Clipped

200

unique matured Shorts scored in the sealed experiment

200/200

frozen scores recomputed to an exact match — zero arithmetic, component or outcome mismatches

0.934

correlation between raw content velocity and the outcome — the founding thesis, confirmed

19 of 20

of the true Top 20 selected by raw views; the composite selected 15

7.37%

of the time the composite was right when it disagreed with raw views

17

release blockers found in one review cycle — 11 by the founder, 6 by a second, independent AI

2,000

creator-block bootstrap resamples, because Shorts cluster by creator

87 → 0

tracking misses, before and after the P0 repair

Clipped began from one belief: the shape of a short-form video's early climb predicts whether it breaks out. If that is true, it can be measured at T+6 hours and acted on before anyone else knows. The product was a scoring engine — CVIE — that would rank Shorts and Reels by breakout likelihood.

The belief turned out to be correct. The product built on top of it did not. This case study is the record of finding that out on purpose, at n=200, rather than after building a business on it.

The engineering came before the question. A BullMQ + Redis + Postgres pipeline with a clean job flow — Discover → Snapshot → Enrich → Publish — ran a working Apify Instagram adapter against real Reels, with self-scheduling snapshot tracking and idempotency through unique constraints and atomic upserts. Proxy velocity was computed from repeated snapshots, growth rate via EWMA, normalised against channel history.

The instrument was unaffordable. Measuring velocity at the frequency the thesis wanted priced out at ₹3.5–8 lakh a month, and could not be obtained retroactively at any price. Ruled unaffordable before a single rupee was spent on it.

The instrument was unreliable. An external audit delivered the fairest possible verdict: the model had not failed — it had not yet been validly tested. Rediscovery every 45 minutes was silently re-adding already-tracked videos, building duplicate tracking chains nobody had noticed.

The project was deleted. The entire clipped-data-pipeline folder was destroyed by a botched terminal command and was not recoverable from Trash. The Postgres database survived. Everything else was rebuilt from scratch to 57/57 tests passing, and version control to a remote was put in place before anything else resumed.

A result you can argue with is not a result.

The methodological centre of the project is a single decision: write down, before any data exists, the exact numbers at which the idea is dead.

The frozen null. Lift ≤ 1.00. Precision advantage ≤ 0. Win rate ≤ 0.50. Committed in writing before the first observation. It is the one element that makes an uncomfortable outcome impossible to narrate into a success afterwards.

Four signals, judged separately. CVIE, raw velocity, channel-normalised and peer-normalised were each evaluated on their own and never merged into a single muddy composite for the purpose of the verdict.

Seven conjunctive gates. A signal had to clear all seven. Hard to pass by chance.

One pre-authorised extension. The sample rule allowed 100 → 200 if inconclusive, and nothing beyond. "Let's collect more data" was bounded to a single approved step rather than left as an open goalpost.

Cryptographic sealing. The rules were hashed and timestamped at commit b580594, checksum-verified before every evaluation run. The seal cannot be edited after the fact without visibly breaking.

Creator-block bootstrap. 2,000 resamples, blocked by creator rather than treating Shorts as independent observations — the choice that avoids a falsely narrow, overconfident interval. Automated evaluation ran every six hours without touching production.

The hardest sentence of the project was operational, not technical: no human reviewer was available and none was affordable. What followed was not a workaround for that gap but a deliberately constructed substitute for it.

The founder's own eleven blockers. Rather than approve the first fix, an eleven-item review closed real distributed-systems gaps: atomic slot acquisition in place of a non-atomic flag; per-claim fencing tokens against zombie writes from expired workers; a reconciler that existed in the codebase but was never called; transactional ordering of derived state after confirmed persistence; and a production-faithful cadence test replacing a synthetic one. All eleven resolved at file level, 151 tests.

Six more from an independent reviewer. A second AI was brought in specifically to audit the first one's claimed-complete work, and found six release-blocking holes the passing tests had missed:

• a migration incomplete in raw SQL, concealed because the local test database had been prepared with prisma db push instead;
• a reconciler registered but never actually scheduled;
• a T+6/T+48 eligibility calculation anchored to the wrong timestamp entirely.

Proof, in the plainest available form, that "tests pass" and "the code is correct" are different claims.

How the loop was stopped. Two adversarial reviewers can spin indefinitely. It converged only once the definition of done was frozen and every new finding was forced through one arbitration question — does it corrupt data, or is it polish? Paired with additive-only writes (nothing ever deletes raw data), a canary deploy, fix-forward-only rollback, and the founder watching live data as the true final gate.

