combo-script-alignment · generalization

v1 no-lookahead 真能泛化吗：6 视频 × 5 reps × 双口径

之前只在 1 个视频(加油站手游 sI4W7c)上测过：给生成 prompt 加一句「只描述这一刻画面里已发生的事，别提前剧透后面镜头」，就把「词跑在画面前」治好了(当时 v1=3/3)。但那是 n=2–3、单一指标、单个 case 的结果。这页把同一改动放到 6 个不同视频、每臂跑 5 次、用两套口径量，看它到底稳不稳。

做法:模型固定 gemini-3.1-pro-preview，温度 1，两臂只差那一段 prompt(v0=线上的，v1=加了 no-lookahead 约束的)，desc/视频/其余完全一样——prompt 是唯一变量。

一眼看结论

22 / 22
单转场口径 v0 vs v1(打平)
8 / 12
全里程碑口径 v0 vs v1
2/6
单转场下 v1 更好的视频
2/5
sI4W7c v1 对齐(原称 3/3)

两条结论先写在前面，不美化:

  • 单转场口径(和原研究可比):v1 没有泛化优势。总对齐 v0 22/30 = v1 22/30，打平。更扎心的是连原 case sI4W7c 都反转了——5 个新 rep 下 v0 对齐 4/5、v1 只 2/5。原来的 v1=3/3 基本是小样本运气。
  • 全里程碑口径(更严，抓全片任何超前):v1 有温和优势但不彻底。v0 8/30 → v1 12/30，6 个视频里 v1 更好 4、持平 1、更差 1。两套口径都改善不了 z5cXvN，且都在某些视频上让 v1 翻车。

两张聚合表

单转场口径 — 每视频只认那个最显著中段转场，±1.5s

视频v0 对齐v1 对齐
z5cXvN5/54/5v1 差
qrIwwY3/55/5v1 好
bdE1hS2/53/5v1 好
rYEP7C5/55/5持平
sI4W7c4/52/5v1 差
c12xUC3/53/5持平
合计22/3022/30打平

单转场(可比原研究) · v1 比 v0：好 2 / 持平 2 / 差 2 个视频

全里程碑口径 — 每视频 2–5 个里程碑，任一首现前 >1s 提及即超前

视频v0 对齐v1 对齐
z5cXvN0/50/5持平
qrIwwY0/51/5v1 好
bdE1hS5/54/5v1 差
rYEP7C1/53/5v1 好
sI4W7c0/51/5v1 好
c12xUC2/53/5v1 好
合计8/3012/30v1 总体更优

全里程碑(更严) · v1 比 v0：好 4 / 持平 1 / 差 1 个视频

口径校准:先证明尺子是准的

单转场口径不是事后凑的——先拿它去量原研究当时那批 sI4W7c 出词(回海洋 vs 上船 ~29s)，看能不能复现原结论。结果:这把尺子定位「回海洋」那句的时间和原研究人工标的 return_start 完全吻合(24.0 / 28.0 / 29.5 / 30.0 / 29.5)，方向也复现(v0 差、v1 好)。尺子可信，再 blind 套到 6 个视频。

原研究那批 reps，用本页尺子重判(±1.5s):

  • v0 rep1:mention 24.0s → premature
  • v0 rep2:mention 28.0s → aligned_borderline
  • v1 rep1:mention 29.5s → aligned
  • v1 rep2:mention 30.0s → aligned
  • v1 rep3:mention 29.5s → aligned

注:v0 rep2 的 28.0s 在 ±1.5s 容差下算 aligned(原研究用硬阈值 29.0 算它超前)。方向一致:v1 仍明显优于 v0。

反例:v1 也会超前(就在原 case 上)

sI4W7c · v1 · rep1 — 玩家 29s 才回海洋，但脚本 18s 就喊「往回走」

[18.0–29.0] “Drop your cash on the pads to unlock new areas! I need more money for the car wash, so let's head back out.”
「head back out」(回海洋)在 18s 就说了，画面 29s 才回去——超前 11 秒。no-lookahead 约束这一 rep 没拦住。这正是 temp=1 下 v1 不稳的真实样本。

