Five numbers someone else can compute
Correction rate, acceptance rate, time-to-trust, mid-run abandonment and escalation rate, each pinned to a numerator, a denominator and a window — and one of the five has no standard definition at all, which the definition has to say out loud.
Picture handing your review-gate metrics to somebody who was not in the room. A data engineer three time zones away, six months from now, with your spec open and no access to you. They read correction rate and they have to write a query. Every place your definition was vague is a place they will make a decision on your behalf, silently, and the number that comes back will look exactly as authoritative as a number you defined properly.
Take the HouseWarm gate. A reviewer opens an extracted document, sees per-field confidence, the source image crop and the raw OCR text, edits two of eleven fields, and approves. Is that one correction or two? Does the document count as corrected, or as 18% corrected? Does it count in the same bucket as an approval, since it was approved? Does it count at all if the reviewer edited a field and then reverted it? Four reasonable engineers produce four different numbers from that single interaction, and all four will be presented to a portfolio company as the correction rate.
This lesson pins down all five metrics tightly enough that the ambiguity is gone. It also does something the metric set is usually not honest about: one of the five is not a real term.
A metric is a numerator, a denominator, and a window
Not a name. The name is the part everyone agrees on and the part that carries no information. “Correction rate” is a label; documents in which at least one field value differed between the agent’s output and the reviewer’s submitted output, divided by documents that reached a terminal reviewer decision, over a rolling 28 days, bucketed by the decision timestamp is a metric. The second one can be argued with. The first one cannot, which is why it survives meetings.
Four things have to be nailed down for each of the five:
- The unit. Per field, per document, per run, per session, per reviewer. Field-level and document-level correction rate are different numbers with the same name, and the difference is usually an order of magnitude.
- The denominator. Everything that reached the gate, or everything that reached a decision? Those differ by exactly the abandoned reviews — which is its own metric, so putting abandonment in the denominator of the other four couples them together.
- The window, and which timestamp buckets it. An item that arrives on the 30th and is reviewed on the 2nd belongs to one month by arrival and another by decision. Pick one, and say which.
- Who the population is. All reviewers pooled, or per reviewer? Pooled, a single high-volume reviewer’s habits become the product’s number.
This is not pedantry imported from analytics. It is the same failure the regulation literature has been documenting for years in the exact adjacent case: Ben Green’s survey of 41 human-oversight policy documents found seven of them requiring “meaningful” human input, and observes that none of the policies proposes a definition of meaningful oversight, so there is no standard for determining whether any particular form of human oversight qualifies. A word everyone endorses and nobody defines does no work. Your metric names are one bad spec away from being the same word.
The five, defined
1. Correction rate at the review gate
Numerator: reviewed items in which at least one field value submitted by the reviewer differs from the value the agent proposed. Denominator: items that reached a terminal reviewer decision. Window: rolling 28 days, bucketed by decision timestamp. Population: reported per reviewer and pooled, never pooled alone.
Two decisions inside that. First, edit-and-revert does not count — the submitted value is what is compared, so a field touched and returned to the agent’s value is not a correction. That is the right call for this metric and the wrong call for the interpretation problem, because a revert is one of the most informative things a reviewer ever does; it gets captured as its own event rather than being folded in here. Second, you need a field-level variant as well, because “three-quarters of documents get touched” and “four percent of fields get changed” can both be true and tell opposite stories to a room.
2. Acceptance rate
Numerator: items approved with zero field values changed. Denominator: items that reached a terminal reviewer decision. Same window, same population rule.
The trap here is definitional overlap, and it is extremely common. If acceptance is defined as “approved” rather than “approved unchanged,” then every corrected-then-approved document counts in both the correction numerator and the acceptance numerator, the two rates sum past 100%, and nobody notices for a quarter. Define the three terminal outcomes as mutually exclusive — accepted unchanged, accepted with corrections, rejected — and derive every rate from that partition.
3. Time-to-trust — this course’s own construct
Say this out loud whenever you use the term: “time-to-trust” has no standardised definition. The Frontiers narrative review that catalogues the field’s trust instruments lists thirty measures of trust in automation — sixteen self-report, nine behavioral, four physiological — and none of them is a “time-to-trust.” The closest thing in that literature is a methodological proposal, attributed there to Yang et al. (2017), to quantify trust as the area under the trust curve, which requires sampling trust frequently enough to capture it dozens of times within a single experimental block. That is a research protocol, not a product metric, and it is not what anyone means when they say time-to-trust in a planning meeting.
So define it as a construct and label it as one. The definition this course uses:
Time-to-trust: the number of completed review sessions after which an individual reviewer’s per-session correction rate first drops below a stated threshold and stays below it for three consecutive sessions.
Sessions, not days — calendar time measures how often someone was rostered, not how much of the agent they have seen. Per reviewer, never pooled, because it is a property of a person’s relationship with a system, and averaging it across a team destroys the only thing it describes. And it inherits every ambiguity of correction rate, which is the point at which you should already be suspicious of it: it is a derived metric resting on a contested one, and it names an internal state — trust — that it never observes.
What you are actually allowed to say
You may say “we define time-to-trust as N sessions to a stable sub-threshold correction rate, and that is our construct, not an industry standard.” You may not say “industry time-to-trust benchmarks are around X.” There are no such benchmarks; anyone quoting one is quoting a house definition they did not disclose. If you genuinely need a validated measure of trust rather than a proxy for it, the instruments exist and have names — Jian, Bisantz & Drury’s twelve-item checklist for trust between people and automation, and Madsen & Gregor’s twenty-five-item human-computer trust questionnaire are both catalogued in the review above. They are questionnaires, though. They cost you a survey, and they measure a self-report, which is a different object from a behavior.
