fix(short): una nota de duración vale una reescritura, no dos
Build & Deploy ResearchOwl / build-and-push (push) Successful in 39s

`editorial_notes` documentaba "se comenta una vez y, si insiste, se renderiza
igual", pero el bucle reintentaba dos veces: las sesiones 166, 167 y 168
gastaron los tres intentos y las tres acabaron renderizando un spec que seguía
pasándose. Un spec que ya cumple el contrato es renderizable; los intentos que
quedan son para el contrato, que sí es binario.

Y de dos specs válidos se guarda el que menos se sale del objetivo, no el
primero. Obedecer a medias es obedecer: antes, una reescritura que bajaba de
70 s a 50 s se tiraba entera y salía el largo.

El prompt, que era la mitad que faltaba. Decirle "20-45 segundos" no le sirve
de nada: la duración real no está escrita en ninguna parte del spec, sale de
sumar plano a plano el mayor entre lo declarado y lo que tarda la voz, y eso no
lo puede calcular quien no sabe a qué velocidad se le lee. Ahora lleva el
ritmo (2,8 palabras por segundo), la cuenta por plano, un presupuesto de 80
palabras de narración, y dos topes que salen del propio ejemplo en vez de estar
inventados: 12 palabras por línea (tope 18) y 6 s por plano. El anterior — "una
línea de menos de 25 palabras" — no describía nada que el canal hubiera
publicado, y el modelo escribía líneas de 25 y 27 sin saltarse ninguna regla.

Medido con cinco generaciones reales sobre el material de la sesión 168:
antes 3 intentos siempre, con el vídeo fuera del objetivo (47,5 s y 45,4 s).
Ahora 2, y el borrador cae dentro: 39,8 / 37,3 s. En dos de las cinco el
reintento ya no fue por duración sino por un fallo de contrato.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
ChemaVX
2026-08-12 21:37:17 +00:00
co-authored by Claude Opus 5
parent 9bfa0fdac2
commit 91ceb3b1ca
2 changed files with 207 additions and 25 deletions
+109 -19
View File
@@ -22,7 +22,8 @@ from typing import Any, Awaitable, Callable, Optional
import structlog
from src.generator.spec_contract import (
SpecInvalid, describe_templates, editorial_notes, validate_spec,
SpecInvalid, describe_templates, editorial_notes, estimated_duration,
validate_spec, NARRATION_WORDS_PER_SECOND,
TARGET_MAX_DURATION, TARGET_MIN_DURATION,
)
@@ -32,6 +33,36 @@ logger = structlog.get_logger()
#: que arreglar es el prompt, no este número.
MAX_ATTEMPTS = 3
#: Reescrituras que se gastan en una nota editorial, no en un fallo de contrato.
#: UNA. Un spec que ya cumple el contrato y sólo se pasa de duración es
#: renderizable: la segunda reescritura no compraba un Short mejor, compraba una
#: generación más. Medido sobre las sesiones 166, 167 y 168 — las tres gastaron
#: los tres intentos por duración y las tres acabaron renderizando un spec que
#: seguía pasándose. Los intentos que quedan son para el contrato, que sí es
#: binario. Ver `_how_to_trim` en `spec_contract`: si la nota no se obedece a la
#: primera, lo que hay que arreglar es la nota.
NOTE_ATTEMPTS = 1
#: Cuánta narración cabe en un Short entero. Comprobación cruzada de la regla
#: de arriba, en la unidad que el modelo escribe: 80 palabras son unos 29 s de
#: voz, y con los respiros y algún plano mudo eso deja el vídeo cerca de 40 s.
#: El ejemplo de referencia habla 74. La sesión 167 habló 97 y salió a 47,5 s.
NARRATION_WORD_BUDGET = 80
#: Lo que mide una línea. No es preferencia de estilo: son las líneas del
#: ejemplo (11, 12, 10, 14, 12, 15 palabras). El tope anterior — "menos de 25" —
#: no describía nada que el canal hubiera publicado, y el modelo escribió líneas
#: de 25 y 27 palabras sin saltarse ninguna regla.
