Open QEC harness – greedy vs. GE, uniform vs. clustered k=4

I'll pull the GitHub repo and HN thread so the paragraphs stay grounded in what's actually there—no invented numbers or dates.An open **QEC Evaluation Suite** from GitHub user **mrblakessinger-rgb** is circulating as a finite-size diagnostic harness for circuit-level and holographic-style quantum error correction. The repo frames itself as digital evaluation and telemetry only—not hardware, not a production decoder, and not an asymptotic threshold paper. On Hacker News under the AI Agents feed it sat at **1 point** with **0 comments**, pointing readers at https://github.com/mrblakessinger-rgb/qec-evaluation-suite. The headline comparison the release pushes is **greedy vs. GF(2) Gaussian elimination (GE)** recovery on a HaPPY-style layer, plus a pattern-stress cut of **uniform vs. clustered** erasures at **k=4**.

Layer **C** is the HaPPY path: a tensor-network model of bulk-boundary encoding with greedy recovery, plus an optional matched-sample **GF(2) GE** ablation on the same erasure masks. Reproducible JSON under `public/data/` reports, at **p_erase = 0.3**, greedy central scores of **0.664** (depth 1) and **0.59** (depth 2) against GE central scores of **0.304** and **0.0575**. Depth-1 uses **n=10, k=6**; depth-2 is capped at **n=20, k=16**. On the pattern stress table (depth-1, k-matched, **400 trials per cell**, seed **42**), uniform **k=4** yields greedy **~0.27** / GE **~0.02**, while clustered **k=4** yields greedy **1.0** / GE **0.0**. The README’s non-claim is explicit: GE is not a recovery superset of greedy on this instance, and the gap widens at depth-2 under the linear face-parameter model.

The rest of the suite is layered so those numbers are not read as a single “decoder ranking.” **Layer A** covers Stim rotated-surface smoke (graphlike DEM, MWPM) and a period-3 honeycomb Floquet schedule that is explicitly not treated as a graphlike MWPM peer of surface. **Layer B** is educational graph invariants, not circuit-level QEC. A separate BP micro-lab on classical Hamming **[7,4]** min-sum BP keeps **ge_unique_recoverability**, **bp_codeword_match**, and OSD syndrome/codeword fields as non-interchangeable metrics. Prefer frozen JSON over re-running heavy sweeps when comparing claims; smoke entry points are `python -m qec_engine.layer_a_smoke` and `python -m qec_engine.repo_rollup`.

For engineers building simulators, decoder prototypes, or evaluation pipelines, the useful signal is methodological: matched erasure masks, decoder × pattern interaction, and honesty language that refuses to promote a finite HaPPY toy into a universal geometry or threshold result. **k=4** is special for greedy—clustered holds, uniform drops—while under GE both modes sit near the floor. That is a concrete warning against ranking recovery methods without fixing the error pattern class, and against treating GE uniqueness as interchangeable with greedy success or BP codeword match.

In market and tooling terms this sits beside Stim-centric surface workflows and production-minded MWPM stacks as a small, open “wind tunnel” rather than a competitor to full-stack QEC stacks or hardware roadmaps. Surface and Floquet artifacts live next to holographic HaPPY ablations so you can see where graphlike MWPM smoke ends and bulk-boundary recovery experiments begin. Claims and non-claims are centralized in `QEC_CLAIMS.md`, with JSON under `public/data/` as the tie-breaker if prose and data disagree—useful when open QEC repos otherwise mix demo UI, asymptotic language, and one-off scripts without a reproducible ledger.

Practical takeaway: clone the suite, treat the matched greedy-vs-GE tables and the uniform/clustered **k=4** stress as the primary artifacts, and re-check any extension against the same seed and trial counts before asserting decoder dominance. Watch whether later commits deepen the depth-2 GE gap, expand pattern stress beyond **k=4**, or keep Layer C strictly finite-size while Layer A remains smoke-only. Source of truth for numbers remains the frozen JSON paths named in the README, not secondary write-ups.