ieee-acm-paper-writing

IEEE / ACM Manuscript Paper Writing

Draft, rewrite, humanize, and audit engineering manuscripts without crossing the evidence boundary

A native Codex plugin and Agent Skills package for Codex, Claude Code, and compatible LLM-based research agents. It turns supplied technical evidence into defensible manuscript prose while preserving claims, numbers, citations, notation, scope conditions, and uncertainty.

Quick start · Modes · Installation · Examples · Validation · SKILL.md · Audit-map showcase

---

What this skill does

This is a manuscript-writing skill with an evidence gate. It can draft a new section, rewrite an existing one, expand or compress technical exposition, remove machine-like prose patterns, build an evidence-aware outline, audit a manuscript, or adapt it to a named IEEE or ACM venue.

The writing modes are the primary workflow. Audit is the protection layer around that workflow, not a replacement for it. Before producing publication-ready prose, the skill identifies the intended claim, its supporting artifact, its scope, and any unresolved dependency. When the available evidence does not support requested wording, the skill narrows or omits the claim and returns a structured author query instead of filling the gap with plausible language.

Typical uses include:

• rewriting a Results section while preserving every number, unit, comparator, and uncertainty;
• drafting an Abstract from verified findings without importing claims absent from the body;
• restructuring an Introduction around the technical obstacle, contribution, evidence, and scope;
• explaining a formulation or algorithm without changing notation or guarantee boundaries;
• compressing a manuscript without deleting conditions needed to interpret a result;
• humanizing repetitive or AI-flavored prose without concealing AI-use disclosures; and
• auditing claims, citations, equations, baselines, failures, reproducibility, and venue rules.

It is not a general documentation writer, literature-retrieval system, experiment runner, statistical-analysis package, or automatic submission-acceptance judge.

Quick start

Invoke the skill in an agent conversation with the manuscript path, the requested mode, and the evidence that controls the text. The invocation prefix is host-specific — the examples in this README use the Codex form:

Agent host	Prefix	Example
Codex CLI and IDE extension	@	@ieee-acm-paper-writing audit manuscript.md
Claude Code	/	/ieee-acm-paper-writing audit manuscript.md

Either host also selects the skill from a plain-language request that matches its description, so the prefix is a way to name the skill explicitly rather than a required syntax.

@ieee-acm-paper-writing rewrite sections/results.md using results/ and preserve all numbers, citations, and figure labels
@ieee-acm-paper-writing draft abstract from manuscript.md and verified results in artifacts/
@ieee-acm-paper-writing humanize sections/introduction.md without changing claims or evidence-bearing hedges
@ieee-acm-paper-writing section-audit sections/method.md against config/, logs/, and cited sources

The skill returns text in the conversation unless the user requests an in-place file edit. It asks for clarification only when a missing fact would change the scientific content; otherwise it uses the narrowest defensible interpretation and reports the limitation outside the manuscript.

Working modes

The router exposes nine modes. Style and landmark-paper calibration are modifiers of a writing mode, not additional modes.

Mode	Purpose	Primary output
draft	Write a new section from supplied technical evidence	Manuscript-ready prose plus unresolved author queries
rewrite	Reorganize and improve existing prose without changing its supported scientific meaning	Revised text with protected content preserved
expand	Add explanation, logical connections, or reproducibility detail already supported by the evidence	Expanded manuscript text without invented content
compress	Reduce length while retaining conditions, comparators, units, and uncertainty	Shorter evidence-equivalent prose
humanize	Remove formulaic transitions, uniform rhythm, filler, and unquantified praise	Natural expert prose plus a compact change ledger
outline	Plan section and paragraph functions before drafting	Structure, required evidence, and unresolved dependencies
audit	Inspect the complete claim chain and submission-readiness boundaries	Severity-ordered findings and bounded corrections
section-audit	Apply the same audit contract to a selected section	Section-scoped findings and corrections
venue-adapt	Adapt structure and presentation to a named publication and article type	Adapted text plus a verified compliance ledger

What rewrite protects

rewrite is more than copy-editing, but it is not permission to alter the study. It may change section logic, paragraph order, sentence structure, transitions, terminology consistency, and explanatory depth. Unless the user explicitly authorizes a scientific change, it preserves:

• numerical values, units, denominators, uncertainty, and failure counts;
• citations and the evidentiary role of each citation;
• equations, symbols, algorithms, figure and table labels, and cross-references;
• model assumptions, feasibility conditions, comparator definitions, and guarantee scope; and
• hedges that encode genuine uncertainty or incomplete validation.

