Antonio Poggi
← All writing

Public alerts · Provenance

Designing Alerts That Preserve the Evidence Behind the Claim

The social post should attract attention without becoming the only surviving record of the measurement, method and context.

27 July 2026 6 min read Citizen science and public evidence
Hourly benzene report for Via Machiavelli showing July 2025 peaks and the 27 micrograms per cubic metre acute reference
An alert image should show the event and the immediate context. The interactive report should preserve the deeper evidence.

A bare alert is easy to produce:

High benzene in Taranto: 42.54.

It is also almost useless as evidence.

Which station? Which hour? Which unit? Was 42.54 an hourly mean or an instantaneous value? High compared with what? Was the source provisional? What happened before and after? Can anybody inspect the observation?

The number gets attention because it is large. The missing context makes it difficult to assess and easy to misuse.

An alert should be an index into evidence, not the end of the evidence.

That principle has shaped the automated air-quality monitoring I have built for Taranto.

The alert is the last step of a data workflow

The visible post may contain one sentence and one image. Behind it sits a complete process:

official source
  → retrieve and timestamp
  → type and validate
  → preserve station and averaging grain
  → apply an explicit reference rule
  → isolate new qualifying observations
  → generate evidence image and text
  → publish
  → link to the interactive report
  → record what was sent

In the original OmniscopeBot implementation, an Omniscope workflow obtained the data, transformed it, compared measurements with configured conditions, generated a report screenshot and passed qualifying rows to a Python block for publication.

The code that called the social API was a small part. The analytical decisions before it were the alert system.

The minimum evidence envelope

Every measurement-based alert should carry or link to an evidence envelope.

1. Station and pollutant

“Taranto” is not a measurement location.

Name the station and pollutant. If the station identifier changed, preserve the identifier used by the source.

2. Observation time and timezone

State the hour or period represented by the value, not only when the post was published.

Use a timezone. Environmental series cross daylight-saving changes, and “03:00” can otherwise become ambiguous.

3. Value, unit and averaging duration

Write:

42.54 µg/m³, hourly mean

not:

42.54

The averaging duration is part of the measure. Hourly, daily and annual values cannot be compared as if they were the same threshold event.

4. Reference and its status

If the alert uses 27 µg/m³, call it the OEHHA acute Reference Exposure Level for infrequent one-hour benzene exposure and non-cancer effects.

Also say what it is not: exceeding the reference is neither proof of individual harm nor an Italian or EU ambient legal breach.

A threshold without its authority, duration and legal status invites a much stronger interpretation than the data supports.

5. Source-data status

Identify the agency and whether the observation is provisional or validated where that information is available.

If a later validated file can revise the value, the publication process needs a way to retain or amend the original alert.

The post should link to a report where a reader can inspect:

  • the surrounding hours;
  • history;
  • station comparisons;
  • source rows;
  • update time;
  • gaps in coverage.

The current Aria Taranto report and benzene history provide that path.

7. Workflow and publication time

The observation may represent 03:00, the source may publish it later, the workflow may retrieve it at 05:10 and the social API may accept the post at 05:11.

Those are different times. Preserve them.

The screenshot is a bridge

A well-designed image gives the reader enough visual context to decide whether to investigate:

  • the current observation;
  • the reference line;
  • the previous hours or days;
  • station name;
  • source update time;
  • units.

The image is not the dataset.

It may be compressed by the platform. Alt text may disappear. The post may be re-shared without its link. A screenshot also freezes the provisional state at publication time while the report may later refresh.

This is why both are useful:

The screenshot creates immediacy. The report preserves exploration.

Context should counterbalance the peak

An alert feed naturally emphasises exceptional events. Without longer-term context, it can leave the false impression that conditions never improve.

In the incomplete 2025 period available when this comparison was published, Taranto reporting showed a falling number and average of high daily benzene readings across several stations from 2023 into 2025, while individual hourly episodes still deserved attention.

Omniscope report comparing high daily benzene readings across four Taranto stations from 2023 through 2025
This chart uses daily means above 5 µg/m³, a different measure from an hourly 27 µg/m³ alert. The 2025 period was incomplete when the image was published. Long-term context and short-term alerts can both be true.

The alert channel and historical report should let a reader hold both facts.

Do not rewrite the historical rules

The first 2023 OmniscopeBot tutorial used a different, lower trigger in its example. The current project uses the 27 µg/m³ acute reference for particular hourly alerts.

It would be inaccurate to describe every historical post as if it used today’s rule.

Store a rule version with each alert:

  • threshold value;
  • comparison operator;
  • averaging period;
  • pollutant and station scope;
  • source of the reference;
  • effective date;
  • workflow version.

When the rule changes, history should explain the change rather than silently reclassifying old publications.

Handle duplicates and revisions

Automated publication creates ordinary distributed-systems problems.

Duplicate protection

A retry after a timeout may publish the same observation twice.

Use a stable key such as:

source + station + pollutant + observation_timestamp + rule_version

Record successful publication separately from attempted publication. Do not assume an HTTP timeout means the social platform rejected the post.

Source revisions

If provisional data is corrected, retain:

  • the value originally alerted;
  • the corrected value;
  • revision time;
  • source file or record version;
  • whether the alert was amended or withdrawn.

Deleting the old post without explanation weakens the public record.

Gaps

If the source or workflow is unavailable, make the gap visible. Absence of alerts should never be interpreted automatically as absence of high readings.

Monitor the monitor.

Keep a record outside the social platform

The Taranto alert channel began on X through OmniscopeBot and later moved to the Aria Taranto Facebook page.

Platforms change API access, link behaviour, image quality and account policy. Accounts can be suspended or removed.

The durable record should remain under project control:

  • alert payload;
  • source observation or snapshot;
  • generated image;
  • rule version;
  • workflow execution;
  • publication result and platform identifier;
  • canonical report URL;
  • correction history.

The social post is distribution. It should not be the archive.

Learn from official communication standards

Annex XVI of Directive 2008/50/EC requires official exceedance information to identify matters such as location, threshold type, start and duration, and concentrations.

Citizen benzene alerts are not statutory EU alerts, and this does not make them so.

The structure is still useful design inspiration. Public warnings should tell people what happened, where, when, at what concentration and against which defined reference.

That is the test I would apply: an alert must explain itself at publication and leave enough state to be corrected later.

The alert gets attention.

The evidence earns trust.