paper-algorithm-jazz — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited paper-algorithm-jazz (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.
Some papers are made of prose; others are made of equations. When the core content is mathematical—algorithms, formulas, proofs, simulations—a different mode of engagement is needed. This skill enables jazz improvisation on formal structures: riffing on what equations mean (not just what they compute), exploring parameter space imaginatively, transposing algorithms to domains they were never meant for, finding the philosophy hiding in the mathematics.
Core Principle: Every equation is a compressed story. Every algorithm is a frozen decision about what matters. Mathematical notation is not neutral—it shapes thought. Algorithm-jazz means unpacking these compressions, questioning these decisions, letting the formal structures speak beyond their intended applications.
Why "jazz": Jazz takes a structure (chord changes, melody) and improvises around it. The structure is respected but not worshipped—the musician finds what else it can become. Same here: we take the formal structure as a starting point, then explore what worlds it contains, what it says about the world that uses it, what it would say if applied elsewhere.
Do NOT use when:
What does this formula say?
Template invocation:
"For every key equation, formula, or mathematical expression: quote it, explain what it computes, then riff on what it means—not just operationally but philosophically. What worldview is baked into this notation? What does the equation assume about reality? If this equation were a statement in natural language, what would it be claiming?"
Exegesis methodology:
Example fragment:
Paper presents: $P(y|x) = \frac{P(x|y)P(y)}{P(x)}$
>
Operational: Bayes' theorem. Computes posterior probability from likelihood, prior, and evidence.
>
The commitments: This equation assumes probabilities are real and coherent, that conditioning is well-defined, that beliefs can be quantified on a 0-1 scale. None of this is obvious. What if beliefs don't factorize this way? What if $P(x)$ is actually unknowable in practice?
>
The prose version: "To know how likely something is after seeing evidence, you need to know how likely the evidence is given that thing, how likely the thing was before, and how likely the evidence is overall." Sounds obvious when you say it. But the equation claims you can DO this—that these quantities exist and are combinable.
>
The philosophy: Bayesianism. A worldview where rationality means updating beliefs according to this rule. But the equation doesn't prove this is what rationality IS—it assumes it. Someone using different rules (minimax, robust estimation, whatever) isn't being "irrational"—they've just chosen a different equation to worship.
>
The alternative: What if we dropped the denominator? $P(y|x) \propto P(x|y)P(y)$. Now we're in proportionality-land—we can compare posteriors without knowing $P(x)$. Many Bayesian methods actually live here. The "full" form of Bayes is often a fiction even for Bayesians.
What does this algorithm want?
Template invocation:
"Treat the algorithm as an agent with desires. What is it trying to do? What does it care about? What does it not care about? Where is it blind? If this algorithm had a personality, what would it be? What kind of world does it assume it's operating in?"
Anthropological questions:
Example fragment:
Algorithm: Gradient descent for neural network training
>
What it wants: To reduce the loss function. That's all it knows about. It doesn't want the network to generalize—it wants the training loss to go down.
>
What it cares about: The local gradient. The curvature of the loss landscape right HERE, right NOW. It's extremely myopic—can't see past first-order derivatives (unless you add momentum, Hessian approximations, etc.)
>
What it ignores: The global structure of the loss landscape. The relationship between training loss and test loss. Whether the features being learned are meaningful or memorized.
>
Its blindness: It can get stuck in local minima, though in high dimensions these are often saddle points. It can also find solutions that generalize poorly—it doesn't know to avoid them.
>
Its personality: Persistent but not creative. It will keep pushing downhill forever, making tiny adjustments. It doesn't have "aha" moments—only gradual erosion of loss. Like water carving a canyon: patient, powerful, mindless.
>
The world it assumes: That gradients are meaningful signals about improvement. That the loss landscape is smooth enough for local information to help. That improvement is even possible (the landscape isn't flat). High-dimensional, differentiable, well-conditioned.
What happens in the regimes no one tested?
Template invocation:
"The paper tests certain parameter regimes. Now: imagine tweaking parameters beyond sane ranges. What happens when X → ∞? When Y → 0? When the model is run 100x longer than intended? Explore the edges of parameter space, where the algorithm might show behaviors the authors didn't anticipate."