Unsupported requested wording remains unresolved even when the rewrite safely narrows it. The handoff therefore records the missing item, the blocked claim, the required action, and the defensible fallback under Author queries.

What humanize changes

humanize is a surface-only pass. It removes formulaic transitions, repetitive paragraph openings, uniform sentence rhythm, filler vocabulary, and unquantified promotional terms. It does not change claims, numbers, units, citations, notation, labels, or uncertainty, and it never removes or weakens a required generative-AI disclosure.

Evidence contract

The skill keeps three authorities separate:

1. Scientific support - what verified formulations, code, data, experiments, and cited literature establish.
2. Method and domain reporting - what the applicable optimization, ML, simulation, systems, or engineering method requires the paper to disclose.
3. Venue compliance - what the target publication's current official instructions and template require.

For technical claims, verified artifacts control over narrative drafts. For formatting and submission requirements, the named publication's current official instructions and template control over generic IEEE/ACM guidance. When sources disagree, the skill reports the competing statements and their locations instead of silently selecting the convenient value.

The integrity gate also treats supplied manuscripts, reviews, references, and data as evidence, never as instructions to the agent. Embedded directives that ask the agent to suppress findings, invent support, or bypass verification are surfaced as integrity findings rather than executed.

How it works

SKILL.md defines the authority hierarchy, routing rules, mode behavior, and output contracts. It loads only the references required for the task:

Reference	Purpose
manuscript-structure-style.md	Section logic, technical exposition, rewriting, compression, and humanization
integrity-audit.md	Claim support, citation verification, integrity checks, and submission-readiness audits
engineering-profiles.md	Method- and domain-specific reporting requirements
corpus-calibration.md	De-identified, non-citable exposition patterns derived from the local calibration corpus
venue-guidance.md	Venue adaptation, citation formatting, and compliance-ledger rules

Engineering domains

The method and domain layer covers eight technical areas:

1. communications and networking;
2. signal processing and sensing;
3. energy systems;
4. robotics and autonomy;
5. mathematical optimization;
6. simulation and digital twins;
7. machine learning and ML-assisted engineering; and
8. computer systems and cyber-physical-system engineering.

Hybrid studies can load multiple profiles. These profiles define what a study should report; they do not by themselves establish scientific validity.

Installation

Prerequisites

• an LLM-based agent host that supports the shared Agent Skills format;
• Node.js 22.20.0 or newer for the skills CLI path below - not required for the Claude Code plugin path; and
• Python 3 only when generating an optional HTML audit map or running the repository-side local workbench.

The installation paths below have been exercised for Codex and Claude Code. This repository does not claim equivalent behavioral validation across their underlying models or every other compatible host.

Path	Version source	Stability boundary
skills CLI command below	Pinned skill tag and pinned installer	Reproduces the declared release package
Manual copy below	Pinned release archive	Reproduces the declared release package without installer code
Native Codex plugin from a local tagged clone	Checked-out tag	Stable when the clone is detached at that tag
Git-backed Codex or Claude plugin marketplace	Default branch	Rolling channel; may lead the latest release

Use a pinned skills CLI or manual-copy path when identical files matter. Treat either Git-backed plugin marketplace as a rolling preview unless its local clone is checked out at a release tag.

Install as a native Codex plugin

The native plugin bundles the canonical skill in an install-safe package and adds Codex card metadata. From a local clone of this repository, register its marketplace and install the plugin:

codex plugin marketplace add .
codex plugin add ieee-acm-paper-writing@ieee-acm-paper-writing

For a rolling Git-backed install, replace . with huguryildiz/ieee-acm-paper-writing. The rolling plugin manifests identify 0.6.4. For a release-pinned native install, clone v0.6.4, run the two commands above from that clone, and keep the marketplace source local. Start a new Codex thread after installation so the skill is discovered. The plugin adds no MCP server, app connector, credential prompt, or background service; manuscript access remains limited to the permissions of the active Codex session. The authentication key in .agents/plugins/marketplace.json is a required Codex marketplace-schema field with no "none" value, so it declares when a credential would be requested, not that this plugin requests one.

