Accuracy · Measured, not claimed
How accurate is Prism?
These are the numbers from our own benchmark, published as measured on 21 September 2026. The documents are synthetic, written to be hard, and hand-checked; results vary from run to run; and a date Prism misses is a miss, not a wrong date, because dates are computed by code from stated rules and a period the document never states is left empty rather than guessed.
Method
- Eight synthetic contract sets written to be difficult: long US office and retail center leases with one to three amendments each; two UK leases with deeds of variation; a purchase and sale agreement; a construction contract; a loan modification; a vendor agreement.
- Expected dates worked out by hand, including the superseded dates that must not appear and the obligations that must stay undated.
- The production model, two runs per document, reference date 21 September 2026.
- Two readings of each set: the base document alone, and the base read together with its amendments.
Deadlines
Across all eight sets. “Single document” is the base contract read alone; “with amendments” is the base read together with its amendments, deeds of variation or change orders.
| Measure | Single document | With amendments |
|---|---|---|
| Expected deadline dates found | 58% (90 of 154) | 75% (120 of 160) |
| Found with the right label | 47% | 64% |
| Deleted or superseded terms correctly left out | 95% (70 of 74) | 97% (105 of 108) |
Abstract (leases only)
Correct fields out of 35, where a field the lease does not state counts as correct only when it was left empty.
| Lease | Correct fields |
|---|---|
| US office lease | 93% |
| US retail center lease | 84% single, 91% with amendments |
| UK office lease | 74% single, 77% with amendments |
| UK retail lease | 59% single (one of two runs returned nothing), 93% with amendments |
| Three shorter leases, earlier the same day | 94% (187 of 198), with no invented quotes |
How to read these numbers
Results vary run to run
The same document read twice does not return the same list. The figures are across both runs; a single run can land above or below them.
Synthetic documents, not customer documents
They were written to carry the traps real documents have (an estimate superseded by a formula, a summary that disagrees with the article, a deed that changes six months to nine), but they are not a sample of what customers send.
A missed date is a miss, not a wrong date
Dates are computed by code from the anchor and rule the model reads. Where a document never states a period, Prism leaves the date empty rather than guessing. The failure mode is an obligation left out or left undated, not a wrong date presented with confidence.
A label miss is a filing error
The date was found and dated correctly but filed under a different kind of obligation than we expected, for example an extension option counted as a renewal option.
What we are fixing
- Obligations whose period is stated as "the last day of the Nth full calendar month" when the anchor falls mid-month.
- Recurring calendar-date obligations ("within 14 days after each 1 January") coming back undated.
- Run-to-run variance.
Use it from an AI agent
Prism is a remote MCP server. In Claude Code:
claude mcp add --transport http prism https://prism.parad1gm.com/api/prism-mcpThen ask: “Read this lease and list every critical date with the clause each one comes from.” The agent sends the document with prism_review_contract and reads the result with prism_get_contract. Setup for Claude on the web, Cursor, VS Code and A2A agents is on the agents page.
$0.50 a contract for an agent, and the first 3 are free. On the website it is $1 per contract, or $9 a month for up to 50.
Questions
Why publish numbers this low?
Because a person or an agent choosing a tool should know what it does before relying on it. 58% of expected dates in a single hard document is what we measured. Every date it does return comes with the clause it came from, so the ones it finds can be checked against the document.
Does a missed date mean a wrong date?
No. Dates are computed in code from the rule and anchor the model reads. When the model does not find the rule, or the document never states a period, the obligation is left out or left undated. It is not given a guessed date.
Why are the amended results higher than the single-document results?
An amendment often restates the date it moves as a plain calendar date, which is easier to find than a formula, and the amended sets have more expected dates. The superseded dates an amendment replaces were left out 97% of the time.
Were these real customer documents?
No. They are synthetic contracts we wrote to be difficult and checked by hand. Customer documents are not used for this benchmark.
How do I report a wrong result?
From an agent, prism_send_feedback; otherwise email support@parad1gm.com. A person reads every report, and a wrong result we reproduce earns an agent account 10 free credits.
Will these numbers change?
This page states the date of the run so you can tell how old the figures are. When the fixes above land and the benchmark is rerun, the figures here are replaced with the new ones, whichever way they move.