PreFlight

Two openings, one account. See which will do better before you post.

A creator often has two takes of a Reel, or two ideas for one, and only finds out which works after posting. PreFlight takes the two videos and says which opening is likely to do better against that account's usual, how often calls like this one were right in testing, and why. It reads the first five seconds, where viewers decide whether to stay: what is seen, what is written on screen, what is said and heard.

• Muse Spark watches each opening and writes it down, second by second.
• Muse Glimmer 30B, fine-tuned with QLoRA on Modal, reads the two write-ups and picks the one that will do better, with how often calls as sure as this one were right in testing.
• Muse Spark then says why, and what could lift the other one.

On 626 test pairs from 90 accounts no model trained on, drawn in two rounds, the fine-tuned Glimmer picks the Reel that did better 67% of the time, against 57% for Muse Spark asked zero-shot (and 50% for Glimmer before fine-tuning). In its surest quarter of calls it is right 82% of the time, and the Reel it picks typically got three times the views of the other.

How it was built

Everything below was written during the hackathon (Oct 26 to Nov 4, 2026); the git history shows it.

Instagram Graph API ──► labels ──► first 5 s ──► Muse Spark ──► pairs ──► QLoRA on Modal ──► held-out tests
(Business Discovery)   vs the      10 frames,    writes it      same         Muse Glimmer 30B     626 pairs,
                       account's   words, loud-  down, second   account,     answers "A" or "B"   90 unseen
                       usual       ness, cuts    by second      ≥2x apart                         accounts

1. Data, through the official API only (preflight/crawl.py, preflight/ig.py). Instagram's Business Discovery returns another professional account's Reels with their view counts and a link to the video. The crawl starts from seed lists and follows @mentions in bios: 1,787 accounts read, 62,482 Reels. No scraping.
2. Labels against the account's usual (preflight/dataset.py). Raw views mostly measure the account, which is fixed before anyone posts. So each Reel's views are divided by its account's median (Reels 14 to 120 days old, age-adjusted): 23,040 labelled Reels from 679 accounts. Instagram withholds the video of rights-claimed Reels, which leaves 8,974 videos.
3. The opening (preflight/describe.py). The first five seconds become ten frames on one contact sheet, one every half second, plus what is heard: the words (faster-whisper), how loud each second is and where the shots change (ffmpeg).
4. Muse Spark writes it down. muse-spark-1.3 fills the same form for every opening: for each second, what is seen, what text is on screen and what is heard, then the hook. About $0.012 an opening; 4,042 openings for training and the first test, 605 for the second.
5. Pairs within one account (preflight/pairs.py). One Reel that beat its account's usual and one that did not, at least 2x apart; closer pairs are mostly luck. A quarter of the accounts, picked by a hash of the name, never train: 207 test pairs (311 Reels, 37 accounts). Training gets 3,414 pairs from 289 other accounts, each shown in both orders. Once all else was fixed, a second test was drawn from the held-out Reels the first left unused: 419 pairs (605 Reels) from 75 accounts, 53 of them new.
6. Fine-tune Muse Glimmer 30B (preflight/finetune.py). QLoRA after Meta's post-training guide for Glimmer (4-bit NF4, r 16, alpha 32, the language model's attention projections) with TRL, on one Modal GPU; the answer is "A" or "B", without reasoning. Two epochs (854 steps) took 4.7 hours and about $10.
7. Score on the held-out pairs: the probability of the answer token, read with the pair both ways round and averaged, so a lean towards A or B cancels out.
8. Demo (preflight/demo.py). A page that takes two videos and runs the same path: cut, listen, Muse Spark writes it down, the fine-tuned Glimmer on Modal makes the call, Muse Spark explains. About half a minute a comparison once the judge's GPU is up.

Results

Every judge takes the same test pairs, from accounts no fine-tune saw: first 207 pairs from 37 accounts, then 419 more, drawn once everything else was fixed, from 75 accounts (53 of them new). 50% is a coin toss. Pairs from one account are not independent, so the 95% ranges come from drawing whole accounts again (preflight.pairs together).