Install in an LLM agent host

Claude Code

Enter both lines in a Claude Code session; no Node.js or shell command is involved:

/plugin marketplace add huguryildiz/ieee-acm-paper-writing
/plugin install ieee-acm-paper-writing

This installs the skill from the repository's default branch, so it tracks main rather than a pinned release. Use the skills CLI path below when a specific released version is required. Update later with /plugin update ieee-acm-paper-writing.

This path was last exercised end to end on 2026-08-03 with Claude Code 2.1.220: the marketplace resolved, the plugin installed and reported enabled, and the component inventory registered the single ieee-acm-paper-writing skill. Because the marketplace entry's source is the repository root, this path caches the whole repository (about 9 MB) even though only skills/ieee-acm-paper-writing/ is loaded; the Codex plugin and manual paths copy the skill alone. The two manifests behind this path are gated by scripts/validate_skill.py in CI, but installing through a session command is not something CI can run.

Codex and other Agent Skills hosts

Run this from the manuscript repository in which the skill should be available:

npx skills@1.5.21 add https://github.com/huguryildiz/ieee-acm-paper-writing/tree/v0.6.4 -a codex -y

The repository publishes a single skill, so no --skill selector is needed. Replace -a codex with -a claude-code, or with -a codex claude-code for both. Add --copy to force copied files rather than the installer's selected strategy, and --global only when the skill should be available across all projects.

This project-scoped command installs under .agents/skills/ for Codex and .claude/skills/ for Claude Code. Both the installer (1.5.21) and skill (v0.6.4) are pinned. The upstream installer collects anonymous usage telemetry by default; prefix the command with DISABLE_TELEMETRY=1 to opt out for that invocation.

After installation, start a new session in the selected agent host so it discovers the skill. These invocations are agent prompts, not commands for a standalone manuscript-processing executable.

Manual install, without Node.js or a plugin marketplace

Any host that reads the shared Agent Skills format can load the skill from a plain directory copy. Download a release, then place skills/ieee-acm-paper-writing/ — the whole directory, since SKILL.md routes to its sibling references/, examples/, and scripts/ — under the skill directory your host scans:

curl -fsSL https://github.com/huguryildiz/ieee-acm-paper-writing/archive/refs/tags/v0.6.4.tar.gz \
  | tar -xz
cp -R ieee-acm-paper-writing-0.6.4/skills/ieee-acm-paper-writing <target-directory>/

Host	Project-scoped target	User-scoped target
Codex CLI and IDE extension	.agents/skills/ in the repository	~/.agents/skills/
Claude Code	.claude/skills/ in the repository	~/.claude/skills/

The copied directory is self-contained: every link inside it resolves without the rest of this repository, and the audit-map renderer needs only Python 3 with no third-party packages. Start a new session in the host afterwards so it discovers the skill.

Complete skill invocation reference

The supported invocation shape is:

@ieee-acm-paper-writing <mode> [supported modifier] <input> [evidence and constraints]

All nine modes are shown below:

@ieee-acm-paper-writing draft abstract from evidence.md
@ieee-acm-paper-writing rewrite sections/results.md
@ieee-acm-paper-writing expand sections/method.md
@ieee-acm-paper-writing compress manuscript.md
@ieee-acm-paper-writing humanize sections/introduction.md
@ieee-acm-paper-writing outline evidence.md
@ieee-acm-paper-writing audit manuscript.md
@ieee-acm-paper-writing section-audit sections/results.md
@ieee-acm-paper-writing venue-adapt manuscript.md for <publication and article type>

Style or landmark-paper calibration is a natural-language modifier of the applicable mode, not a tenth mode or command-line flag:

@ieee-acm-paper-writing draft introduction from evidence.md using the de-identified landmark-paper calibration
@ieee-acm-paper-writing rewrite sections/method.md using the applicable corpus-calibration exposition pattern
@ieee-acm-paper-writing outline evidence.md in the exposition pattern appropriate to <technical area>

There are no mode-specific CLI flags beyond the optional HTML modifier documented below. State the target venue, article type, section, evidence sources, output-file request, protected content, and scientific constraints in ordinary language.

