Skip to content
allanninal.dev/build
Collaborate

Part 3 of 7 · Lost property matcher series ~6 min read

How a claim gets matched to an item

A claim is a description of something the claimant is not looking at, recalled several days later, and it will be wrong in at least one detail that feels certain to them.

Key takeaways

  • Zone and date are reliable. Colour, brand and size are not.
  • Match structurally first; text is a tie-breaker, never the filter.
  • The model normalises the claim into fields. It does not decide.
  • Return a shortlist to a person, never a single confident answer.
  • No match today is a subscription, not a rejection.

What a claimant actually knows

Reliability of different details in a lost property claimA bar chart with five bars showing how often each detail is recalled correctly. Where: eighty-eight percent. Which day: seventy-nine percent. Category: ninety-two percent. Colour: fifty-four percent. Brand: thirty-one percent. A note says the three tall bars are the entire matching strategy and the two short ones are what most systems search on.050100150200~88Where~79Which day~92Category~54Colour~31BrandRecalled correctly, %The three tall bars are the entire matching strategy. The two short ones are what most systems search on.
Fig 1. How reliable each part of a claim is. Building the match on the left-hand bars is the difference between finding things and telling people you have nothing.

Colour is the striking one. Roughly half of people describe their own coat as a colour it is not, which is unsurprising the moment you consider how rarely anybody looks at their own coat, and completely fatal to a system that filters on it.

Structure first, text last

The match runs on the fields that survive memory: a zone or its neighbours, a date window of a day either side, and a category. That produces a handful of candidates from a cupboard of two hundred items.

Free text then ranks that handful rather than filtering it. A claim mentioning a particular badge or a torn pocket promotes the item that has one, but the absence of a match on wording never removes a candidate, because the claimant’s wording was never reliable in the first place.

Where the model earns its place

How a lost property claim is normalised and matchedA vertical chain of four steps entered by a box labelled A claim, however it arrives: form, email or phone note. Step one, Normalise free text into the same fields, with a side box noting one model call for extraction only. Step two, Structural query on zone, date window and category. Step three, Rank on detail, using distinctive words rather than colour. Step four, Shortlist to a person, three to five items. A note says the model turns prose into fields, every decision after that is a query, and the last one is a human.AWS ACCOUNTA claim, howeverit arrives: form,email, phone noteNormalisefree text intothe same fieldsOne model callextraction onlyStructural queryzone, date window,categoryRank on detaildistinctive words,not colourShortlist to a personthree to five itemsThe model turns prose into fields. Every decision after that is a query, and the last one is a human.
Fig 2. How a claim in any format becomes a shortlist. The model does the one job models are reliably good at and nothing else.
  • App integration
  • Machine learning
  • Analytics
  • Front-end & mobile
  • People

Why the model does not decide

A model asked whether a claim matches an item will produce a confident answer and a score, and the score will be well calibrated for umbrellas and badly calibrated for the one case that matters, which is two similar black holdalls found on the same night.

Normalising a description into fields is a different job: it is extraction, it is checkable, and when it is wrong the error is visible rather than buried inside a judgement. That distinction is worth holding onto in every system in this series, and it is unusually clear here.

The shortlist

What goes to the person handling the claim

  • Three to five candidates, ordered, with the photograph visible to staff only.
  • Why each one is a candidate — same zone, one day out, same category — rather than a percentage.
  • What the claim said that does not match, stated plainly. A colour mismatch is information, not a disqualification.
  • An ownership question suggested from the item record: what is in the pocket, what the keyring has on it, what the lock screen shows.
  • And no automatic release, ever. The screen ends in a person deciding.

Showing why an item is a candidate matters more than ordering them well. Staff make the final call in fifteen seconds when they can see that the item was found in the right place on the right night, and they make it badly when all they have is a ranked list with no reasoning.

When there is no match

Most claims match nothing, because most lost things are genuinely lost somewhere else. The response to that should not be a dead end: the claim stays open, and new items logged over the following weeks are checked against it automatically.

That is where a surprising number of returns come from. An item found on Tuesday and logged on Wednesday matching a claim from the previous Saturday is a common pattern, and it only works if claims persist instead of being answered once and closed.

Closing a claim honestly

A claim that has matched nothing for the full retention period should be closed with a message that says so plainly. Leaving it silently open is worse than a clear ’we did not find it’, because the claimant is still hoping.

Next: getting the item back to its owner.

All posts