[braboj]

Round 5c was scored on Python 3.14, which refused three builds. Scored again on 3.12, the spec's version, it reads differently. Both new readings are labelled post-hoc; the registered report's verdicts stand.

File	What it is
2026-11-01-efficacy.md	Registered report, plus one paragraph linking the two readings below
2026-11-02-efficacy*.md, radars	5c on Python 3.12, read by 5c's own rules
2026-11-02-efficacy-posthoc.md, radars	The same scores, read by round 5d's rules

How the verdicts move on 3.12:

Question	Registered (3.14)	Post-hoc (3.12)
hybrid against no file	Worse	Better
inline against no file	No clear gain	Better
reference against no file	Worse	Worse
skills against no file	Worse	Worse
H3, a reference alone delivers less	Not settled	Yes

• hybrid-3, inline-5 and skills-4 install on 3.12 and were judged for the first time (9 judgings, evidence found 100 %).
• Archived and verified in braboj/tariff-runs as rounds/2026-11-02-round-5c-py312b.

Tracked in solid-ai-dev/solid-ai-benchmark#103.

🤖 Generated with Claude Code

https://claude.ai/code/session_01DpT5xPFWEQr1bipxbbKBRF

[braboj]

Codex review, round 1: one finding.

Severity	Finding	Outcome
P2	The post-hoc radar values do not match "bound nearest zero over the row's full scale"	By design: the full scales are round 5d's registered ones (round-5d.md 13d.8), 1 point for a judge score and 20 points for coverage, not 4 and 100. inline's Tests axis: (0.578 / 1 + 14.567 / 20) / 2 = 0.65, as printed

[braboj]

Codex review, round 2: the same finding again, not a new one. Now fixed rather than only answered: the caption under the post-hoc radar names each row's full scale (solid-ai-dev/solid-ai-benchmark#114).

[braboj]

Codex review, round 3: no findings. The loop ends here.