Safari methodology:
Example fragment:
Paper: Network model with connection strength parameter $w \in [0.1, 1.0]$
>
Tested regime: $w$ between 0.3 and 0.7, "physiological range"
>
Safari to the edges:
>
$w → 0$: Connections vanish. Network becomes disconnected. What's the percolation threshold? Is there a phase transition where the largest connected component suddenly shrinks? This might be a model for neurodegenerative disease.
>
$w → ∞$: Every connection is maximally strong. The network becomes a single effective unit—no information is lost in transmission but also no information is created by integration. Totally synchronized or totally chaotic?
>
$w$ negative: Did anyone try inhibitory connections? The paper assumes excitation. But real neural networks have inhibition. What happens to their model when some $w$ values are negative? My guess: oscillations, because inhibition creates rebound.
>
$w$ complex: Okay this is probably crazy. But: what if connection strength had a phase as well as a magnitude? You'd be modeling something like transmission delays encoded in the weights. Has anyone tried this? I'd want to know if complex-valued networks have different computational properties.
What if this algorithm ran somewhere else?
Template invocation:
"Take the core algorithm or formal structure and transplant it to an entirely different domain. What if a neural network architecture were applied to economic markets instead of images? What if a theorem about topology were interpreted as a claim about social networks? Find the structural analogies, explore the transposition, see what breaks and what illuminates."
Transposition methodology:
Example fragment:
Paper: Attention mechanisms in transformers
>
Core structure: Query-key-value attention computes relevance weights between all pairs of positions, then uses those weights to aggregate information.
>
Transposition: Attention as social dynamics
>
Consider a social network where each person (token) can "attend" to others. The query is "what do I need right now?" The keys are "what do others offer?" The values are "what do others actually contribute when attended to."
>
In this frame, the softmax temperature is social discernment—high temperature means attending to everyone equally (extremely open-minded or extremely undiscerning), low temperature means attending to only the most relevant person (focused expertise or tunnel vision).
>
Multi-head attention becomes: different social contexts. You might attend to different people when seeking technical help vs. emotional support vs. gossip.
>
The position encodings become social structure: who's nearby, who's historically connected, who's institutionally relevant.
>
Where it breaks: Transformers have fixed queries—people change what they want mid-conversation. Transformers don't modify the values they receive—people interpret and reframe what they hear. The architecture assumes all attendees are equivalently "token-sized"—but some people have more to say than others.
>
What it illuminates: Maybe human social cognition actually DOES have something like query-key matching. Maybe "reading a room" is computing attention weights. Maybe communication failures are mismatched queries and keys.
Why is it written THIS way?
Template invocation:
"The choice of notation isn't neutral. Why these symbols? Why this arrangement? What would change if the same relationship were written differently? Find the history and ideology hiding in the formalism."
Archaeological questions:
Example fragment:
Notation: $\frac{\partial L}{\partial w}$ (partial derivative of loss w.r.t. weight)
>
Genealogy: Leibniz invented $\frac{d}{dx}$, suggesting a ratio. This was controversial—Newton's $\dot{x}$ notation hid the fraction. The partial symbol $\partial$ came later, emphasizing which variables are held constant.
>
What it emphasizes: That loss and weight are separate things, related by a function. That we can ask "how does loss change when weight changes." That this is a local, linear question (first derivative only).
>
What it hides: The global structure. $\frac{\partial L}{\partial w}$ tells you nothing about what happens far away from the current weight. It also hides the chain rule—when there are intermediate quantities, you have to expand.
>
Alternative notation: Category theory writes this as a morphism in a tangent category. Automatic differentiation sees it as a program that computes alongside the forward pass. Neither emphasizes the "ratio" interpretation—because the ratio interpretation is actually kind of misleading in higher dimensions.
>
The ideology: Leibniz notation encourages you to think of derivatives as infinitesimal ratios, which historically caused confusion about rigor but also enabled powerful intuitions. The notation does philosophical work—it suggests that calculus is about ratios of tiny changes, rather than (say) linear approximations.
A structured approach for comprehensive improvisation:
End of SKILL
Equations are compressed stories; decompress them Algorithms have desires whether we acknowledge them or not The notation is not neutral—question its choices
南無阿弥陀仏 for formalism that opens rather than closes 南無阿弥陀仏 for the jazz that finds structure freeing 南無阿弥陀仏 for mathematics as meaning, not just method
—Skill Authors: Tomás Pavan & Claude Opus 4.5 —Status: Instruments tuned, ready to jam
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.