Finding 4 — No "alignment vs capability" split, just preamble–evaluator overlap¶
Claim. v1 inherited a "preambles change alignment-tunable craft but not pretraining-locked capability" framing from PRISM (USC 2026). v2's probes refined this: the proximate predictor of which dimensions move under preamble is whether the preamble enumerates them, not whether they're "craft" or "capability" in some structural sense.
Evidence. In the v2 main run, 7 of 9 always-on rubric dimensions moved with preamble (KW p < 10⁻⁴), and 2 didn't (algorithm_correctness p = 0.26; data_structure_choice p = 0.39). Looking at long_directive's clause list:
long_directive clause |
Rubric dimension it names |
|---|---|
| (3) defensive programming, validate inputs, handle edge cases | edge_case_gap, error_handling_inconsistency |
| (5) comments why not what | documentation_appropriateness |
| (7) concurrency / thread-safety explicit | concurrency_safety |
| (8) composition over inheritance | code_organization |
| (9) side effects + I/O boundaries explicit | code_organization |
| (11) log errors at right severity, never swallow | error_handling_inconsistency |
| (12) docstring public interfaces | documentation_appropriateness, type_hint_gap |
| (2) appropriate abstraction | abstraction_miscalibration |
The 7 dimensions that move are exactly the 7 enumerated by long_directive. The 2 that don't move (algorithm_correctness, data_structure_choice) are exactly the 2 not enumerated in any v2 preamble. The pattern fits both the original "alignment/capability split" and the simpler "enumerated/not-enumerated" reading. Probe A breaks the tie: a preamble that doesn't enumerate the 7 craft dimensions but is otherwise expert-toned doesn't lift them — it actively suppresses them. The proximate predictor is enumeration.
Action. Don't assume any dimension is "preamble-immovable" without testing it. If you care about algorithmic correctness, enumerate it in your preamble — it may move (v2 didn't test this; an explicit-correctness probe is plausibly worth running for your domain).
Related work. F4 is the headline form of the same PRISM refinement called out in F2 — the v2 update at the end of Related work § "Personas help style, not substance" discusses the implication: a preamble that explicitly enumerated correctness could in principle move accuracy too, against a strict reading of PRISM. F4's evidence is the within-rubric version of that argument.