NARRATION_WORDS_PER_LINE = 12
NARRATION_WORDS_PER_LINE_MAX = 18
#: Lo que dura un plano como mucho. También del ejemplo (6,0 s el más largo,
#: 5,6 de media). Sin este tope el modelo declaraba 42 s en seis planos de siete
#: segundos y LUEGO les colgaba la narración encima: el primer borrador salía a
#: 50 s las tres veces que se midió, y hacía falta una reescritura entera para
#: bajarlo.
MAX_SHOT_DURATION = 6.0
EXAMPLE_PATH = Path(__file__).parent / "examples" / "jal1628.json"
__all__ = ["ShortSpecWriter", "SpecResult", "SpecWriteFailed", "NARRATIVE_SHAPES"]
@@ -84,10 +115,18 @@ exists, and a prop name that is not listed is a parse error, not a nuance.
# 3. Rules
- Total duration 20-45 seconds. The contract allows 180; that is a ceiling, \
not a target. Aim for {target_min:.0f}-{target_max:.0f}.
- Typically 6-9 shots. Give a shot the seconds its content needs to be read: \
a card with four rows needs longer than a headline.
- **Total duration {target_min:.0f}-{target_max:.0f} seconds, and the voice is \
what decides it, not the durations you declare.** The contract allows 180; that \
is a ceiling, not a target. A narrated shot runs as long as its line takes to \
say — the renderer never cuts the voice off, it grows the shot — so the whole \
video is really about {word_budget} words of narration and no more. That is the \
number to hold: **count the words of every `narration` you write, and stop at \
{word_budget}.** Section 3b has the arithmetic behind it.
- Typically 6-9 shots, and **none of them longer than {max_shot:.0f} seconds** \
— that is the example's longest, and its average is 5.6. Give a shot the \
seconds its content needs to be read: a card with four rows needs longer than a \
headline. A {max_shot:.0f}-second shot with a short line on it is not a \
generous shot, it is a shot the viewer has already finished reading.
- Every string is drawn as given. Write them the way they should appear: \
SHORT, UPPERCASE, no trailing punctuation. A headline is 2-5 words.
- **"CABE ~N caracteres dibujados" is a width, and it is the one limit nothing \
@@ -116,9 +155,12 @@ aloud by the renderer and burned in as captions. Write it for the ear.
- **The first line is the whole hook.** Two seconds decide whether anyone \
watches the rest, and the opening shot's narration is those two seconds. Lead \
with the strangest true thing you have, not with a preamble.
- Keep a line under 25 words. Long sentences lose the listener and stretch the \
shot; the renderer will not cut your voice off, it will make the shot longer \
instead, and a Short that drifts past 45 seconds is a Short people leave.
- **Keep a line to {words_per_line} words, hard stop at {words_per_line_max}.** \
That is the example's own average, and it is not a style preference: a 25-word \
line is three seconds of your whole budget spent on one shot. Long sentences lose the listener and \
stretch the shot; the renderer will not cut your voice off, it will make the \
shot longer instead, and a Short that drifts past {target_max:.0f} seconds is a \
Short people leave.
- **Do not read the screen aloud.** The captions already show your words and \
the template already shows its own. If the shot draws "35,000 FT", the voice \
says what that altitude meant, not the number again.
@@ -132,10 +174,22 @@ silent exactly the two that draw a quotation, where the voice would only be \
competing with words already on the frame. Chosen silence is an edit; a spec \
with one narrated shot out of eight is not a Short with a voice, it is a Short \
that forgot to speak.
- Narration costs seconds. A shot is never cut short to fit the voice — it \
grows instead — so a line that needs six seconds in a four-second shot pushes \
your whole total past the target. Write the line, then give the shot the time \
the line actually takes.
- **Give every narrated shot enough time for its own line, and work it out \
rather than guessing.** The voice reads about {words_per_second:.1f} words a \
second and pauses a quarter second at every full stop, so:
duration ≥ words ÷ {words_per_second:.1f} + half a second
A twelve-word line needs five seconds; give that shot 5.0, not 4.0. This is \
the one rule that makes your own arithmetic true: a shot runs for the LONGER of \
its declared duration and its line — never shorter, the voice is never cut off \
— so a shot that declares less than its line silently grows, and the video ends \
up longer than the durations you wrote. Hold this rule and the total you \
declare IS the video's length; break it once and nothing you counted means \
anything.
