The Gate We Rebuilt After a Retired Player

The gate was supposed to answer one question: is this player available to compete tonight? It had been answering that question without incident for long enough that we had stopped thinking about it as a question at all. Then, one Tuesday in the middle of a hockey stretch, it passed a player who had not competed in four months — because he had not been formally listed as unavailable. He had retired. Quietly, in a brief statement that the data feed we relied on had not categorized as a status change. It had categorized it as nothing.

The rating that came out the other end was a gem by every internal measure. The corpus was deep — over two full seasons of tracked performance, well above our minimum. The line read cleanly. Shrinkage pulled the estimate toward the base rate, as it should. The grade was ready to post. Kieran caught it about forty minutes before we would have published, not because the process flagged anything, but because he happened to search the player's name for an unrelated reason and found the retirement announcement sitting in a news archive from mid-autumn.

We had built a gate that checked whether a player was injured, suspended, or unconfirmed in a lineup. We had not built a gate that checked whether a player still existed as an active competitor. That distinction had never come up before. It came up once, expensively, and now it is its own step.

How a Retired Player Cleared Every Check We Had

The availability gate, as it stood at the time, pulled from two sources: a real-time injury and suspension feed, and a lineup confirmation aggregator that tracked whether a player was expected to feature in the night's event. A player who was injured showed a flag. A player who was suspended showed a flag. A player who was simply not in the lineup showed as unconfirmed and was held pending confirmation.

A player who had retired showed as none of those things. His status in the injury feed was clean — no injury on record. His status in the lineup aggregator was absent, which the system read as unconfirmed rather than ineligible. Unconfirmed candidates are held, not eliminated. The gate was designed to wait for confirmation, not to treat absence as disqualification. So the system waited, found a historical lineup entry from earlier in the season that had never been purged, and resolved the confirmation ambiguity in the player's favor.

This is the kind of failure that feels obvious in retrospect and invisible in advance. The gate had been leaking in a different direction earlier that year — passing players whose listed status was active but whose participation was genuinely uncertain — and we had spent most of our gate-review energy on that problem. We had tightened confirmation logic. We had added a secondary source requirement. None of that work touched the retirement case because retirement had never been a failure mode we had observed. It was not in the taxonomy.

What We Tried First: Flagging Inactive Roster Status

The first repair was narrow. We added a roster-status field to the gate — a flag that would trip if a player's entry in the primary roster database showed anything other than "active." Retired, released, unsigned, on a reserve list with no return date: all of those would now disqualify a candidate before the corpus was even consulted.

This worked for about three weeks. Then it passed a soccer player who had been transferred between clubs mid-season and whose roster entry in the primary database still showed his old club as active. The new club had not yet propagated to our feed. He was active — just not where the system thought he was. The gate correctly identified him as active and passed him through. The corpus it consulted was from his old club's tactical context. The rating was not wrong enough to be embarrassing, but it was measuring something that no longer existed in the form we thought it did.

Kieran's note from that week, which I have kept because it is the most useful single sentence anyone in this shop has written about data infrastructure:

"Roster status is not a fact. It's a claim made by a specific source at a specific moment, and the source is almost always behind the event."

He was right, and it reframed the whole repair. We were not trying to read a reliable field. We were trying to triangulate between several unreliable ones and decide how much disagreement was tolerable before the gate closed. That is a different problem, and it does not have a clean solution.

What the Narrow Fix Got Wrong About the Actual Problem

The cost of the first repair was that it gave us false confidence in a field that was not designed to carry the weight we were putting on it. Roster status in most data feeds is updated on a league-administrative timeline, not a real-world timeline. A player can retire on a Wednesday and remain listed as active in the feed until the following Monday's administrative update. A player can be released and show as active for longer than that if the club's paperwork is slow. We knew this abstractly. We had written about corpus entries that survived every check and still produced bad output — the problem of a clean record attached to a player who had changed in ways the record did not reflect. We had not applied that same skepticism to the gate inputs themselves.

There was also a subtler cost. By adding the roster-status check as a hard gate — pass or fail, no gradation — we created a new failure mode on the other side. Players returning from long absences sometimes showed a brief window of "inactive" status during reinstatement processing. The gate was now eliminating candidates who were genuinely available but administratively mid-transition. We caught two of those before they became published misses, but only because the volume of eliminated candidates that week looked anomalously high and someone went looking for the reason.

We had fixed the retirement case by introducing an over-elimination case, and we had not noticed until the numbers looked strange. This is, I think, the standard shape of gate repair: you close the gap you can see and open a smaller one somewhere adjacent. We have been through that loop before, and I still do not have a reliable way to see the adjacent gap before it shows up in the grades.

The Two-Source Disagreement Rule We Kept From the Rebuild

What we kept from the rebuild was not a new field or a new flag. It was a rule about disagreement. If a candidate's status across two independent sources does not agree — one says active, one says inactive; one has a current club, one has a prior club; one has a lineup entry, one has no entry at all — the gate now closes automatically, regardless of which source is nominally more authoritative. The candidate is held for manual review, which in practice means it does not get rated that cycle.

This is conservative to the point of being occasionally frustrating. We lose candidates who are genuinely available and whose status disagreement is a pure data-lag artifact. The gate does not care. Its job is not to maximize the number of candidates who pass through — it is to ensure that the ones who do pass through are actually who the corpus thinks they are. We have accepted that we will miss some valid candidates in exchange for not rating ghosts.

The retirement case, specifically, now has its own sub-check: any candidate whose most recent competitive appearance is more than ninety days prior to the rating date is flagged for manual confirmation, regardless of what the status feeds say. Ninety days is not a principled number — it came from looking at the distribution of long absences in our corpus and finding that legitimate absences of that length almost always had a documented return date attached. Absences without a documented return date, at that length, were overwhelmingly retirements, transfers, or administrative errors. The threshold has held for two seasons without a known miss. That is not the same as saying it is right.

We also tightened the relationship between the availability gate and the corpus recency requirement. A player with a deep corpus but no activity in the relevant window used to pass the availability gate on the strength of the corpus alone. Now the two checks are linked: corpus depth counts for nothing if the availability signal is ambiguous. This was a direct response to what happened — the original miss was possible precisely because a rich historical record papered over the absence of any current signal. The method should not let the past stand in for the present, and for a while, ours did.

The gate we have now is better than the one we had, and I am reasonably confident it would catch the original miss. What I am less sure about is whether the next retirement — or the next transfer, or the next administrative anomaly we have not seen yet — will look enough like the ones we have already seen for the gate to recognize it. The gate is a pattern-matcher, and it only knows the patterns it has been shown.

Note: PlayerGem is a fictional analytics shop and these accounts are invented. Nothing here is a pick, a recommendation, or betting or investment advice, and the players, teams and competitions described do not exist.