The mistake that led to your death happened 3 minutes ago

A player dies at 11:55. The obvious thing to review is what happened around 11:55.

Were they out of position? Did they enter the fight too late? Was there a safer route? Those questions can produce useful advice. They can also produce a very polished explanation of the wrong mistake.

While planning visual analysis for LeagueLoop, my post-match coaching project, I described a seemingly efficient approach: use match metadata to identify a few promising moments, then inspect replay or VOD footage around those windows.

The correction was immediate. In League of Legends, a death may be the end of a decision chain that started minutes earlier. The player stayed for one extra wave, recalled late, returned with broken tempo, accepted the wrong lane assignment, rotated after the play was already compromised, and only then died.

The death is observable. The mistake is upstream.

The nearest explanation is often the easiest one

LeagueLoop's current prototype is deliberately narrow. It uses Riot match and timeline data to find early-game decisions, builds deterministic evidence, and rejects coaching claims the available data cannot support. It does not inspect replays or VODs.

That boundary is honest, but the first proposal for expanding it carried an important assumption: metadata would remain the index. A kill, recall, objective, purchase, or change in gold could nominate a short section of video for closer inspection.

That works when the important question is local. It can help explain whether a player had enough health to continue a fight or whether a nearby objective was available after a takedown.

It is much weaker when the outcome was shaped by a sequence like this. This is an illustrative product-design counterexample, not a measured LeagueLoop match:

1. 018:20 — stay for one more wave
2. 029:05 — recall late
3. 0310:15 — return with broken tempo
4. 0410:50 — accept the wrong lane assignment
5. 0511:30 — rotate after the setup is compromised
6. 0611:55 — die in the resulting fight

If the system inspects only 11:40–12:10, it may confidently recommend better positioning. That advice can be reasonable and still miss the more valuable lesson: the player should never have arrived at that fight under those conditions.

This is not only a game-coaching problem. Customer churn may materialize at cancellation even though the decisive friction began during onboarding. A production incident may surface at the request that timed out even though the system became fragile several deployments earlier. A sales deal may be lost in procurement after expectations were set incorrectly in the first call.

Visible failures create convenient timestamps. Root causes do not respect them.

The unit of analysis had to change

The correction changed more than the size of a video window. It changed what the system would eventually need to represent.

A "moment" contains nearby state and an outcome. A decision chain contains an initial state, the choice made, reasonable alternatives, the constraints created by that choice, later decisions made inside those constraints, and the point where the consequence finally became visible.

That distinction also changes evaluation. An expert should not be shown only the moments the existing system selected. Doing so would let the candidate generator define the supposed gold standard and hide its own blind spots.

The expert needs enough of the match to identify the root decision independently. The system can then be evaluated on whether it found substantially the same chain—not merely whether it said something intelligent about the death.

Metadata still matters. It can synchronize footage, verify recall and purchase timings, calculate resource differences, and attach hard evidence to consequences. What it cannot do is decide, by itself, which parts of the match were eligible to matter. Wave state, camera information, spacing, hesitation, and lane assignments may exist only in the visual record.

Efficiency can preserve the wrong architecture

The original selected-window proposal was attractive because it bounded cost and complexity. Full-match visual analysis is expensive. Short, metadata-selected clips are easier to store, inspect, and send through a model.

But a cheap pipeline is not efficient if it systematically removes the evidence needed for the answer.

The better direction is not necessarily one enormous model call over an entire high-resolution VOD. It could use a coarse pass to reconstruct major state transitions, form multi-minute strategic episodes, trace backward from outcomes, and reinspect the most important sections at higher resolution. Code can still calculate what code knows. Models can still be constrained to evidence. Processing can still be staged and budgeted.

The non-negotiable part is that every relevant period must be eligible for inspection before the system knows where the lesson began.

This remains a product-design correction, not a shipped multimodal system. As of this Note's publication date, LeagueLoop is still metadata-only, and its existing replay proposal still describes metadata-selected windows. The value of the correction is that it arrived before that convenient assumption hardened into implementation.

The death at 11:55 is still worth reviewing. It is just not permission to start the story there.