NX — N-version revival on a WEAK model (qwen2.5-coder:7b, k=2)

Ran 2026-09-21. Ollama qwen2.5-coder:7b (GPU), 48 generations, 0 gen-failures, 2 dead versions (46 live). No Docker/DB. Contrast: the frontier run (Sonnet, RESULTS.md) was a clean negative.

Result

Problem	Category	Baseline defect-rate	N-version flagged	Majority wrong	Disagreement-catch	False-positive
money-split	subtle	0%	0%	0%	—	0
merge-intervals	subtle	0%	0%	0%	—	0
business-days-add	subtle	0%	26%	26%	1.00 (100%)	0.00
reschedule-conflict	subtle	0%	0%	0%	—	0
sum-array	trivial	0%	0%	0%	—	0
reverse-string	trivial	0%	0%	0%	—	0

Read — honest, still largely negative, with one real signal

Baseline defect-rate is 0% on every problem for qwen7b too — like Sonnet, a single qwen generation gets all 6 right. So on 5 of 6 problems the exposure λ ≈ 0 and N-version has nothing to catch → still dead. These problems are within a competent coder model’s single-shot competence; model-weakness alone (7B vs frontier) did not open λ.

The one exception is a genuine mechanism signal. On business-days-add — the most edge-case-dense problem — qwen’s implementations were inconsistent: across the triple, versions landed on different handling of the n=0 / weekend / holiday-adjacency edges. Where 2 of 3 landed on a wrong variant (majority wrong on 26% of inputs, averaged over the two triples; 52% in the triple that diverged), disagreement flagged 100% of the majority-wrong cases with ZERO false positives (catch=1.0, fp=0). When λ>0 appeared, the N-version mechanism worked perfectly.

The subtle point that matters most: the baseline shipped 0% on business-days-add while the triple was majority-wrong 52% on one rep. A single generation is a coin flip — it sometimes samples a correct implementation and looks clean; the disagreement across three draws surfaces the model’s underlying uncertainty on those edge inputs that the one lucky single-shot hides. So on a weak/ uncertain model, N-version’s real value is exposing variance (“the model is unsure here”) exactly where silent defects live — not just catching a defect a single version already shows.

What this says about revival (and the thesis)

Revival is gated by λ, and λ = f(model, problem-difficulty). Weakening the model (frontier→7B) was not enough to open λ on easy problems — qwen is a good coder. The lever that reliably opens λ is problem difficulty beyond single-shot competence, and there N-version pays.
Capacity-relative, refined: the frontier had zero variance even on business-days (0 flags); the weak model had variance on the hardest problem (26% flagged, perfect catch). The gradient is real but the effect is small because the problems don’t stress either model enough.
Calibration datum: in the λ>0 window, N-version’s coverage of logic-error is effectively 1.0 with 0 false positives — so the revival model should treat N-version as high-coverage, low-residual-once-triggered, but its effective benefit = cov × λ, and λ collapses on easy tasks. This is the model-conditioned-λ refinement from the frontier run, now with a number.

Verdict

Not a clean revival demonstration (problems too easy → λ≈0 for both models), but the mechanism is validated in the one window where λ>0 appeared (perfect catch, zero false positives), and it revealed model uncertainty a single-shot hides. The missing lever is harder problems, not a weaker model.

Caveats

k=2 pilot; single weak model; 6 problems all within single-shot competence (0% baseline everywhere); AI-version diversity via framing/temperature is a weak independence proxy. Mechanism/direction check, not a powered result.

Author harder problems (edge-dense enough that a single competent generation ships defects → λ>0 even on Sonnet), then re-run NX at k≥3 across the capacity ladder. The prediction: catch rises with λ, and λ rises as problems get harder and models get weaker — the two axes of the revival surface.