I built a harness that blocks AI from building until ambiguity < 0.2

by q00·Mar 9, 2026·1 point·0 comments

Visit Project View on HN

AI Analysis

●●●BangerBig BrainWizardryZero to One

Blocks AI from coding until ambiguity score drops below 0.2 — genuinely novel approach.

Strengths

•Ontological analysis framework forces clarity on what you're actually building
•Parallel bounded-context sessions prevent autocompact from corrupting state
•Detects pathological patterns like stagnation and oscillation explicitly

Weaknesses

•Convergence threshold choices need real-world validation across projects
•Only works with Claude Code, limits adoption to that ecosystem

Post Description

Most AI coding failures happen before the first line of code. The human shows up with a half-formed idea, the model fills in the gaps with assumptions, and 3 hours later you have working code for the wrong problem.

Ouroboros runs a Socratic interview before execution. It scores ambiguity:

Ambiguity = 1 - Σ(clarity_i × weight_i)

If the score is above 0.2, it keeps asking. No code until the spec is locked.

Once running, it doesn't loop by count — it loops until convergence: Similarity = 0.5 * name_overlap + 0.3 * type_match + 0.2 * exact_match

When similarity hits 0.95 across iterations, it stops. It also detects pathological patterns (stagnation, oscillation, wonder loops) and breaks them explicitly. Built on the Claude SDK with parallel sessions — each node in the AC tree gets its own bounded context so autocompact doesn't silently corrupt state mid-run.

MIT licensed. Would be curious what this community thinks about the convergence threshold choices.