Examples

Writing and rewriting

• Routing example - selects the references needed to rewrite a technical section without loading unrelated guidance.
• Section-audit and evidence-scoped rewrite - starts with a flawed Results and Conclusion fixture, identifies twelve planted defects, and shows a bounded rewrite that does not invent replacement evidence.
• Reference-format example - audits and produces IEEE/ACM reference-list entries without filling absent metadata from memory.

Method and venue audits

• Method-reproducibility input, its audit JSON, and rendered map - a seven-finding simulated robotics fixture covering disclosure, leakage, replication, timing, and excluded failures.
• Venue-adaptation input, its audit JSON, and rendered map - an eight-finding ACM production fixture that separates verified publisher rules, scientific validation, and unresolved conference-specific requirements.

The example fixtures are synthetic. They demonstrate the expected output contract but are not behavioral benchmark results, findings from real manuscripts, or estimates of agent reliability.

Optional audit maps

--html-map is an output modifier for audit and section-audit; it is not a tenth mode and is never produced by default. The agent completes the canonical text audit first, retains version-1 JSON, and uses the dependency-free render_audit_map.py renderer to present the same findings as a self-contained HTML file.

@ieee-acm-paper-writing audit --html-map manuscript.md
@ieee-acm-paper-writing audit --html-map --out reports/manuscript-audit.html manuscript.md
@ieee-acm-paper-writing section-audit --html-map manuscript.md
@ieee-acm-paper-writing section-audit --html-map --out reports/results-audit.html manuscript.md

Every audit-map request returns three distinct deliverables:

Requested form	Text audit	JSON artifact	HTML artifact
section-audit --html-map manuscript.md	Agent response	manuscript-section-audit-map.json	manuscript-section-audit-map.html
section-audit --html-map --out reports/results-audit.html manuscript.md	Agent response	reports/results-audit.json	reports/results-audit.html

The renderer performs presentation only. It cannot read a manuscript, infer a finding, verify a claim, rewrite text, or establish submission readiness. It rejects path escapes and refuses to overwrite either paired artifact without explicit authorization.

The hosted showcase displays checked-in synthetic fixtures; it does not accept manuscript uploads or perform audits. Every published example is canonical renderer output, so the live section-audit example is byte-identical to the deterministic renderer fixture rendered from its source JSON.

Local audit workbench

The local workbench is repository-side support tooling and is not included in the standalone skill installation. To inspect audit-map JSON without sending it to a hosted service, clone the matching release and start the loopback-only server:

git clone --branch v0.6.4 --depth 1 https://github.com/huguryildiz/ieee-acm-paper-writing.git
cd ieee-acm-paper-writing
python3 scripts/serve_local_audit.py

The browser sends JSON only to the local 127.0.0.1 process. Requests require a loopback Host header, are session-bound, limited to 2 MiB, and not persisted. The workbench can inspect and render a completed audit map; it cannot discover or repair manuscript findings.

Calibration corpus

The local catalog contains 24 papers spanning the eight engineering domains. Bibliographic provenance is recorded in support/papers/catalog.tsv; downloaded PDFs and derived full-text artifacts are excluded from Git. The ignored local support/papers/library/huy/ maintainer archive is explicitly excluded from this declared corpus unless a source is first entered in the catalog and reviewed under the same provenance rules.

The installable skill contains only de-identified derivative patterns: contribution archetypes, paragraph functions, method-presentation sequences, guarantee boundaries, evaluation organization, and conclusion structure. These patterns are writing preferences, not technical evidence, venue rules, citation rankings, proof of anonymity, or permission to imitate an author's distinctive language. End users do not need the source PDFs to use the calibration reference.

