citation ledger
before registering sources; for a continuation, preserve the existing ledger
- ids. Render the Sources block with `sources.py render --replace-in <draft>`
- before running `sources.py verify --strict`; stale registered sources can make
- strict verification fail even when the prose citations are correct. Ledger
- validation proves that an id exists; it does not prove that the page supports
+ ids. Re-render the Sources block from the ledger, then run the ledger's strict
+ verification; stale registered sources can make strict verification fail
+ even when the prose citations are correct. Ledger validation proves that an id exists; it does not prove that the page supports
that claim.
10. **Deliver directly usable prose.** Default to concise Simplified Chinese,
formal but readable report style, short paragraphs, balanced numbered sections,
and a short methodology/coverage caveat when the input contains shaky figures.
Put long source lists after the body; do not bury the requested deliverable
under research narration.
## Negative findings and evidence backfill
When a question is answered by the *absence* of evidence, the negative finding is
itself a claim and must be auditable:
- Record it as a search-type conclusion (检索型结论): state the search window and the
inclusion criteria — "在 <窗口> 公开材料中未见案例" — never as a capability
statement in either direction.
- Attach an exclusion table: every item that superficially matches but fails the
criteria, one row each with the disqualifying reason (enterprise/civilian source,
theory article, modeling study, threat-role appearance, wrong service or echelon,
foreign case). Without it, a later session rediscovers one of those items and
cites it as evidence for the very claim the search refuted.
- Never substitute a foreign, theoretical, civilian, or adjacent-service example to
fill a domestic gap; the gap is the finding.
Incremental re-verification (backfill) after an evidence packet or report is frozen:
- Do not mutate the frozen packet/report that already produced drafts; append an
addendum and extend the verification log with a new dated batch, then state
explicitly whether any existing claim was invalidated and whether the delivered
report needs revision — "no change needed" is a valid recorded outcome.
- For each re-verified candidate, fetch the original full text and quote the exact
phrase before logging it; fix provenance — an aggregator or reprint (a portal
carrying another outlet's story) is not the original publisher.
- Some candidates exist only as a **bibliographic record** — title, authors, journal, volume,
DOI — with the full text behind a paywall, a login wall, or a thesis catalogue. Log those as
bibliography-level verification and say so in the batch, and forbid citing their findings: the
record proves the paper exists, not what it concluded. Marking the tier keeps them usable as
leads without letting a later draft borrow a conclusion nobody read.
## Final-report sourcing style (简单出处)
Final reports can use light sourcing by design: name the outlet inline in the prose
(《解放军报》、中国军网、国防部网 …) and close with a single `来源:` line naming the outlets and
the year window. No numbered reference list and no `[n]` markers in the final deliverable —
numbered citation machinery belongs to working drafts and process records only. When a numbered
draft is converted to this style, invert the QA checks: verify zero `[n]` residue, a present source
line, and a complete closing sentence, instead of keeping a marker-count gate.
## Multi-model draft iterations
Long-form reports can run as a multi-model draft line: a frozen packet (context + material +
skeleton × style variants) goes to several writer models; each draft's filename encodes what
produced it; the operator reviews with line-level comments and dictates each iteration.
- **Frozen stays frozen.** An approved final is not edited in place — it is `final v01` and stays
byte-identical. Improvements run on a separate draft line (`drafts/`), and only the operator
promotes a draft to a new final.
- **Name drafts by provenance dimensions:** `draft-vXX-<skeleton>-<style>-m-<model>` — iteration
number, skeleton version, style version, model short name. Record the version map (which code =
which skeleton/style, model short names) in the project's drafts README the first time; never leave
the mapping only in chat.
- **The pattern's dimensions are the scope.** When the operator supplies such a pattern and asks for
drafts, they mean every combination of the stated dimensions; an example model written inside the
pattern is a format sample, not the target set. Generate the full matrix (skipping combinations
that already exist), not one sample file.
- **The loop is comment-driven.** After a draft lands, the operator returns comments (quoted text +
instruction); apply them as the next `vXX+1` for that combo, preserving the prior draft and
receipts. Do not self-iterate or auto-promote to the final line.
- **Reuse the frozen packet; do not re-run completed discovery.** Draft waves build on research that
already exists — redoing established discovery is waste the operator explicitly rejects.
## Always-on quality rules
- Never fabricate a current inventory, percentage, type count, performance figure,
or effectiveness result to make the argument sound complete.
- Put units, time period, platform/status, and source scope beside every important
number; a bare number is not evidence.
- Treat “can theoretically do X” as a capability implication, not proof that X
has been operationally achieved.
- Preserve a clear boundary between fact, inference, and judgment; signal the
boundary in the prose when it affects the conclusion.
- For “约N字,” aim for the requested length within roughly 10–15% unless the
evidence needed to satisfy the brief makes that impossible; state the actual
approximate length if materially over or under.