Glossary¶
Short definitions for the recurring terms in this investigation. Each entry links to the reference page where the term is defined in full.
CQS-craft¶
The v2 Composite Quality Score (craft-only). A weighted bounded composite on [0, 1]:
Computed per sample after self-judgment exclusion. The headline statistic for the v2 main run. Pre-registered weights are (0.45, 0.45, 0.10); the weight-sensitivity panel confirms the headline does not depend on this choice.
The "craft" qualifier marks the deliberate exclusion of static-analysis metrics — v1 showed those are flat across preambles, so they are reported only as a diagnostic panel.
Rubric severity¶
The 0–5 graded measure that each rubric dimension is scored on. Not a present-vs-absent flag — see the calibration anchor: severity 0 is reserved for "genuinely exemplary," severity 1–2 is the realistic baseline for most algorithmic code, severity 4–5 means "pervasive or material defect."
Always-on dimension¶
A rubric dimension that is scored on every sample, regardless of task. Nine of the 11 rubric dimensions are always-on: data_structure_choice, algorithm_correctness, error_handling_inconsistency, api_ergonomics, abstraction_miscalibration, code_organization, type_hint_gap, edge_case_gap, documentation_appropriateness. See rubric.
Conditionally-N/A dimension¶
A rubric dimension that judges return null for when the dimension does not apply to the code or task. Two such dimensions in v2: concurrency_safety (null when the task does not involve concurrency) and example_quality (null when no examples are requested). N/A judgments are excluded from that judge's per-sample mean severity. See rubric.
Rubric overlap density¶
The fraction of rubric dimensions that a preamble explicitly enumerates. long_directive's 12 clauses name 7 of 9 always-on dimensions — a high overlap density — and the confound probes show that this overlap drives most of its CQS lift over none. The refined v2 mechanism reads preamble effects as proportional to preamble–rubric overlap density, not to "expertness" or preamble length.
Attention-allocation¶
The post-confound-probe reading of the v2 mechanism: preambles direct the model's finite craft-attention budget toward whatever dimensions they enumerate, at the cost of other behaviors. CQS-craft is real and reproducible, but rubric-dependent — a preamble's lift over none is proportional to the overlap between (what the preamble enumerates) and (what the rubric measures). See Finding 2 and CONCLUSIONS.md §"Refined hypothesis: H-attention-allocation".
Self-judgment exclusion¶
The rule that a judgment is dropped from primary CQS whenever the judge model and subject model share a provider-prefix family (model.split("/", 1)[0]). So deepseek/deepseek-v3.2 judged by deepseek/deepseek-v4-flash counts as a self-judgment even though they are different models. Self-judgments are retained in the dataset for the F3 self-vs-cross hygiene stratification report, but never enter the headline CQS. See models and judge protocol.
ETA (empirical-test-agreed verdict)¶
A debate-scorecard verdict used in the ml-lab investigation cycle: the critic and defender agree the question cannot be resolved by argument alone, so it is deferred to an empirical pre-flight probe whose outcome is binding. In v2, findings F2, F4, and F5 closed via ETA — the trap task was dropped after Phase B, the anchored-vs-unanchored prompt confound was resolved by a Phase C re-probe, and the rubric was redesigned after Phase D found 0/11 dimensions active on modern code. See CONCLUSIONS.md §"Pre-flight debate scorecard".
defense_wins¶
A debate-scorecard verdict: the defender's position prevailed at debate close without requiring an empirical probe. In v2, findings F3, F6, and F7 closed defense_wins — the v1 self-vs-cross protocol was preserved, and two minor design concerns were settled in debate.
critique_wins¶
A debate-scorecard verdict: the critic's position prevailed at debate close without requiring an empirical probe. In v2, finding F1 closed critique_wins — the judge OOR-clamp bug was conceded by the defender at severity 9 (FATAL) and fixed before main run (drop-not-clip with a structured OOR log; main-run OOR rate = 0). See CONCLUSIONS.md §"Pre-flight debate scorecard".