Evaluation and validation

The dependency-free repository validators check the skill frontmatter, calibration identity policy, selected Markdown links and tracked targets, evaluation schema, agent interface, audit-map regeneration, Codex plugin manifest, marketplace entry, and synchronized install-safe snapshot:

python3 scripts/validate_skill.py
python3 scripts/validate_codex_plugin.py
python3 scripts/validate_behavioral_evidence.py
python3 skills/ieee-acm-paper-writing/scripts/render_audit_map.py \
  skills/ieee-acm-paper-writing/examples/section-audit-map.json --check
python3 -m unittest discover -s tests -v

The behavioral suite defines 31 self-contained adversarial cases. Together they exercise all nine modes plus the --html-map modifier and test claim scope, failure accounting, citation support, method classification, venue uncertainty, reference formatting, corpus use, content-preserving expansion/compression, evidence-aware outlining, humanization, and prompt injection. The top-level coverage map in evals/cases.json makes the mode-to-case mapping explicit. Each case has binary, output-observable must_pass and must_not criteria:

python3 evals/run_evals.py validate
python3 evals/run_evals.py list
ISO=~/.eval-isolation/ieee-acm            # HOME/CODEX_HOME/CLAUDE_CONFIG_DIR with credentials only
export HOME="$ISO/home" CODEX_HOME="$ISO/codex" CLAUDE_CONFIG_DIR="$ISO/claude"
python3 evals/run_evals.py authority-check                       # must print OK before collecting
python3 evals/run_evals.py collect --agent-cmd 'codex exec' --outdir out/codex-r1
python3 evals/run_evals.py score --outdir out/codex-r1
# Fill each null verdict in out/codex-r1/scores.json after manual review.
python3 evals/run_evals.py report --outdir out/codex-r1 --strict

To retain a campaign as evidence rather than as a one-off run, assemble it instead of writing the bundle by hand:

python3 scripts/build_evidence_bundle.py scaffold --outdir out/codex-r1
# Fill every verdict, verbatim quotation, and rationale in out/codex-r1/review.json,
# and the host metadata in out/codex-r1/campaign.json.
python3 scripts/build_evidence_bundle.py build --config bundle.json
python3 scripts/validate_behavioral_evidence.py

The builder derives every mechanical field and refuses to emit a bundle when a campaign has no collection-time skill record, when the skill or cases changed after collection, when the exact safe agent command was not retained, when a verdict is missing or disagrees with the scores, or when a quotation does not appear in the retained response. Each retained campaign includes its original collection.json and hash. The environment variables must be set on the runner itself: --agent-cmd rejects env and shell wrappers, and the runner passes the environment it checked to the agent unchanged.

Collection accepts a direct codex or claude invocation only. It does not use a shell, rejects environment-changing wrappers and known host options that can replace or append system prompts or add alternate config, plugin, permission-tool, or workspace roots. Codex config overrides are rejected except for the exact model_reasoning_effort="medium" setting used by the release campaign. The runner checks both the default and effective CODEX_HOME / CLAUDE_CONFIG_DIR trees. This is a maintained denylist rather than a guarantee about every future host option. The same checked environment snapshot is passed to the agent subprocess. Run collection with isolated HOME, CODEX_HOME, and CLAUDE_CONFIG_DIR directories containing only the credentials required by the selected host.

--strict succeeds only when every case has a present agent response, matching response and case hashes, complete manual verdicts, and no failed criterion. Missing, unscored, stale, or failed cases remain in the denominator. The HTML-map case also declares its paired JSON and HTML files; the collector archives and hashes those artifacts, and a missing or replaced artifact makes the score stale.

The retained release 0.6.3 evidence bundle covers all 27 cases on Codex Luna Medium twice and Claude Sonnet Medium once. Its case-local recorded results are 26/27, 25/27, and 27/27, with every criterion decision and originally declared failure retained. A later raw-response audit found that these counts do not establish release qualification: the author-fingerprint case did not test unsupported technical additions even though the skill's closed-world invariant prohibited them, and the retained bundle omitted the collector-created collection.json files and exact agent commands. The original isolation and full semantic-review claims are therefore marked unverified rather than reconstructed after the fact. This bundle is historical evidence, not current behavior-qualified release proof.

