Porchlight your friendly neighborhood agent
Sent to this zone, Parcel lockers, bldg 3
Over the past 13 hours, from yesterday evening through this morning, there have been several reports of unusual lingering near the parcel lockers at building 3. This is a spot that normally sees very little activity of concern. If you're expecting deliveries, you may want to retrieve packages promptly, and report anything that seems out of place.
3 reports · 3 reporters · 13h · z=6.5
That is the whole of what Porchlight sent. One message, to the people who live on that street, drafted by the agent and released by a person.
38 reports · 27 logged silently · 9 correlated and declined · 1 suppressed as duplicates · 1 alert
The three ringed marks belong to the same situation as the lit one, and produced no second message. The other thirty-four were read, weighed and left where they were. Nobody was notified, nobody was woken, and nobody was described. Hover any mark to see what it was.
These arrived as separate messages, hours and days apart, from people who had not spoken to each other. Read down the left column and they share almost no words. Read down the right and the agent has worked out which of them are the same situation.
One of the locker doors will not shut properly. It has been like that a while now.
A parcel locker door at building 3 will not shut properly.
The recycling has not been collected on our side of the road for two weeks.
Recycling has not been collected on the south side of Elm St for two weeks.
There has been a guy hanging around the mailboxes the last couple of evenings. Does not live in our building as far as I know.
A person has been seen near the parcel lockers at building 3 on multiple evenings.
Saw someone loitering by the post boxes again tonight when I got back from work. Just stood there on their phone.
A person was seen standing near the parcel lockers at building 3, using a phone.
Somebody was messing about near where the packages get dropped when I left this morning. Could not tell what they were doing.
A person was seen near the parcel lockers at building 3 on the morning of April 7 at approximately 07:50.
Person waiting around by the delivery lockers early on again. That is the third time this week someone has brought it up.
A person was seen waiting near the delivery lockers at building 3 at 6:20 AM.
4 ways of saying it, not one content word in common, and one place: Parcel lockers, bldg 3
Only the right column is indexed or correlated. The left column is held briefly and deleted on a retention timer. Every report here is invented fixture data, and the raw text is off by default.
Four reports, one place, a tight window, and four people behind them. The agent found the pattern, looked at who filed it, and did not send it. Nobody heard about this. Nobody got talked about.
Correlated, not sent, Birch Ln
Four distinct reporters is meaningful — that's real corroboration — and the reports do share a loose theme (vehicle/person loitering on Birch Ln in the evening). But the decisive evidence against alerting is the combination of temporal spread and anomaly score. The four suspicious reports span 21 days (March 14 to April 4). That's roughly one report per week, which describes what a residential street looks like over a month, not a developing situation. The newest report is 12 days after the previous one, so there is no tightening cluster. Each incident resolved on its own (the van drove off, the vehicle left, the person moved on). None escalated to a crime report. The anomaly score of 1.21 confirms this: Birch Ln is not measurably busier than its own baseline. The correlation stage itself flagged that there's no volume spike. When the zone's own history says this is normal background activity and the reports are spaced weeks apart, calling it a pattern risks turning ordinary neighbourhood noise into something it isn't. The broken streetlight is an interesting contextual detail, but it doesn't change the calculus — it's a maintenance issue, not evidence of a threat pattern. Four people each saw one thing, weeks apart, that looked a bit odd to them. That's a street, not a situation. Logged silently.
4 reports · 4 reporters · 504h · z=1.2
Triage
Haiku 4.5
reads one report, strips the person, writes one sentence
global.anthropic.claude-haiku-4-5-20251001-v1:0
Correlation
Sonnet 4.6
chooses its own lookups, decides what is related
global.anthropic.claude-sonnet-4-6
Escalation
Opus 4.6
weighs the counted evidence and makes the call
global.anthropic.claude-opus-4-6-v1
A held-out set of 20 cases, written before any tuning began, passed 20 of 20. It was run once, on 31 August 2026, and has not been run since: re-running a holdout after a prompt change turns it back into tuning data. It produced zero declines, so it does not show the agent refusing a cluster that looked real.
Reading every report costs about 5 cents each. Around $1.90 to work through a month like this one, across all three models. Measured rather than estimated: $9.39 of Bedrock spend across roughly 190 reports processed, read from the AWS credits page on 2 September 2026.
What this page does not show. Every report here is invented fixture data. There are no real users, no real incidents and no deployment behind it. The split between reports logged silently and reports correlated then declined moves by one between runs, so it is reported, not promised; only the totals are stable. And the agent is not always the thing that keeps a report quiet. Some are never linked to anything in the first place, which is retrieval separating them rather than judgment refusing them.