Judge	First test (207)	Second test (419)	Both (626)
Muse Glimmer 30B, as it comes	104 (50%)
Muse Glimmer 30B, fine-tuned, after 1 epoch	131 (63%)
Muse Glimmer 30B, fine-tuned (2 epochs)	133 (64%)	285 (68%)	418 (67%, 61–72%)
Muse Spark 1.3, zero-shot	135 (65%)	222 (53%)	357 (57%, 52–62%)

On the first test the two looked level, and right on different pairs: each alone got about 40 that the other missed (38 and 40, McNemar p = 0.91). The second test, twice the size, settled it. The fine-tuned Glimmer alone was right on 117 pairs Muse Spark got wrong, Spark alone on 54 (p < 0.0001), and Glimmer held up on the 53 accounts new to the second test (197 of 293, 67%). Across both tests, its pick got a median 1.52x its account's usual and the Reel passed over 0.76x: typically (the geometric mean) 1.69x the other's views.

How sure it is

Glimmer's own probability is not worth showing: trained to near-zero loss, it is 99% sure or more of nearly every pair. How sure it is still ranks its calls, though. Split into quarters by it, over both tests, it was right 60%, 61%, 65% and 82% of the time:

Glimmer's calls	Right (95% range)	The pick's views, typically
Its surest quarter	128 of 157 (82%, 73–89%)	3.03x the other's (1.83–5.02x)
The rest	290 of 469 (62%, 56–67%)	1.39x (1.03–1.87x)

So the demo quotes each call the record of its band, never a made-up percentage. This held up out of sample: the line for a sure call was drawn on the first test (its median sureness), and on the second, the calls past it were right on 156 of 209 (75%), those below it on 129 of 210 (61%). Views are each Reel's against its account's usual, and the test pairs are at least 2x apart, so between two closer openings a right call gains less.

Two judges, tried and dropped

On the first test, asking Glimmer and Muse Spark together looked like the signal: where they agreed they were right on 95 of 129 pairs (74%), where they did not on 38 of 78 (49%), so the demo called those pairs too close. The second test did not bear it out: where they disagreed, Glimmer was right on 117 of 171 (68%). Over both, agreement adds little to Glimmer's own sureness (Glimmer is right on 70% of the pairs where they agree, 62% where they do not), so the demo now lets Glimmer make every call.

What did not help

Measured, then left out:

• One Reel at a time. Asked "did this Reel beat its usual?", Glimmer learned nothing: every held-out Reel got exactly 0.5. Comparing two Reels from one account takes the account out of the question.
• The first frame as an image. Fine-tuned with each Reel's first frame above its write-up, Glimmer got 129 of 207 right against 127 without it (a one-epoch practice run), at 1.7 times the GPU time; Muse Spark shown both first frames got 135, as without them. The write-up already says what is in the frame.
• More training pairs. 3,414 pairs (133 right) did no better than practice runs on 1,282 pairs written down by another model (127 to 139): the limit seems to be what five seconds written down can tell, not how much the model has seen.
• Sound beyond words and loudness. AudioSet labels, CLAP embeddings, beat and brightness: on their own, at chance on the held-out pairs. Reels using a sound another account used first (a "trending" sound) did no better within their account.
• What is known before posting. Caption words, length, hashtags and posting time, and CLIP embeddings of the cover, score an AUC of 0.49 to 0.53 on accounts the model never saw (preflight/baseline.py): a coin toss.

Meta AI used

• Muse Spark 1.3 through the Meta Model API (preflight/meta.py): writes down every opening (low effort), is the zero-shot judge every test is set against (medium effort), and explains the demo's call. Every call goes through a ledger that refuses anything past the credit.
• Muse Glimmer 30B (open weights, meta-models/Muse-Glimmer-30B): fine-tuned with QLoRA and served from Modal (preflight.finetune.Judge), where it makes the demo's call.
• Instagram Graph API (Business Discovery) for the data.