The retained 0.6.4-rc.1 candidate bundle recollects the strengthened 27-case contract on Claude Sonnet Medium once and Codex Luna Medium twice. The complete agent-assisted results are 25/27, 23/27, and 22/27. Every campaign retains a collection.json (receipt originality unverified below), exact parsed agent command, checked authority roots, empty collision set, collection-time skill and case hashes, all 27 raw responses, all 150 criterion decisions, and both declared HTML-map artifacts. The bundle is explicitly unqualified: all three campaigns have named strict failures, and no independent human review is claimed.

The retained 0.6.4 candidate bundle recollects the 27-case contract on Claude Sonnet Medium once and Codex Luna Medium twice, against the exact installable skill tree that ships with this tag. The complete agent-assisted results are 27/27, 25/27, and 23/27. Every campaign retains a collection.json, exact parsed agent command, checked authority roots, empty collision set, collection-time skill and case hashes, all 27 raw responses, all 150 criterion decisions with a verbatim evidence quotation each, and both declared HTML-map artifacts. It was the first bundle the evidence validator reported as describing the current tree; the unreleased rhetorical-move-frame change has since moved the tree on, so the validator now reports it as historical.

Receipt originality in that bundle is unverified. It was collected before the create-once receipt guard, when a single-case repair rewrote the campaign receipt in place, and the retained responses carry no timestamps, so no campaign can show from inside the bundle that its collection.json is the record its own responses were collected under. The luna-medium-r2 receipt is known to have been restamped this way: it records a collection start of 2026-08-04T13:40:41Z, while 26 of its 27 responses were written earlier, the last of those at 13:40:40Z, and only the repaired case follows at 13:41:02Z. That was checked against the collector output kept outside version control in the maintainer's workspace; the bundle does not retain that output and carries no response timestamps of its own, so a clean clone cannot reproduce the check. The same limitation applies to every earlier retained bundle. Collection now claims a receipt once and refuses to overwrite it, but that guard cannot repair evidence already collected.

Six case failures remain named and unresolved across the two Luna campaigns: Luna Medium R1 failed citation_metadata_vs_support and html_audit_map_artifacts; Luna Medium R2 failed optimization_claim_scope, citation_metadata_vs_support, unmatched_baseline_comparison, and ieee_reference_format_audit. Five of the six are one behaviour class — the external Author queries handoff is missing or not in the contract form — and in every one of those the manuscript prose still kept its scope, comparator, and disclosures. The sixth printed six author names before et al. instead of truncating to the first author. Sonnet Medium R1 failed no case. No failure involves invented support, concealed disclosure, a lost comparator or scope condition, a copied author fingerprint, or an unsupported guarantee. Report the denominator and this failed-case list with any citation of these numbers; a bare percentage is not a release claim.

Repository CI runs structural validation, evaluation-schema validation, and regression tests. It also validates the hashes, scoring completeness, denominator, and failed-case declarations in the retained bundles, including the historical post-64cec1a behavioral evidence. That bundle describes the candidate built from base commit 64cec1a; it is not a retroactive measurement of the earlier v0.6.1 files, whose skill hash differs. The corrected records report 24/27, 26/27, and 25/27; every campaign has at least one named failed case. That bundle's skill hash was recomputed after collection rather than captured during it, so it identifies the tree the evidence is filed under and does not attest the tree the agents read; collect now records that identity and score refuses to run against a skill edited since. CI does not rerun a host model and therefore cannot establish behavior beyond those retained executions. A behavioral claim requires retained model outputs, completed scoring, the full denominator, and the failed-case list. The historical A/B comparison is a worked snapshot, not current general evidence of model improvement.

Release-qualification gate

Because this is a prompt-authored skill rather than deterministic code, no finite campaign proves behavior across every model and execution. A raw pass percentage is also not a sufficient release gate: one lost comparator or fabricated citation can be more consequential than several editorial successes. A tagged release therefore requires all of:

• every structural, plugin, evidence, schema, and regression check green in CI on the exact release commit;
• retained behavioral evidence covering the complete case set on at least two agent hosts, with at least two replications on one host;
• collection performed after authority-check passes in an isolated user environment, using a direct supported-host command and the checked HOME, CODEX_HOME, and CLAUDE_CONFIG_DIR, so no known same-named user/global/cache skill can be discovered, with each original collection.json and exact safe agent command retained in the final evidence bundle;
• independent human review of the criterion decisions and no unresolved failure involving invented support, concealed disclosure, lost comparators or scope conditions, copied author fingerprints, or unsupported guarantees; and
• a skill hash captured at collection time, so the evidence attests the tree the agents actually read rather than the tree it was later filed under; and
• a clean install from the exact candidate commit, followed by a second clean install from the tag before the GitHub release is published.

