The attention cost — the metric nobody has
"Six pages an hour" is a limit, not a measurement. It tells you when to shut up; it never tells you what the silence bought. Beeper holds the one asset no other agent tool has — a human who decides, on purpose, on a device that does nothing else — and it is the only asset nobody prices. This is the pricing.
page ──▶ created_at ──▶ delivered_at ──▶ read_at ──▶ decided_at
│ │ │ │
└── the wait ──┘ └─ the look ─┘
(read − delivered)
└─ latency_ms ─┘
(the decision)
The bill is read_ms + decide_ms: how long a human waited to look, plus how
long they spent deciding. Nothing else counts. An agent's CPU minutes are
cheap; a human's are not, and this is the only place the distinction is made.
What is measured
| Number | Meaning | Why it matters |
|---|---|---|
human_minutes |
The bill, in minutes | The only unit a manager reads |
denial_rate |
Denied over decided | A gate that never denies is decorature |
interruptions_per_hour |
Pages over the span of the samples | The real cost of noise |
decision_median_ms |
Median, never mean | One 89 s hesitation must not hide a 1 s habit |
read − decided |
Read and never answered | The pages you saw and dropped |
| per station | The same, broken down | Which agent is worth its attention |
From the journal, never from a counter
Every figure is derived from the pages and decisions that already exist — one SQL join, no parallel state. Three consequences, all tested:
- nothing is counted twice: the sum of the stations is the total, exactly;
- nothing is invented: a page nobody read contributes no reading time, a page undecided contributes no decision time;
- the window ages with the clock: advance the clock and the bill follows, because the window is a query, not a remembered number.
A dashboard that invents a figure is worse than none: it gets read as a measurement.
The two traps
A mean lies. Three decisions of 1 s, 2 s and 89 s have a mean of 30.5 s and
a median of 2 s. The habit is 2 s; the 89 s is one hesitation. The median is
reported, and the total stays visible separately (decide_ms), because both are
true and they answer different questions.
A quiet week dilutes a busy hour. Six pages in one minute, then seven days of silence, average to well under one interruption per hour. So the rate is measured over the span of the samples, not over the calendar: the six-in-a- minute stay the fact that matters, and the silence does not launder them.
On the wire
A dashboard is not something you scrape, so the bill is an endpoint:
curl -s -H "authorization: Bearer $TOKEN" "https://…/v1/metrics/attention?hours=24"
# → {"total":{…},"stations":[…],"series":[{"at":…,"pages":3,"human_ms":41000},…],
# "window_hours":24}
hours is clamped to 1…168 (a week) and falls back to 24 on junk — the query
belongs to the caller, the budget belongs to the Central. Authenticated like
every other metric, and it carries no payload: the bill is made of timestamps
and verdicts, never of commands (BIP-N-03).
On the dashboard
/dashboard → Attention cost. The total in minutes, the five numbers above,
a table of stations worst-bill-first, and a spark of one bar per hour. Empty
hours are drawn as empty bars: a gap is a fact, and a dashboard that hides it
reads as a busy product.
What it is not
Not a leaderboard. The bill is per station because the question is "which agent is worth its attention", not "which colleague is the noisiest" — and it is self-hosted, so the numbers never leave your machine. Not a quota either: the budget (BIP-F-09) decides what may ring; this decides what it cost you, so the next budget is set from a measurement instead of from a guess.