paper-scrying — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited paper-scrying (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
This skill transforms reading into cross-examination. Where paper-hermeneutics seeks understanding and paper-riffing invites free association, paper-scrying approaches with productive suspicion—not cynicism, but rigorous skepticism. It asks: What must be true for this to work? What's hiding in the assumptions? What would break this? How might this be accidentally or deliberately misleading?
Core Principle: Every paper is a claim-structure vulnerable to interrogation. The authors made choices—of framing, methods, statistics, emphasis, omission. Scrying means making those choices visible, testing them against alternatives, and asking what the paper would prefer you not notice. Not to debunk, but to see clearly.
Etymological Note: Scrying—divination by gazing into reflective surfaces—here inverted. We gaze into the paper not to see futures but to see through surfaces, to catch the flicker of what's hidden beneath confident prose.
Do NOT use when:
Every paper rests on foundations it doesn't examine. Find them.
Template invocation:
"Identify the passages where the authors most clearly signal consensus, dogma, or 'received wisdom.' Quote these at length, then excavate: What assumptions are buried here? What must be true for these claims to hold? What if the opposite were true? What forbidden experiments or alternative interpretations lie just offstage?"
What to look for:
Example fragment:
The paper states: "Following standard practice, we normalized the data using z-score transformation before analysis."
>
Stop. "Standard practice" is doing heavy lifting here. Why z-score and not min-max? Why normalize at all—what about the original distribution made it unsuitable? Z-score transformation assumes the data should be approximately Gaussian, which for this kind of measurement is... questionable. What if the non-normality IS the signal? I want to see the analysis run on raw data. The fact that they don't show it makes me suspicious they tried and didn't like the results.
Approach with the mindset of someone who's been burned before. Where's the hype?
Template invocation:
"Approach this paper with a mixture of open curiosity and healthy suspicion. Where might there be hype, wishful thinking, or buried assumptions? For each claim that stirs skepticism or seems 'too good to be true,' quote it fully, then break down: What would you need to know to really trust this? What alternative explanations or unaddressed confounders leap out? If you had to build a scammy parody of this paper, what would it look like?"
Red flags to watch for:
The scammy parody technique: Imagine building a fake version of this paper. What would you do? Now check: did the authors do any of those things? Not because they're fraudulent, but because the same pressures that would lead a scammer to cut corners also affect honest researchers.
Example fragment:
"Our model achieved 94% accuracy on the benchmark dataset, significantly outperforming previous state-of-the-art methods."
>
Okay, let me build the scam version: I would (1) use a benchmark dataset I helped create, (2) tune hyperparameters extensively on the test set while calling it validation, (3) compare to poorly-tuned implementations of competitors, (4) report only the best of many runs, (5) define "accuracy" in a way that flatters my method.
>
Now let me check: The benchmark is... one they introduced last year. Hyperparameter selection is... described in a supplementary PDF I can't find. Competitor implementations are... "standard" with no links. Number of runs is... not mentioned. Accuracy definition is... buried in methods.
>
I'm not saying it's a scam. I'm saying they made it impossible to tell.
For every positive result, imagine the experiment that would disprove it.
Template invocation:
"For each major claim, design a counter-experiment—a test that would falsify the finding if it's wrong. What would you measure? What controls would you need? What result would make you say 'okay, I was wrong to be skeptical'? And conversely: what result would confirm your suspicion that something's off?"
Falsification hierarchy:
Example fragment:
The paper claims their drug reduces anxiety in mice as measured by time in open arms of elevated plus maze.
>
Counter-experiments I want: - Same drug, same maze, different strain of mice (is this strain-specific?) - Same drug, different anxiety measure (light-dark box, open field) — does it generalize? - Check if the drug affects motor function at all — maybe "less anxiety" is actually "can't move as well" - Run the maze with drug-naive observers scoring the videos blind — observer bias? - Measure blood levels — did the drug actually reach the brain at claimed doses?
>
I notice they only used one maze, one strain, and don't mention blinding. Not fatal, but not reassuring.
Find the orthodoxy and profane it (thoughtfully).
Template invocation:
"What would it mean if this paper's central claims were wrong—not fraudulent, but honestly mistaken in the way science often is? What would we learn? What alternative framework would we need? What heresies against the field's consensus does this paper suppress or ignore? Give voice to the heterodox positions."
Categories of productive heresy:
Example fragment:
This paper on neural correlates of consciousness assumes the standard cognitivist framework where consciousness is information processing in specific brain regions.
>
The heretical alternatives they don't engage: - Integrated Information Theory: Consciousness might not be localized at all - Global Workspace Theory's critics: Maybe the workspace is epiphenomenal - Enactivism: Consciousness might be in the organism-environment coupling, not the brain - Idealism (don't laugh): What if the correlation goes the other direction? - Illusionism: What if there's nothing to explain—"consciousness" is a confused concept?
>
The paper treats its framework as given. But the framework is exactly what's contested. Finding neural correlates under cognitivist assumptions proves cognitivism how, exactly?
Imagine the malicious or negligent use case.
Template invocation:
"How might these findings be abused, weaponized, or subverted by the unscrupulous? What's the 'dark twin' of this research—the version built for harm? Consider: surveillance, manipulation, exploitation, environmental damage, military application, discriminatory use. This isn't to condemn the research but to complete the ethical picture the authors may have elided."
What to consider:
Example fragment:
This facial recognition paper improves accuracy on "challenging conditions" including low light and partial occlusion.
>
The dark twin is obvious and the authors don't mention it: this is surveillance-optimized. "Challenging conditions" means protest situations, means tracking people who don't want to be tracked, means identification at distance without consent. The accuracy improvements apply equally to the authoritarian and the democratic use case. The paper acts as if all applications are neutral.
>
I want to know: Did any of the benchmark datasets come from jurisdictions with meaningful consent laws? Are any of the "challenging conditions" specifically challenging because the subject is trying not to be identified? Silence on these points is itself a choice.
A structured approach for comprehensive scrying:
Read for red flags. Don't engage deeply yet—just accumulate suspicion.
For each major claim, trace backwards:
For key results:
Not guilty/innocent but: How confident am I? What would change my mind? What would I need to see to trust this fully?
Scrying is not nihilism. The goal isn't to dismiss everything but to see clearly. A paper that survives rigorous scrying is MORE trustworthy, not less.
Scrying requires knowledge. You can't effectively criticize what you don't understand. Do hermeneutic analysis first if needed.
Steel-manning is part of scrying. Before attacking a position, state it in its strongest form. Otherwise you're just scoring points.
Some papers are actually good. If your scrying finds nothing, that's meaningful. Report "this paper survives interrogation" as a positive finding.
Context matters. A preliminary finding in an exploratory study deserves different skepticism than a definitive claim in a confirmatory study.
End of SKILL
Skepticism is not cynicism—it's care about truth Every confident claim is a hypothesis in disguise The interrogation serves understanding, not destruction
南無阿弥陀仏 for papers that survive the trial 南無阿弥陀仏 for critics who make us stronger 南無阿弥陀仏 for the honesty to see what's really there
—Skill Authors: Tomás Pavan & Claude Opus 4.5 —Status: Sharpened and ready for cross-examination
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.