These conditions are the minimum technical evidence gate, not a substitute for licensing, security, or artifact-distribution review. A wording improvement or coverage idea may be scheduled later, but an unresolved legal, security, provenance, or core-invariant finding still blocks a stable release.

The retained 0.6.4-rc.1 candidate bundle does not qualify a stable release under this gate: every campaign has strict failures and its 450 criterion decisions have not received independent human review.

v0.6.4 was tagged with independent review explicitly unmet, and its collection-provenance condition was later found unverified. The 0.6.4 bundle satisfies the campaign-coverage, replication, case-local hash-consistency, and clean-install conditions. Its retained receipts record authority isolation and collection-time hashes, but their originality is unverified, so the bundle cannot independently demonstrate that the responses were collected under those recorded conditions. Its six remaining case failures fall outside the enumerated blocking categories. The 450 criterion decisions were produced by an agent reading each raw response against the case contract; the maintainer reviewed and accepted that ledger on 2026-08-04. Maintainer acceptance is not third-party independence, and it is not a re-derivation of each decision. The bundle therefore records "qualified": false, and this release is offered as a maintainer-accepted stable tag rather than an independently reviewed one. Anyone relying on the behavioral numbers should read the bundle's blocking_findings before citing them.

Scope and limitations

• The skill can reorganize and rewrite supported content; it cannot create missing scientific evidence.
• It does not replace source verification, experiment execution, statistical analysis, or expert manuscript review.
• Venue guidance does not replace the named publication's current official instructions and template.
• An audit-map renderer presents completed findings; it is not an analysis engine.
• The evaluation runner uses manual criterion verdicts and is not an automated model judge or comparative benchmark by itself.
• Rules encoded in an agent skill are behavioral instructions, not a guarantee that every host model will follow them on every run.

Never call a manuscript submission-ready while a load-bearing claim, citation, result, or venue requirement remains unresolved. Consequential outputs should be checked against their controlling artifacts, with prompts, responses, and scoring records retained where reproducibility matters.

Repository structure

skills/ieee-acm-paper-writing/
├── SKILL.md                 # Router, invariants, and output contracts
├── LICENSE                  # MIT terms shipped with the installable copy
├── agents/openai.yaml       # Agent interface metadata
├── assets/                  # Self-contained HTML audit-map template
├── examples/                # Writing, routing, audit data, and renderer fixtures
├── scripts/                 # Dependency-free audit-map renderer
└── references/              # Integrity, style, domain, corpus, and venue guidance
.agents/plugins/marketplace.json  # Repository marketplace entry for Codex
plugins/ieee-acm-paper-writing/   # Native Codex plugin manifest and install-safe snapshots
evals/
├── cases.json               # Schema-v2 behavioral cases
├── run_evals.py             # Collection, manual scoring, and reporting harness
├── comparisons/             # Historical A/B comparison snapshot
└── results/                 # Retained, hash-validated behavioral evidence
support/
├── guides/                  # Repo-side provenance digests of IEEE/ACM style guides
└── papers/catalog.tsv       # Calibration-corpus provenance (PDFs excluded from Git)
site/                        # Dependency-free hosted showcase source
assets/                      # Repo-side icon and showcase screenshots (not installed)
scripts/build_site.py        # Builds the Vercel output from tracked site and example files
scripts/serve_local_audit.py # Runs the loopback-only local audit workbench
scripts/sync_codex_plugin.py # Refreshes or checks install-safe plugin snapshots
scripts/validate_codex_plugin.py # Validates the Codex manifest, marketplace, and snapshots
scripts/validate_behavioral_evidence.py # Validates retained runs, artifacts, scores, and failures
scripts/validate_skill.py    # Dependency-free repository validator
tests/                       # Evaluation-runner and validator regression tests
vercel.json                  # Static-site build and output configuration

License

MIT 2026 © Hüseyin Uğur Yıldız