How accurate Verifi is
These figures are counted from what Verifi has recorded: test runs of its AI and rule agents, and the decisions people made on Verifi's work. Nothing on this page is an estimate. Where too little has been recorded, it says so.
0 of 7 figures have enough data to show.A figure is shown only when at least 30 decisions or test cases are recorded. Figures from customers' own work also need at least 3 organisations, with no single organisation over half of them, and count only organisations that agreed to share outcomes for aggregation. A figure changes only once at least 30 more are recorded. Only totals are used; no organisation's data appears here.Every figure below describes the past. None of them says that a particular report, document or shipment is right.
AI and rule agents on labelled test sets
- No runs recorded yet. Not enough data yet.
- How it is counted
- Each agent version is run on Verifi's labelled test cases by the evaluation harness. The figure is the share of cases whose output matched the expected answer exactly. The latest run per agent is shown.
- Source
- Runs of Verifi's evaluation harness
- What it means
- How often this version got Verifi's own test cases exactly right.
- What it does not mean
- It is not how often it is right on your documents. Test cases are chosen by Verifi and may be easier or harder than real ones.
Values read from shipment documents, kept unchanged by the exporter
Read by Verifi's own reader (text, QR code, OCR; no AI)
Not enough data yet
Suggested by AI
Not enough data yet
- How it is counted
- Every time a person accepts, edits or rejects a value Verifi read from an uploaded document, the decision is recorded. The figure is accepted ÷ (accepted + edited + rejected) in the period. Values a person typed with no suggestion are not counted. Only organisations that agreed to share outcomes for aggregation are counted; test organisations never are.
- Source
- Each accept, edit or reject decision on a value read from a shipment document, by how the value was read
- What it means
- How often people kept a read value as it was.
- What it does not mean
- A kept value was not independently checked, and a person may keep a wrong value. It shows how often people stepped in, not how often the reading was right.
Shipment drafts from photos: drafted values kept
Drafted by AI
Not enough data yet
Read by Verifi's own reader (no AI)
Not enough data yet
- How it is counted
- When a person confirms a shipment drafted from a photo, each drafted value is recorded as kept, edited or cleared. The figure is kept ÷ (kept + edited + cleared). Only organisations that agreed to share outcomes for aggregation are counted; test organisations never are.
- Source
- Shipments confirmed from a photo draft
- What it means
- How often a drafted value survived the person's check.
- What it does not mean
- It says nothing about values the draft missed, and a kept value is the person's decision, not a second check.
Public refusal-feed rows spot-checked by staff
Fields checked by a second staff member
Not enough data yet
- How it is counted
- A second staff member checks a sample of rows taken from official public feed documents, field by field. The figure is correct ÷ checked fields.
- Source
- Staff spot checks of rows read from official public feeds
- What it means
- How often the extracted fields matched the official document in the checked sample.
- What it does not mean
- Only a sample is checked; rows outside it are not measured.
Report claims a reviewer changed or removed
Claims decided by a named reviewer
Not enough data yet
- How it is counted
- Every claim Verifi drafts for a report from its rules is decided by a named reviewer: kept as drafted, edited, or removed. The figure is (edited + removed) ÷ (kept + edited + removed). Claims decided as a reviewer's judgment are counted separately and not in the figure. Only organisations that agreed to share outcomes for aggregation are counted; test organisations never are.
- Source
- Named reviewers' decisions on report claims
- What it means
- How often a reviewer had to change or remove what the rules drafted.
- What it does not mean
- A kept claim is still the reviewer's signed decision. A low figure does not mean reports need no review.
Reported shipment outcomes that were refusals
Outcomes exporters reported
Not enough data yet
- How it is counted
- Exporters who agreed to share outcomes report whether a shipment cleared, was held or was refused. The figure is refused ÷ all reported outcomes in the period.
- Source
- Shipment outcomes reported by exporters who agreed to share them
- What it means
- The share of reported outcomes that were refusals.
- What it does not mean
- It is not a refusal rate for all shipments: only some exporters report, and those with problems may report more often. It does not show that a report caused an outcome.
Test sets in Verifi's code
- HS spices golden set (29 cases, not labelled; not scored). Product inputs only: no expected HS codes are committed yet, so this set cannot be scored. It is not a measurement.
- AskVerifi question set (100 cases, labelled; scored only through recorded runs). Questions with the sources an answer should cite. Its scores appear above only once a run is recorded by the evaluation harness.
Computed 25 Oct 2026, 7:36 pm IST · period 2025-10-25 to 2026-10-25 · method version 2026-10-12.4 · snapshot 77db203f9a60. The same counts always produce the same page, so any published version can be rebuilt and checked.
How Verifi uses AI, feature by feature: model cards (sign-in needed).