From "asymmetry" to "load-bearing channel" — refining Finding 1¶

This page documents the interpretive history of Finding 1. The headline moved from a "negative effect is larger than the positive effect" reading to a sharper "the preamble channel can push output below the no-preamble baseline at all" reading. The shift matters because the two readings license different actions for practitioners.

The original asymmetry reading¶

The v2 main run produced two effects vs none that, on first inspection, looked like an asymmetry:

Read naively, the bad preamble hurt by 0.060, the good preamble helped by 0.046, and the difference suggested the channel was more efficient at damaging output than at improving it.

The naive interpretation was: "negative preambles are more dangerous than positive preambles are useful." On that reading, the headline takeaway becomes "be careful what you write" — a risk-management framing.

Why that reading is misleading¶

The asymmetry framing has two problems.

First, it confuses sign with strength. What the v2 data actually demonstrates is not that negatives are stronger than positives. It demonstrates that the preamble channel reaches below baseline. If preambles were inert or weakly additive, you could not write one that moves output below what the model produces with no system prompt at all. The fact that negative_control lands at −0.060 — strictly below none — is the demonstration. The magnitude is secondary; the sign and direction are the load-bearing observation.

Second, the magnitude gap is partly an artifact of ceiling effects. none already scores near the top of several rubric dimensions (documentation, type hints, organization on tasks where the model defaults to writing them). A preamble that targets those same dimensions has limited headroom to push them higher — the dimension is near-ceilinged. A preamble that pushes them down has the full range of the scale to work with. The asymmetry, where it appears, is partly a mechanical consequence of where baseline sits on the rubric, not evidence that the channel is structurally biased toward damage.

The refined reading: a load-bearing channel¶

The current Finding 1 framing reads:

The preamble channel is powerful enough that content choices move outputs measurably above and below a no-system-prompt baseline. The cleanest evidence is the degradation case: framings like "junior developer", "still learning Python", or "don't worry too much about style" produce worse code than supplying no system prompt at all. That's the sharp test — the model isn't just amplified by good preambles or unaffected by bad ones, it's actively responsive to content in both directions. Preambles are not decorative; they steer.

This is the same data with a different load-bearing claim. The claim is not "negatives hurt more than positives help". The claim is none is not a floor. The channel carries signal in both directions, with sufficient magnitude that content choices on the bad end of the spectrum demonstrate the channel exists at all.

What this licenses¶

Two practical implications follow from the load-bearing reframe that the asymmetry framing obscured.

1. Treat every preamble change as bidirectional. A new clause, a reworded directive, or an added persona is not assumed to be neutral or positive. If preambles can push below baseline, then any change to a working preamble is a candidate to make output worse, not just to leave it unchanged. The v2 evidence that supports this is none not being a floor — that is the load-bearing fact, not the relative magnitudes of long_directive vs negative_control.

2. The realistic failure modes are smaller in magnitude than the synthetic probe but in the same direction. negative_control was a deliberately blunt synthetic probe ("junior developer still learning Python") whose purpose was to prove the channel can move output below baseline at all. Production prompts rarely contain language that direct. The realistic failure modes — rubric mismatch (see Finding 2) and verbose dilution (see Finding 3) — produce the same directional effect at smaller magnitude. The probe is upper-bound evidence for a channel that operates with the same sign in production conditions.

The secondary observation, properly bounded¶

The negative-vs-positive magnitude difference is real, but it does not support the strong asymmetry claim:

• Real component. negative_control shifts the model's stylistic register downward — "learning Python", "don't worry too much about style" — and judges (blind to the preamble) score the resulting code lower on the dimensions that register affects. This is a content-level effect, not a measurement artifact. See the judge protocol methodology.
• Ceiling-bounded component. The dimensions where none already scores near the top have less room to lift than to drop. The apparent "negative bigger than positive" gap is partly a consequence of where the no-preamble baseline lands on the rubric.

The two components are not separately quantified in v2; a v3 design with a more saturated mid-range rubric could in principle decompose them. For practical purposes, the takeaway is that the asymmetry should not be treated as a structural fact about preambles — it is partly real and partly a property of where none sits on this particular rubric.

Why the reframe matters¶

The asymmetry reading risks two errors. The first is over-claiming: "preambles are more dangerous than helpful" implies a defensive posture (avoid changes, keep preambles minimal) that the data does not support. The second is under-claiming the mechanism: the asymmetry framing treats the negative as the special case to be explained, when in fact both directions are evidence of the same load-bearing channel, just with different headroom on this rubric.

The load-bearing channel reading collapses the two effects into a single claim — none is not a floor — and lets the per-direction magnitudes be whatever they are, without making the magnitude gap do interpretive work it cannot support.

Sources¶

• README.md Finding 1 — current framing.
• preamble_quality_experiment_v2/CONCLUSIONS.md §"Mixed-effects models", interpretation paragraph: "The sharpest reading is not a positive/negative asymmetry but a demonstration that the preamble channel is load-bearing: none is not a floor — negative_control pushes CQS-craft below what the model produces with no system prompt at all. … The negative effect being larger in magnitude than the positive effect is a secondary observation, partly real and partly bounded by ceiling effects on rubric dimensions where none already scores near the top."
• preamble_quality_experiment_v2/CONCLUSIONS.md §"Verdict", bullet 2.
• See Finding 1 for the evidence summary in finding form, and the attention-allocation mechanism for the underlying mechanism story.