- As a cross-check, all the narration in the spec together should come to about \
{word_budget} words. The example below speaks 74. A spec that spoke 97 rendered \
at 47.5 seconds and had to be cut.
- Everything in section 4 applies to narration word for word. It is prose you \
compose rather than a label you copy, which makes it the easiest place to \
slip in a figure no source gave you — and it is checked exactly like the rest.
@@ -359,6 +413,25 @@ def _format_notes(notes: list[str]) -> str:
"Return the adjusted JSON object.")
def _off_target(spec: dict) -> float:
"""Segundos fuera de la ventana editorial. 0 = dentro."""
total = estimated_duration(spec)
return max(0.0, TARGET_MIN_DURATION - total, total - TARGET_MAX_DURATION)
def _closer_to_target(a: Optional[SpecResult], b: SpecResult) -> SpecResult:
"""De dos specs válidos, el que menos se sale del objetivo.
Antes se guardaba el PRIMERO válido y punto, con lo que una reescritura que
obedecía la nota a medias — 53 s en vez de 58 — se tiraba entera y salía el
largo. El empate se lo lleva el anterior: sin razón para cambiar, no se
cambia.
"""
if a is None:
return b
return a if _off_target(a.spec) <= _off_target(b.spec) else b
#: (system, prompt) -> texto del modelo.
LLMCall = Callable[[str, str], Awaitable[str]]
@@ -388,6 +461,11 @@ class ShortSpecWriter:
domain=domain,
target_min=TARGET_MIN_DURATION,
target_max=TARGET_MAX_DURATION,
words_per_second=NARRATION_WORDS_PER_SECOND,
word_budget=NARRATION_WORD_BUDGET,
words_per_line=NARRATION_WORDS_PER_LINE,
words_per_line_max=NARRATION_WORDS_PER_LINE_MAX,
max_shot=MAX_SHOT_DURATION,
example=_load_example(),
article=article,
context=context,
@@ -404,6 +482,8 @@ class ShortSpecWriter:
#: Un spec que cumple el contrato pero se pasa de duración. Se guarda
#: para que un intento posterior peor no lo tire: es renderizable.
best: Optional[SpecResult] = None
#: Reescrituras ya gastadas en notas editoriales.
note_rounds = 0
for attempt in range(1, MAX_ATTEMPTS + 1):
if on_progress and attempt > 1:
@@ -440,23 +520,33 @@ class ShortSpecWriter:
notes = editorial_notes(spec)
result = SpecResult(spec=spec, attempts=attempt, notes=notes,
history=list(history))
if notes and attempt < MAX_ATTEMPTS:
# Nota editorial, no violación del contrato: se comenta una vez
# y, si insiste, se renderiza igual.
best = best or result
if not notes:
logger.info("short spec válido", attempts=attempt,
shots=len(spec.get("shots", [])), notes=0)
return result
best = _closer_to_target(best, result)
if note_rounds < NOTE_ATTEMPTS and attempt < MAX_ATTEMPTS:
# Nota editorial, no violación del contrato: se comenta y, si
# insiste, se renderiza el intento que menos se pase.
note_rounds += 1
history.append(notes)
feedback = _format_notes(notes)
continue
logger.info("short spec válido", attempts=attempt,
shots=len(spec.get("shots", [])), notes=len(notes))
return result
logger.info("short spec válido pero fuera de objetivo",
attempts=attempt, shots=len(best.spec.get("shots", [])),
off_target=round(_off_target(best.spec), 1))
best.attempts = attempt
best.history = history
return best
if best is not None:
# Un intento anterior sí cumplía el contrato. Vale más un Short
# largo que ningún Short.
logger.info("short spec: se recupera el intento válido anterior",
attempts=MAX_ATTEMPTS, notes=best.notes)
best.attempts = MAX_ATTEMPTS
best.history = history
return best