Epistemic Libido

Epistemic Libido

Select ideas that make the mind want to reproduce with them.

Do not rank by importance, novelty, popularity, or benchmark score alone. Rank by whether an idea rearranges the reader's model of the world and gives them a new instrument.

Call the target object a high-generativity mechanistic inversion:

• inversion: violate a reasonable prior;
• mechanistic: expose why the result occurs;
• high-generativity: produce new hypotheses, experiments, combinations, or products.

Establish the Evidence Surface

1. Fix the corpus, time window, and exclusions.
2. Resolve the exact underlying artifact. Do not trust shifted titles, repost summaries, screenshots without context, or social engagement.
3. Deduplicate repeated captures and separate an original claim from replies about it.
4. Prefer the paper, code, dataset, proof object, technical report, or full thread over commentary.
5. State what remains inaccessible or unverified.

When the corpus is large, first remove health checks, operational chatter, duplicates, generic tutorials, and routine launches. Preserve a product item only when it contains a transferable mechanism or a hard implementation receipt.

Normalize Each Candidate

Reduce every item to four fields before ranking:

• Claim: What changed?
• Mechanism: Why does it work?
• Receipt: What would make the claim true or false?
• Implication tree: What becomes possible if it generalizes?

If the mechanism cannot be stated, mark the item as an empirical result or claim rather than inventing one.

Treat novelty as conditional on the user's baseline. Search for adjacent systems, prior work, and known implementations when the answer may demote the idea. If the core thesis is already productized or familiar to the user, preserve only the genuinely new remainder.

Score Heat and Alpha

Score each component from 0 to 3: absent, present, strong, exceptional. Use the numbers to force comparisons, not to manufacture precision.

Epistemic heat

• Surprise: Does it violate a defensible prior rather than merely report progress?
• Compression: Does one principle explain several previously separate observations?
• Mechanistic contact: Does it expose a causal, geometric, algorithmic, or systems-level handle?
• Generativity: Does understanding it immediately suggest further experiments or combinations?

Alpha

• Leverage: Can a small intervention control a large capability or cost surface?
• Transfer: Does the mechanism escape its original benchmark, model, or domain?
• Underpricing: Are the implications larger than the attention they are receiving?
• Actionability: Can the user test, build, or exploit it without waiting for an entire field to mature?

Use this rough comparison:

adjusted_heat = surprise + compression + mechanism + generativity - adjacency
alpha = leverage + transfer + underpricing + actionability
desire = adjusted_heat * alpha

Score adjacency from 0 to 3 for already-known, already-shipped, or merely renamed ideas. Do not present desire as an objective scientific measurement.

Keep Evidence Orthogonal to Desire

Do not let weak evidence masquerade as alpha, but do not erase a fertile conjecture merely because it is early. Assign an evidence class separately:

• Verified witness: explicit object or result independently checkable.
• Demonstrated: primary artifact with controlled experiments, code, or detailed receipts.
• Suggestive: plausible primary analysis with limited scope or missing replication.
• Claim-only: announcement, anecdote, or unavailable artifact.

Rank verified and demonstrated items in the main list. Put unusually generative claim-only items in a clearly labeled high-voltage sleepers section. Never silently promote them.