Skip to content

Planned before any planted quote-style result is read: the exchange rate on a code-level convention

# research(2026-10). δ is carried from exact match to s at the rate of the adapters' gain over the base model on each scale (0.4372 / 0.2384 over 92 own-half development arms, 2026-10-09), and most of that gain is the adapter learning the task's output format, which says little about the rate for small differences between adapters (Stage 1 review panel, item 15). This measures the rate on a convention of the kind the study asks about, fixed here before any run:

  1. The convention: each half of a partition run writes its Python refinements in its own quote style, half a single and half b double, on plain literals only (planted.to_single, to_double: literals found by Python's tokenizer, single-line, holding no quote or backslash, never prefixed or inside an f-string; a fragment that does not tokenize is left alone), in every window (placebo_corpus.py --plant-convention quotes, jobs PLANT_CONVENTION=quotes, tagged -plantquotes<f>). In PHP or shell the quotes differ in meaning, so only .py rows are eligible, about one in twelve of OpenStack's refinements on each side; Wikimedia has too few, so the rate is read on OpenStack and stands for both organizations, as the registered rate does.
  2. The runs: OpenStack's first ten admissible splits at f = 1.0, and its first five at f = 0.5, each scored by likelihood_score.py.
  3. The rate (planted_exchange.py): on each half's held-out examples the plant changed, own minus sibling adapter in s and in exact match; the rate is the summed gain in s over the summed gain in exact match across every half of the ten f = 1.0 runs, with a 95% percentile interval from resampling runs. The five f = 0.5 runs give the same rate again, as a check that it does not depend on the dose.
  4. The rule: if the f = 1.0 interval excludes the registered rate 1.834, δ is carried at the planted rate, disclosed in the Stage 1 design history as a change made after the development pilots were read, and every bound on δ is recomputed. If it includes 1.834, the registered rate stands and the planted rate is reported beside it as a check. If exact match moves by nothing, no rate is read and the registered rate stands; if more than 1% of resamples move it by nothing, no interval is read and the registered rate stands. A Python fragment the tokenizer cannot read (a hunk cut inside a call or a string) is never planted, so the planted examples lean toward whole statements, and each half's realised share is reported.