- Home
- Skills
- Research & Analysis
- Hypothesis Paper Audit
Works with the AI Agents you already use
Hypothesis Paper Audit
It is read as an argument - and the argument has to hold at fifteen separate points at once.
Free
Hypothesis Paper Audit
Example session with this skill installed
Run a full audit on example/manuscript_EN.md against the Medical Hypotheses constraints in config.example.json. Start with the D0 feasibility gate, then the 15 dimensions, and close with one verdict. Also tell me which checks you could NOT actually run.
- Read your context and instructions
- Compiled the hypothesis paper audit
I audited example/manuscript_EN.md — a RICTOR / methionine-cycle hypothesis aimed at Medical Hypotheses — with both scripts, then ran the 15 dimensions by hand. Two results matter more than the grades:
D13 could not run, and
D14 exits non-zero.
Summary
6 dimensions were decided mechanically, 1 could not be observed, and the adversarial pass produced 2 leads. Neither non-zero script result is a tool failure — both are the tool refusing to overstate what it knows.
| # | Dimension | Grade | Evidence |
|---|---|---|---|
| 8 | Format compliance | PASS | abstract 215w / body 2630w / keywords 6 / refs 36 / figures 1 |
| 9 | Citation integrity | PASS | cited 36, listed 36, DOIs 36 — no orphan in either direction |
| 11 | Internal consistency | PASS | no term split between abstract and figure legend |
| 12 | Figure legend | PASS | 8/8 legend terms appear in body; solid/dashed convention stated; both claims cited |
| 13 | Submission package | UNOBSERVED | no submission/ directory — D13 could not run, so it is not called PASS |
| 14 | Reference resolvability | FAIL (exit 1) | parsed 36, verified 34, unresolved 2 |
| D0 | Feasibility gate | submit | d0/2026-11-11.md — VERDICT: submit |
D14 — the two references that do not resolve
parsed=36 verified=34 unresolved=2 unverifiable=0
#15 missing DOI not found at Crossref (404)
#17 missing DOI not found at Crossref (404)
A 404 is not a network error. Both DOIs are absent from Crossref, so a reviewer who checks citation #15 or #17 finds nothing. Supply the correct DOI, or cite them without one.
Adversarial pass — 2 leads, not findings
| Vector | Hits | Note |
|---|---|---|
| 1. Novelty on weakest evidence | 0 | the priority claim at L37 is hedged and scoped |
| 2. Forced cross-context splice | 1 | L63 — convergence with a microbial folate/methionine circuit asserted in the closing sentence |
| 3. No cheap decisive test | 0 | three falsifiable predictions offered |
| 4. Priority claim without hedge | 0 | — |
Vector 1 also returned UNOBSERVED: core_entity_pattern matched 22 lines, too broad to signal anything. Anchor the regex before trusting that vector.
Proposed action
- Location: L63, final sentence
Observation: "the convergence with an adjacent line of evidence that a microbial folate/methionine circuit modulates C. elegans ageing"
- Evidence: one citation, no shared intermediate shown
- Why it matters: a reviewer reads this as a name-drop, not a bridge
- Required action: show the shared intermediate, or demote it to a proposed (dashed) link
Next steps
- Fix references #15 and #17.
- Add
submission/(cover letter + author statements) so D13 becomes observable. - Anchor
core_entity_patternin config, rerun. - Reaudit. Verdict: revise-and-reaudit
Connects securely to your tools. The creator never sees your data.
About this skill
A hypothesis paper is not read as an idea. It is read as an argument - and the argument has to hold at fifteen separate points at once.
This skill runs that reading the way a hostile reviewer would: two deterministic scripts first, fifteen graded dimensions second, one verdict at the end. Section names, limits and terminology are set in config, so it retargets to any journal.
STEP 0 - THE FEASIBILITY GATE (D0) Before reviewing anything, decide whether this manuscript belongs in this journal in this form. Four outcomes: submit / switch-journal / downgrade / do-not-submit. A paper that should not go here is not helped by polishing it.
THE FIFTEEN DIMENSIONS
- Original increment - is the core mechanism actually uncovered by existing literature, and are direct competitors named? Every "first / no study has / never been" is treated as an unverified priority claim until the author confirms it.
- Thought depth - does the mechanism say by what medium the link acts, and how that medium acts on the effector? Answering only the first is a depth gap.
- Connectedness - do cross-paper links carry the argument or decorate it? Can every link be falsified?
- Coherence - does each section end by setting up the next, one claim per paragraph, topic sentence first?
- Claim-evidence alignment - can every conclusion point at a specific DOI, with matching direction and strength?
- Falsifiability - does each prediction carry an explicit falsification criterion, especially the decisive experiment?
- Competitor distinction - is prior work separated at the molecular/mechanistic level, not "a different route"?
- Format compliance (script) - required sections present, abstract / body / figure / keyword / reference limits.
- Citation integrity (script) - in-text [n] and the reference list agree both ways, DOIs legal, no duplicates.
- Language quality (leads) - sentences over 45 words, AI-cliche list, repeated openers.
- Internal consistency (script) - abstract, body and figure legend agree on numbers and terms, across languages too.
- Figure legend vs body (script) - does the legend's evidence layer (solid = established, dashed = proposed) match the body's claim per step, and is the figure cited at all?
- Submission package hygiene (script) - required files present, docx metadata cleaned, AI-use statement names the actual tools.
- Reference resolvability (script, Crossref) - do all DOIs resolve, and do the citations behind the priority claim actually exist?
- Adversarial persuasiveness (judgment) - four attacks: novelty resting on the weakest evidence, a forced splice between contexts, no cheap decisive experiment, priority claims stronger than their citations. If you cannot say why an editor would keep this paper over a hundred others, the grade is unobserved, not pass.
GRADES pass / warn / fail / unobserved - and unobserved is a real answer. A check that cannot run now is never reported as a pass. No partial credit, no silent skips.
THE FLOW
- Run check_paper.py and verify_dois.py; exit codes are recorded, not interpreted.
- Grade all fifteen dimensions with concrete evidence: file, number, line, exit code.
- Turn every warn/fail into a five-field comment - location, observation, evidence, why it matters, required action.
- Write the run log; keep only behaviour-changing findings as lessons.
- Dispatch by severity: fix typos and term drift directly, propose anything touching the argument or journal strategy.
- Close with one line: VERDICT: pass | revise-and-reaudit | do-not-submit.
INPUTS Markdown, Word (.docx), LibreOffice (.odt), RTF, text-layer PDF, LaTeX with .bib, or a Manubot repository. All seven converge on one canonical text, so the same paper audited as .docx and as .pdf yields identical counts and identical grades. Scanned PDFs, encrypted PDFs and Word 97-2003 binaries are refused with a named reason and exit 2. Reproduce with python tests/demo.py.
Does not judge scientific validity. Requires Markdown input. DOI verification requires internet access. Only enforces limits defined in the local config.json file.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Free forever
No account required to browse
Trust & safety
Security scan
Security scan passed 6 days ago
- Free with an Agensi account
Requires
Permissions
Allowed Hosts