Phase 6.3: add the LoRA #2 mix builder and refine the linter

07_build_lora2_mix.py assembles train_lora2.jsonl (900), eval_lora2.jsonl
and calibration_v2.jsonl from the new Penpot seeds plus a filtered replay
sample of data/train.jsonl. It never writes data/train.jsonl or
data/eval.jsonl: those are the provenance of the model in production and
the gate 1 baseline, and regenerating them is not idempotent anyway, since
stratified_split shuffles one RNG over the concatenated list, so touching
the penpot bucket reshuffles every other bucket's split too.

Two things worth flagging in the mix:

The 45 "corrected penpot basics" the plan lists inside the replay portion
do not come from data/train.jsonl. 21 of its 41 penpot seeds teach
findShapeById(page, id) and 5 use shape.layout, so sampling that bucket
would re-teach the exact bug this phase removes; the forbidden-pattern
filter would drop them anyway. They come from the new corpus instead. This
is a conscious deviation from the plan text and is recorded in the
docstring.

Variation comes only from hand-written meta.paraphrases, never from
automatic value substitution. That is the v1 lesson: perturb_value rewrote
only tool_calls.arguments and left the tool results and the final answer
saying something else, producing 30 self-contradictory examples. A
perturbed Penpot code payload is just broken code.

Linter fixes, both false positives found by running it against the real
corpus:
- flex evidence for a bare appendChild is now scoped to the whole seed
  rather than the single payload. A multi-call seed builds the flex board
  in call one and stashes helpers in storage, so by the time call two does
  main.appendChild(...) neither addFlexLayout( nor .flex appears in that
  payload. The old scope flagged exactly the storage-persistence pattern
  that execute_code's own description asks for.
- a grey hex is a problem when it is applied, not when it is searched for.
  The repair seeds have to name the greys they are about to replace, so
  greys are allowed in that group inside a comparison context.
This commit is contained in:
2026-07-30 17:11:36 +00:00
parent 63da20c031
commit 19eb50f351
2 changed files with 411 additions and 9 deletions
+29 -9
View File
@@ -155,12 +155,22 @@ FORBIDDEN = [
# `appendChild` solo es correcto sobre un board con flex (o `grid.appendChild(s, r, c)`). Sobre un
# padre sin layout hay que usar insertChild. No se puede resolver estaticamente en general, asi que
# la heuristica es: si el payload usa `X.appendChild(` con un solo argumento, tiene que haber
# evidencia de flex en el mismo payload.
# la heuristica es exigir evidencia de flex.
#
# El alcance de esa evidencia es EL SEED ENTERO, no el payload. Un seed multi-llamada arma el board
# con flex en la primera llamada y guarda los helpers en `storage`; para cuando la segunda llamada
# hace `main.appendChild(...)`, ni `addFlexLayout(` ni `.flex` aparecen en ese payload. Buscar solo
# dentro del payload marcaba como error justo el patron de persistencia en `storage` que la
# descripcion de `execute_code` pide explicitamente.
APPEND_ONE_ARG_RE = re.compile(r"(\w+)\s*\.\s*appendChild\s*\(\s*[^,()]+\s*\)")
FLEX_EVIDENCE_RE = re.compile(r"addFlexLayout\s*\(|\.flex\b")
GRID_APPEND_RE = re.compile(r"\.grid\s*\.\s*appendChild\s*\(")
# Un gris en el `code` es un problema cuando se APLICA, no cuando se BUSCA. Los seeds de reparacion
# del grupo B8 tienen que poder detectar los grises que van a reemplazar, y para eso necesitan
# nombrarlos. Se permite unicamente en ese grupo y solo en contexto de comparacion/pertenencia.
GREY_COMPARISON_RE = re.compile(r"===|==|!==|!=|\.includes\s*\(|\.indexOf\s*\(|\.some\s*\(|\.has\s*\(")
def is_placeholder_grey(hex_str):
"""R2: saturacion <= 10 Y 100 <= max <= 220. Los casi-negros y casi-blancos pasan."""
@@ -420,6 +430,9 @@ def main():
seen = True
return False
seed_code_blob = "\n".join(code for _, code in ordered_calls)
seed_has_flex = bool(FLEX_EVIDENCE_RE.search(seed_code_blob))
code_by_call = {}
for m in msgs:
for tc in m.get("tool_calls") or []:
@@ -460,7 +473,7 @@ def main():
problems.append(f"{tag}: [{pname}] ...{frag}...\n -> {why}")
# appendChild sobre receptor sin evidencia de flex
if (APPEND_ONE_ARG_RE.search(code) and not FLEX_EVIDENCE_RE.search(code)
if (APPEND_ONE_ARG_RE.search(code) and not seed_has_flex
and not GRID_APPEND_RE.search(code)):
problems.append(
f"{tag}: usa `X.appendChild(shape)` sin ninguna evidencia de flex en el "
@@ -478,13 +491,20 @@ def main():
f"-- el servidor lo prohibe explicitamente (llega duplicado)"
)
# R2: grises de placeholder en el codigo (nunca permitidos, ni en B8)
for hx in HEX_RE.findall(code):
if is_placeholder_grey(hx):
# R2: grises de placeholder en el codigo. Un gris que se APLICA es siempre un
# error; uno que se BUSCA es legitimo, y solo en el grupo de reparacion, que
# necesita nombrar los grises que va a reemplazar.
for code_line in code.split("\n"):
for hx in HEX_RE.findall(code_line):
if not is_placeholder_grey(hx):
continue
if grupo == GREY_INPUT_GROUP and GREY_COMPARISON_RE.search(code_line):
continue
problems.append(
f"{tag}: gris de placeholder {hx} en el `code` (invariante R2). "
f"El gris solo puede aparecer como INPUT en los tool results del "
f"grupo {GREY_INPUT_GROUP}."
f"{tag}: gris de placeholder {hx} aplicado en el `code` "
f"(invariante R2). El gris solo puede aparecer como INPUT en los "
f"tool results del grupo {GREY_INPUT_GROUP}, o nombrado en una "
f"comparacion dentro de ese mismo grupo para detectarlo."
)
# cobertura