Try it

https://natanr123--preflight-web-web-site.us-east.modal.run

Drop in two videos you might post on one account. The first comparison after a quiet spell waits about a minute while the judge's GPU starts; then each takes about half a minute. The live demo (preflight/web.py) runs on Modal: one CPU container for the page, ffmpeg and speech, the fine-tuned Glimmer on its GPU, and Muse Spark through the Meta Model API. It spends the hackathon's credit, so it keeps its own ledger with a hard limit, and allows the public 20 comparisons a day within a monthly budget for the GPU (judges get a link without either). Videos are deleted within 15 minutes.

Run it

Needs Python 3.13, uv, ffmpeg, a Meta Model API key and a Modal account.

uv sync
echo "META_API_KEY=..." > secrets/meta.env
uv run python -m preflight.meta arm --limit 100     # the most the ledger will let it spend
uvx modal token new
uv run modal deploy -m preflight.finetune           # the fine-tuned judge, on Modal
uv run python -m preflight.demo                     # Muse Spark + glimmer:pairs-spark

Then open http://127.0.0.1:8008 and drop in two videos. The first comparison waits about a minute while Modal starts the judge's GPU.

To host it as above, put the key in a Modal secret and deploy the web app:

uv run --with python-dotenv modal secret create preflight-meta --from-dotenv secrets/meta.env
echo "PREFLIGHT_CODE=$(openssl rand -hex 6)" > secrets/web.env  # judges' code
uv run --with python-dotenv modal secret create preflight-web --from-dotenv secrets/web.env
uv run modal deploy -m preflight.web

The whole pipeline, from the crawl to the fine-tune (needs an Instagram professional account's token in secrets/ig.env):

uv run python -m preflight.crawl seed seeds/instagram.txt
uv run python -m preflight.crawl run --hours 10
uv run python -m preflight.dataset
uv run python -m preflight.videos --every 30 --hours 10
uv run python -m preflight.describe prepare
uv run python -m preflight.describe run --with meta:muse-spark-1.3:low --pairs 15
uv run python -m preflight.pairs run --judge meta:muse-spark-1.3:medium --describer meta:muse-spark-1.3:low
uv run modal run -m preflight.finetune --run pairs-spark --describer meta:muse-spark-1.3:low --model glimmer --epochs 2
# the second test, drawn once all else is fixed: draw it, write it down, judge it
uv run python -m preflight.pairs run --set fresh --judge meta:muse-spark-1.3:medium --describer meta:muse-spark-1.3:low
uv run python -m preflight.describe run --with meta:muse-spark-1.3:low --pairs 0
uv run python -m preflight.pairs run --set fresh --judge meta:muse-spark-1.3:medium --describer meta:muse-spark-1.3:low
uv run modal run -m preflight.finetune --run pairs-spark --describer meta:muse-spark-1.3:low --model glimmer --no-retrain --no-baseline --test-set fresh

Data and rights

• Official APIs only: Business Discovery for Reels and view counts, the Meta Model API and Modal for the models.
• The Reels belong to their creators. The dataset, videos, frames, write-ups and fine-tuned weights stay private and are not published, and the raw crawl is deleted after the hackathon. The demo video uses the author's own Reels.
• The Meta Model API is used in standard mode (not contributor mode), on the hackathon's credit.
• The code is under the MIT license (LICENSE).

Limitations

• 626 test pairs from 90 accounts: the ranges, drawn by account, are still about ±5 points (61–72% for the fine-tuned Glimmer).
• The pairs are at least 2x apart. Between two openings that would do about equally well, PreFlight is guessing; it still makes a call, quoting the record of calls as sure.
• Only the first five seconds, and only what can be written down: no pacing beyond the half-second frames, no music beyond loudness and words, and not the frame's shape (a landscape and a vertical cut of one opening are written down alike).
• Professional accounts large enough to be found through mentions; small creators may behave differently. On the author's own channel (animated children's stories, 57 followers), Glimmer got 13 of 29 same-story pairs right and Muse Spark 19, and Glimmer was never in its surest quarter there: three stories say little, but it is not the kind of account it learned from.