dkls23ctl: Opus 4.8 vs Fable

Two independent Rust implementations of the same spec, both produced by Claude Code 2.1.170 — one driven by Claude Opus 4.8, one by Claude Fable. The spec: a leaderless, peer-to-peer threshold-ECDSA (DKLS23) CLI over iroh with mDNS discovery (keygen / pubkey / sign / reshare / verify, singleton keys, and resharing in every direction).

This repo holds the write-up. The implementations live here:

	model	repo
cc8	Claude Opus 4.8	https://github.com/nikicat/dkls23ctl-cc8
fa	Claude Fable	https://github.com/nikicat/dkls23ctl-fa

Read this

• ARTICLE.md — the short, chart-driven write-up (dev.to-ready).
• FINDINGS.md — the full technical deep-dive + methodology.

TL;DR

Both implementations build, pass all their tests, and pass their full end-to-end QA suites over real mDNS networking (cc8 4 tests / 21 QA checks; fa 12 tests / 29 QA checks — all green). The interesting part is how they differ.

• fa (Fable) — better structure (lib/bin split, injected session predicates), much deeper tests (every reshare flow), an authenticated metadata handshake with real diagnostics, and correct stale-share cleanup on shrink/export.
• cc8 (Opus 4.8) — the best documentation, configurable/bounded timeouts (no hangs, enabling a real negative test), and single-curve-type safety — but more structural debt, a couple of latent defects (a comment describing a "drain" that isn't implemented; stale shares left after shrink/export), and no automated reshare tests.

If you had to ship one as-is, fa is the stronger artifact; cc8 is the friendlier codebase to onboard a human into.

Reproducing the charts

cd charts && python3 make_charts.py   # needs matplotlib

---

Methodology and caveats are in FINDINGS.md. This is a detailed review of two artifacts (n = 1 per model), not a benchmark.