Epistemic Deconstructor — Development skill for Claude Code

Epistemic Deconstructor

[](LICENSE) [](CHANGELOG.md) [](tests/) [](src/scripts/) [](https://www.electiconsulting.com)

**Turn "I don't know how this works" into a quantified, validated model of how it works.**

Epistemic Deconstructor is a [Claude](https://claude.com/claude-code) skill. You hand Claude a black box; it runs a structured scientific investigation — Bayesian hypothesis tracking, falsification tests, parametric model fitting, conformal prediction — and writes every step to disk so context-window loss can't erase progress. The deliverable is a validated predictive model with parameter uncertainty bounds, not a summary document. Because every hypothesis, probe, and update is persisted to a session directory, a context-window reset mid-analysis is a non-event: the next turn resumes from disk, not from memory.

You don't run the tools by hand. You talk to Claude; Claude adopts the orchestrator role, picks a depth tier, and drives the whole protocol for you.

What it sounds like

**Engineering — black-box reverse engineering.**

*"Help me figure out how this API's rate limiter works — I only have request/response logs."*

Claude seeds three competing mechanisms (token bucket / sliding window / leaky bucket) with priors, designs differential probes (burst vs. steady traffic), fits a parametric model with capacity and refill-rate uncertainty, and returns allow/reject predictions with conformal intervals plus a head-to-head comparison against a "count requests" baseline.

**Security — side-channel investigation.**

*"Does this login endpoint leak whether a username exists through response-time variance?"*

Claude