Shadow mode — measure the interruption before you take it
Intercepting a human is the product. It is also the cost. Nobody can say what a gate would have saved, because turning it on is the only way to find out — and the first thing it finds out is on a Friday afternoon, in front of someone trying to ship.
Shadow mode breaks that circle: intercept, display, simulate the agreement, never block. Seven days of it in production and the score is a measurement, not a promise. You are not selling an interruption any more, you are selling a number.
agent ──▶ POST /v1/pages ──▶ Central
│
┌───────────────┴────────────────┐
▼ ▼
enforce (default) shadow (BIPEUR_SHADOW=1)
compose → route → block compose → route → SHOW
wait for a signed decision resolve: approved, SIMULATED
│ │
▼ ▼
the agent waits the agent goes, at once
the pager still shows the page
the log says `page.shadowed`
Three rules, and they are the whole design
A simulated verdict is never indistinguishable from a real one. The wire
carries simulated: true next to the status; the audit carries
page.shadowed; and no row ever enters the decisions table. So the metrics
of A.3 — median decision time, unblocked time, filter rate — stay what a human
did, and a month of shadow does not quietly inflate them with six hundred
approvals nobody made. An adapter that ignores the flag and acts as if a person
had approved is no longer protected by anything, and the log says which verdict
it was handed.
A page a human actually decides wins over the simulation, whenever it
arrives. Shadow mode resolves the wait; it never mutes the pager. The page
is on the screen, it can be read, it can be decided for real, and the instant a
signed decision lands, simulated flips to false — because that verdict now
has a person behind it.
The counters are read back from the journal, never accumulated in memory. A counter that dies with the process is not a measurement either. Restart the Central and the score is what it was, because it is the log.
What the score says
curl -s -H "authorization: Bearer $TOKEN" https://…/v1/shadow
# → {"enabled":true,"simulated":412,"would_have_held":37,
# "would_have_been_denied":9,"really_decided":14}
| Number | Meaning | What you do with it |
|---|---|---|
simulated |
Pages that would have interrupted you | The volume of the gate |
would_have_held |
Pages a saturated budget would have held | The noise you were spared |
would_have_been_denied |
Pages the rules would have refused | The disasters avoided |
really_decided |
Pages a human decided anyway, during shadow | The ones that needed a person |
would_have_* is computed before the simulation, so it describes the world
where Beeper interrupts — not this one. That is the point of the exercise: you
are measuring the counterfactual, and the log keeps the two apart.
Turning it on and off
# At boot:
BIPEUR_SHADOW=1 bipeur-central
# Live, without restarting:
curl -s -X PUT -H "authorization: Bearer $TOKEN" https://…/v1/shadow -d '{"enabled":true}'
The toggle is journalled as shadow.toggled — a distinct event type, not a
page. Who flipped it, and when, is part of the record: a mode that silently
changed in the middle of a week would make the score worthless.
Off is the default, and it stays the default. Measuring what an interruption would cost must never be the reason an interruption goes missing.
What it is not
Not a trial period, not a demo, not a way to soften the product. Shadow mode is the deployment strategy of a gate: you do not hand someone a device that stops their agents until you can show what it stopped. It is also the only honest answer to the objection that kills this category of tool — "what if it blocks something important?" — because the answer is a number from your own journal, not a slide.
The pages are still composed, still routed, still shown. Nothing about the product is simulated: only the agreement is.