逐视频:左边播、右边对词

右栏展示样例 rep1 的完整出词。蓝色高亮 = 该句超前;「单转场超前」按可比口径，「里程碑超前」按更严口径。两套口径各自的 5-rep 对齐结果用圆点标在每臂标题(● 对齐 / ○ 超前)。

z5cXvN · 59.0s

单转场 v0/v15/5 · 4/5
里程碑 v0/v10/5 · 0/5
单转场目标(盲选):45.0s Tiles Survive! gameplay
里程碑(更严口径锚点):
  • 4.0s Gold grinding begins
  • 14.0s Gold bar production
  • 24.0s Robotic arms added
  • 42.0s Level failed
  • 45.0s Tiles Survive! gameplay

v0 · 线上 prompt

v0单转场 5/5●●●●●里程碑 0/5○○○○○
样例 rep1:
[0.0–4.5]Watch this giant gold block smash through the gate!
[4.5–10.0]Let's upgrade these gears and grind it into pure gold.
[10.0–15.0]Look at it rain nuggets! We are totally rich.
[15.0–21.0]Buy more lanes so the trucks can load up faster.
[21.0–28.0]Wait, it's piling up way too fast! Drivers are getting angry.
[28.0–34.0]Quick, buy the robot arms and upgrade them to max speed!
[34.0–44.0]Oh no, there's a warning! The whole stack is gonna explode! Ah man, I failed.
里程碑超前 +8.0s：提前提到「Level failed」(画面 42.0s)
[44.0–52.0]Let's try the actual game. Build a power plant and connect the water.
[52.0–59.0]The base is powered up! Can you survive? Download Tiles Survive today!

v1 · no-lookahead

v1单转场 4/5○●●●●里程碑 0/5○○○○○
样例 rep1:
[0.0–3.5]Woah, look at this massive gold drop!
[3.5–9.0]Grind it up into nuggets! Let's hit that upgrade button immediately.
[9.0–14.0]Oh yeah, max out those gears! The payout is so satisfying.
[14.0–20.0]Stack those bars! Let's buy more lines to speed up the process.
[20.0–25.0]The trucks are moving, but we need way more capacity!
里程碑超前 +4.0s：提前提到「Robotic arms added」(画面 24.0s)
[25.0–32.0]Add robot arms to load the trains. Keep upgrading them to match the flow!
[32.0–38.0]Wait, a bottleneck! The drivers are getting mad, and it's backing up!
[38.0–44.5]Oh no, we completely overflowed the lot! That was crazy! But watch this instead.
单转场超前 +7.0s：在「Tiles Survive! gameplay」(画面 45.0s)出现前就提到
里程碑超前 +4.0s：提前提到「Level failed」(画面 42.0s)
[44.5–52.0]In Tiles Survive, you actually chop down real resources to build your own survival base!
[52.0–59.0]Power up the crops and expand your territory! Download now and play for free!

qrIwwY · 29.2s

单转场 v0/v13/5 · 5/5
里程碑 v0/v10/5 · 1/5
单转场目标(盲选):16.0s Truck loading area revealed
里程碑(更严口径锚点):
  • 7.5s Wood chipping machine revealed
  • 16.0s Truck loading area revealed
  • 25.0s New conveyor belt area revealed

v0 · 线上 prompt

v0单转场 3/5●●○○●里程碑 0/5○○○○○
样例 rep1:
[0.0–4.5]Watch me build the ultimate lumber empire right now!
[4.5–11.0]First, we strip the logs and slice them up. Look at that stockpile grow!
里程碑超前 +3.0s：提前提到「Wood chipping machine revealed」(画面 7.5s)
[11.0–16.5]Let's upgrade the saw blades! Oh yeah, maximum production speed!
[16.5–21.0]Uh oh, a bottleneck! The delivery trucks are stuck.
[21.0–25.5]Upgrade the belts! Boom, watch that cash flow!
里程碑超前 +4.0s：提前提到「New conveyor belt area revealed」(画面 25.0s)
[25.5–29.2]This is so satisfying! Play for free now!

v1 · no-lookahead

v1单转场 5/5●●●●●里程碑 1/5○○●○○
样例 rep1:
[0.0–4.5]Oh, so satisfying! Watch these logs get perfectly shaved.
[4.5–9.5]Drop them in the shredder, let the chips pile up!
里程碑超前 +3.0s：提前提到「Wood chipping machine revealed」(画面 7.5s)
[9.5–15.5]We need more speed. Boom, upgrade! Look at that massive output.
[15.5–20.5]The delivery trucks are literally backed up at the gate!
[20.5–29.2]Let's invest this cash to open a brand new processing line. Build your own empire, play now!
里程碑超前 +4.5s：提前提到「New conveyor belt area revealed」(画面 25.0s)

bdE1hS · 10.1s

单转场 v0/v12/5 · 3/5
里程碑 v0/v15/5 · 4/5
单转场目标(盲选):5.0s Cake slice separated
里程碑(更严口径锚点):
  • 2.0s Caramel center revealed
  • 5.0s Cake slice separated

v0 · 线上 prompt

v0单转场 2/5●○○○●里程碑 5/5●●●●●
样例 rep1:
[0.0–4.0]Finally, an ASMR slicing game that's actually real.
[4.0–7.5]Look at that caramel ooze! So satisfying!
[7.5–10.1]Download and play for free!