The stranding risk, caught before deploy. Shipping the code fix alone would have silently stranded all 235 already-tracked videos for 72 hours: their in-flight legacy jobs would no-op against the new guard with nothing re-arming them until the stale-chain timer expired. The fix and its re-arming bootstrap had to deploy together. Separately flagged by both reviewers and actioned the same day: rotate the leaked database credential, shred the plaintext scratchpad.

Final state at review: 172/172 tests, 11/11 Postgres integrity suite, production boundary intact — and correctly treated as review-ready rather than approved. "The branch's description of itself earns the diff a serious review, not a signature."

The first look inside the seal, correctly marked DIRECTIONAL_ONLY at 70 of the required 100 observations.

The thesis held. Raw content velocity: 0.917 correlation with T+48 views, 100% precision in the top group.

Timing and monetization are not weak predictors. They are inversely related to the outcome, because they answer a different question — "is this good for the business?" — than the composite claimed to answer — "will this break out?" Averaging a strong signal with negatively correlated components does not dilute it; it actively drags it down. A category error in the architecture, not a tuning problem.

Channel normalisation came back actively harmful. 0.187 correlation, 0% top-K precision, after weeks of dedicated engineering built specifically to stop raw velocity being fooled by channel size.

Peer normalisation was left unresolved, not refuted. 0.825 and promising, but only 30 Shorts of coverage against a pre-registered floor of 50. Held at SYSTEMS_ONLY.

One catch worth naming, correctly scoped. T+48 cumulative views partially contain the T+24 signal being tested against them. That makes "CVIE lost to raw views" a valid result on the sealed question but a harder test than the more interesting one — does velocity predict acceleration rather than correlate with a number already containing part of its own answer. Filed as grounds for a separate follow-up experiment, not used to discount the result in hand.

The instruction that mattered most: change nothing until n=100. The pull to quietly drop channel normalisation, having already seen it hurt, is precisely the temptation pre-registration exists to survive.

Arithmetic verdict: fully correct. 200/200 recomputed scores match their frozen values exactly — zero arithmetic mismatches, zero component mismatches, zero outcome conflicts. Everything the P0 fight, the adversarial review and the cryptographic sealing were built to protect held across the full sample.

Predictive verdict: the composite does not beat raw views. Formal labels issued: NO_VERDICT_DIVERSITY_GATE, NULL_DOES_NOT_BEAT_RAW_VIEWS, REDESIGN_AWAITING_APPROVAL.

The concrete face of the failure. One Short ranked #5 by CVIE at a score of 32.46 held 5,901 actual T+48 views and a true rank of 84 — the exact failure mode diagnosed three weeks earlier at n=70, now locked and unambiguous. Not a new problem. The same one, proven.

The prescribed fix, sketched before the verdict and unchanged by it: split the composite into a purely predictive Breakout Score and an honestly non-predictive Brand Fit / Risk Score, shown side by side and never averaged into one number again.

Clipped's visual identity was developed 5–10 July, drawing on Indian newsstands and independent publishing rather than the flat SaaS conventions of the category. A "Print Era" direction, logo and motion exploration, landing-page iterations, and a set of house-style rules. The identity is treated as work in its own right, independent of the product hypothesis it was built to dress.

Assets to attach: logo lockups, Print Era boards, landing-page iterations, motion tests.

• Write the null down before the data exists. A result you can argue with is not a result.
• One pre-authorised sample extension, never two.
• Judge signals independently; a composite is a claim, not a convenience.
• Block the bootstrap by creator. Observations that cluster are not independent.
• "Tests pass" and "correct" are different claims; only an independent reviewer separates them.
• Deployed ≠ working. Verify with data, never assume — the pipeline's heartbeat logs looked healthy for the entire time four silent bugs were live, because the heartbeat never touched the database.
• Additive-only writes. Nothing ever deletes raw data.
• Arbitrate scope with one question: does it corrupt data, or is it polish?
• A number disagreeing with expectation gets investigated, never quietly recalculated until it matches hope.
• Ship the fix and its bootstrap together, or strand the fleet.

The signal this project exists to find is real, and it strengthened under the hardest test it faced — 0.917 at n=70, 0.934 at n=200. The engineering is sound and independently verified. The product built on top of the signal was wrong, and the machinery built to find that out worked exactly as designed, on schedule, against its own author's interest.

Very few commercial products are ever tested this way. That, rather than the score, is what this case study is for.

Materials not yet approved for publication. Identity work developed 5–10 July, independent of the product hypothesis it dressed.