4. Abandonment mid-run
Numerator: review sessions in which the reviewer opened an item, produced at least one interaction with it, and reached no terminal decision before the session expired. Denominator: review sessions in which an item was opened. Window as above, bucketed by session start.
“Mid-run” is doing dangerous work in that name. In a streaming, multi-step agent, mid-run means the agent was still working and the human left. In HouseWarm’s gate — app-level, request/response, no resume — there is no run to be in the middle of; the extraction has already finished before a human sees anything. What is being abandoned is the review, not the run. Write that distinction into the definition, because the same metric name will be used on a portfolio product whose agent does stream, and the two numbers are not comparable.
Note also that this metric is only computable client-side. A review opened and walked away from emits nothing to a backend that only hears about approvals. If the UI does not emit the open, the interaction and a heartbeat, abandonment is not a number you have — it is a number that reads as zero.
5. Escalation rate
Numerator: items the reviewer routes to a named exception path — a second reviewer, a supervisor queue, a can’t-decide state. Denominator: items that reached a terminal reviewer decision, with escalation counted as a fourth terminal outcome alongside the three above.
The whole metric is conditional on an affordance existing. A gate with approve and reject and nothing else has an escalation rate of zero forever, and that zero is not evidence of a confident reviewer — it is evidence that the interface offered no way to say I am not the right person to decide this. That is a design finding disguised as a metric, and it is the one place where the EU AI Act is directly useful as a design brief rather than a compliance chore: Article 14(4) requires that a person assigned oversight of a high-risk system be enabled to decide not to use the system, or to disregard, override or reverse its output, and to interrupt it through a stop button or similar procedure. Those are affordances, and each one is an event, and each event is a metric you could not otherwise have. (That page is the AI Act implementation project’s reading copy of Regulation (EU) 2024/1689, not the Official Journal — check the OJ text before relying on it for anything with legal consequence.)
Check your recall
Answer from memory — no scrolling back.
Retrieval check
Why does this lesson insist that acceptance rate and correction rate be derived from a partition of mutually exclusive outcomes, rather than defined independently?
Check your answer
Because defined independently, they overlap. The natural reading of “acceptance rate” is the share of items the reviewer approved — and a corrected document is still approved, so it lands in both numerators. The two rates then sum to more than one, and because each is individually plausible, nobody checks the sum.
Deriving both from a single partition — accepted unchanged, accepted with corrections, rejected, escalated — makes the arithmetic self-checking: the four shares have to total one, so a definitional drift in any one of them shows up immediately as a total that doesn’t close. It also forces you to decide, once and in writing, what happens to the fifth case — the review that ended with no decision at all — instead of leaving it to whoever writes the query.
Hands on
Fill the metrics table against a real gate
Done when: Table 2 of INSTRUMENTATION.md has five rows, each with a numerator, denominator, window and population specific enough that a stranger writing the query has no decision left to make — and the time-to-trust row is labelled as this course’s own construct in the row itself, not in a footnote.
- Open
learning/agent-evaluation/INSTRUMENTATION.mdand work in section 2. Leave Second reading and What would distinguish the readings blank for now — those columns are the next lesson’s work, and filling them in early will make you write the comfortable answer. - Write all five rows against the HouseWarm gate specifically, not against an abstract review surface. Its real shape — per-field confidence, source image crop, raw OCR text, approve-or-correct, request/response with no resume — should force at least two of your definitions to say something they would not otherwise say.
- For correction rate, write both variants: per document and per field. Then write the sentence you would say in a meeting to stop someone quoting one as if it were the other.
- For time-to-trust, put the words “this is our construct; the term has no standard definition” inside the row. Then name the threshold and the number of consecutive sessions you chose, and write one line on why those numbers and not others. If the honest answer is “arbitrary,” write “arbitrary.”
- For escalation rate, check whether the gate you are describing actually has an escalation affordance. If it does not, write the row anyway and add a line in section 4 — open questions — naming the missing affordance and who has to agree to add it.
- Bring the table into the chat. I will read it as the data engineer who was not in the room, and ask you every question your definitions left open.
What this does not cover
Every one of these five is now computable, and not one of them is yet interpretable. A precisely defined correction rate that falls from 40% to 12% is a precisely defined number with two completely different stories behind it, and this lesson has given you no way to tell them apart — it has only guaranteed that everyone is computing the same ambiguous thing.
The interpretation lesson takes that seriously: it works through the second reading of each of the five, shows that the flattering reading and the alarming one are frequently indistinguishable in the data as specified here, and asks what extra signal would separate them. That question — what would I have to observe to tell these two apart — is what the instrumentation module then turns into events.
Read this next — primary source
Measurement of Trust in Automation: A Narrative Review and Reference GuideSpencer C. Kohn, Ewart J. de Visser, Eva Wiese, Yi-Ching Lee & Tyler H. Shaw, Frontiers in Psychology, 2021 — open access (CC BY)
This lesson takes one thing from it — that the field already has thirty catalogued measures of trust in automation and none of them is called “time-to-trust.” The review itself is a reference guide: it sorts those thirty into self-report, behavioral and physiological families, says what each one actually measures and at what cadence, and argues that infrequent single-point surveys miss the evolution of trust entirely. Read in full, it is the fastest way to stop inventing trust metrics from scratch and start borrowing instruments that have been validated, or at least to know precisely which validated thing your own construct is standing in for.
Stuck, curious, or think this lesson is wrong? Ask your teaching agent. The lessons are the scaffold; the conversation is where the learning gets unstuck.