v1 · no-lookahead

v1单转场 3/5●○○●●里程碑 4/5●○●●●
样例 rep1:
[0.0–3.5]I could watch this all day!
[3.5–7.0]Look at that perfect caramel pour.
[7.0–10.1]So satisfying. Play for free now!

rYEP7C · 56.6s

单转场 v0/v15/5 · 5/5
里程碑 v0/v11/5 · 3/5
单转场目标(盲选):14.0s Truck loading area
里程碑(更严口径锚点):
  • 4.0s Log cutting begins
  • 10.0s Chainsaw upgrade
  • 14.0s Truck loading area
  • 42.0s Coastal road and tunnel

v0 · 线上 prompt

v0单转场 5/5●●●●●里程碑 1/5○○○○●
样例 rep1:
[0.0–5.0]Watch this massive log! Let's slice it up and make bank.
里程碑超前 +4.0s：提前提到「Log cutting begins」(画面 4.0s)
[5.0–10.0]Upgrading the saw! Oh yeah, now we're churning out the wood.
里程碑超前 +5.0s：提前提到「Chainsaw upgrade」(画面 10.0s)
[10.0–14.0]Chainsaw unlocked! That is a mountain of logs!
[14.0–20.0]Uh oh, the pile is huge! Let's buy trucks to move it.
[20.0–28.0]Spam that buy button! Go, go, go! We've gotta clear this out before it overflows.
[28.0–35.0]Look at them go! This is so satisfying. Just endless trucks loading up now.
[35.0–43.0]We are finally making a dent! Alright, let's see where they're taking all this wood.
[43.0–49.0]Wait, the tunnel is overflowing! We're losing logs in the water, not good!
[49.0–56.5]Upgrade the processor! Boom, problem solved. Look at those coins rolling in! Download and play now!

v1 · no-lookahead

v1单转场 5/5●●●●●里程碑 3/5○●○●●
样例 rep1:
[0.0–4.0]Check out this massive log! Let's slice it.
里程碑超前 +4.0s：提前提到「Log cutting begins」(画面 4.0s)
[4.0–13.0]Upgrading the saws! Bro, this is so satisfying to watch. We are literally printing money right now!
[13.0–20.0]Look at that huge stack of wood! We need some trucks to transport it all.
[20.0–28.0]Wait, the path is completely blocked. They can't get past! Let me spawn some more.
[28.0–36.0]Come on, move guys! This traffic jam is getting totally out of control. Just keep buying!
[36.0–42.0]Why aren't they moving at all? Let's see what is happening over here.
[42.0–50.0]Oh no, they are completely backed up at the tunnel! The logs are falling into the water.
[50.0–56.5]I'm losing so much cash! Can you fix this? Download and play now!

sI4W7c · 71.5s

单转场 v0/v14/5 · 2/5
里程碑 v0/v10/5 · 1/5
单转场目标(原研究锚点):29.0s Player goes BACK to the ocean / re-enters the boat for the mid-roll oil-cleaning, AFTER the on-land building+selling phase
里程碑(更严口径锚点):
  • 3.0s Gas station built
  • 6.0s Boat cleaning oil
  • 45.0s Car wash unlocked
  • 52.0s Oil pump jack built
  • 69.0s Defeat screen

v0 · 线上 prompt

v0单转场 4/5●●●○●里程碑 0/5○○○○○
样例 rep1:
[0.0–4.5]Whoa, a massive oil spill! We gotta clean this!
[4.5–12.5]Hop in the vacuum boat and start sucking up that black gold. Look at it go!
里程碑超前 +1.5s：提前提到「Boat cleaning oil」(画面 6.0s)
[12.5–20.0]Turn the crude oil into fuel barrels and serve these angry customers. Get that cash!
[20.0–30.0]Look at those stacks of money! Time to expand the station, unlock more pumps, and keep the business booming!
[30.0–38.0]Back to the spill! The more we clean, the more we earn. This is incredibly satisfying!
[38.0–45.0]Dropping off another massive haul! Let's reinvest and build a brand new car wash.
里程碑超前 +7.0s：提前提到「Car wash unlocked」(画面 45.0s)
[45.0–53.0]Oh yeah, the car wash is printing money! Our gas station empire is growing so fast!
[53.0–60.0]Wait, we can build an automated oil rig? Now the barrels are pumping themselves!
[60.0–66.0]Oh no, it's producing way too fast! We just can't keep up!
[66.0–71.5]Everything is overflowing! Can you do better? Download and play now!

v1 · no-lookahead

v1单转场 2/5○●○●○里程碑 1/5○●○○○
样例 rep1:
[0.0–3.2]Dude, this massive offshore oil spill...
[3.2–11.0]...is actually an absolute goldmine! Let's grab the boat, clean it up, and make some bank!
里程碑超前 +2.8s：提前提到「Boat cleaning oil」(画面 6.0s)
[11.0–18.0]Back on land, we carry fuel barrels to our angry customers. Then we upgrade!
[18.0–29.0]Drop your cash on the pads to unlock new areas! I need more money for the car wash, so let's head back out.
单转场超前 +11.0s：在「Player goes BACK to the ocean / re-enters the 」(画面 29.0s)出现前就提到
里程碑超前 +27.0s：提前提到「Car wash unlocked」(画面 45.0s)
[29.0–38.0]Vacuuming black gold gets us crazy rich. Just look at this massive stack of cash on the boat!
[38.0–48.0]Time to expand! I'm dropping all this money to build the car wash. Boom, now we're getting passive income!
里程碑超前 +7.0s：提前提到「Car wash unlocked」(画面 45.0s)
[48.0–58.0]Next, let's build an oil pump to automate our fuel supply. But wait, these barrels are stacking up too fast!
里程碑超前 +4.0s：提前提到「Oil pump jack built」(画面 52.0s)
[58.0–68.0]I literally cannot keep up with all this demand! The whole station is overflowing with barrels, it's complete chaos!
[68.0–71.5]Oh no! Download and try it yourself!

c12xUC · 15.7s

单转场 v0/v13/5 · 3/5
里程碑 v0/v12/5 · 3/5
单转场目标(盲选):9.5s Multi-story building and money collection
里程碑(更严口径锚点):
  • 3.4s Drill close-up
  • 6.5s Completed small building
  • 9.5s Multi-story building and money collection

v0 · 线上 prompt

v0单转场 3/5●○●●○里程碑 2/5○○●●○
样例 rep1:
[0.0–3.0]Watch me build this from scratch!
[3.0–6.5]Finish the wiring, and boom, new shop!
里程碑超前 +3.5s：提前提到「Completed small building」(画面 6.5s)
[6.5–10.0]Hit upgrade and watch the money rain!
里程碑超前 +3.0s：提前提到「Multi-story building and money collection」(画面 9.5s)
[10.0–13.5]Collect that cash all the way up!
[13.5–15.7]Play Tiles Survive now!

v1 · no-lookahead

v1单转场 3/5●○○●●里程碑 3/5●○○●●
样例 rep1:
[0.0–3.5]Building from scratch is seriously so satisfying!
[3.5–6.5]Even the wiring feels super realistic.
[6.5–9.5]Look at this cozy little base.
[9.5–15.7]Watch cash turn this tiny house into a massive tower! Play Tiles Survive!

总裁定(分两条写)

① 单转场口径(和原研究同一问题):v1 没帮上忙，也没泛化。 总对齐 v0 22/30、v1 22/30，打平。6 个视频里 v1 更好 2、持平 2、更差 2。关键是原研究的 v1=3/3(n=2–3)在 5 reps 下站不住:同一个 sI4W7c，新样本里 v0 4/5、v1 2/5，反而 v0 更稳。也就是说，最初那个干净结论很大程度上是小样本+temp=1 的运气。

② 全里程碑口径(更严，看全片任何超前):v1 有温和但不彻底的改善。 总对齐 v0 8/30 → v1 12/30，v1 更好 4、持平 1、更差 1 个视频。方向是正的，但远没到「每个视频都稳定治好」——而且这套口径比原研究严，绝对数都很低，v0/v1 都漏一堆。

给 haonan 的一句话: no-lookahead 这个改动方向没错(更严口径下确实少一点超前)，但它不是一个稳定、可泛化的修复，在可比口径下甚至看不出收益，单测 case 的「3/3」别再当成定论。要上线前建议:扩大样本、或把约束写得更强/配合后处理校验，而不是凭这一版 prompt 直接拍。

方法:gemini-3.1-pro-preview · temp 1 · 6 视频 × v0/v1 × 5 reps = 60 次生成;真值标注 + 两套口径判定各一轮 · 未改动 prompts.ts · 实验脚本与原始数据在 experiments/script-alignment-study/gen31/