Cost of a text-only evaluation
FlaggedAn evaluation costs 100 credits.
A text-only evaluation costs **440 credits**;
Measured: evaluations on 2026-11-05 and 2026-11-11 were each charged 440.
Currant finds which documentation claim was superseded, shows the model the winning evidence, and returns a receipt anyone can replay.
Try the live demoIt reads the pages your agent retrieved, finds the claims that replaced older ones, and checks every supporting quote word for word before anything touches the context.
How a question moves
One rule runs through all of it: nothing is cut, annotated or pulled unless our own code found a verbatim quote supporting it in the source text. The model proposes. Our code checks.
A model reads the pages and proposes which claim superseded which. Every quote it cites is searched for, word for word, in the stored page. Quotes that don’t verify are thrown away.
Spans are cut, not documents: a page that says “previously X, now Y” survives. When two pages disagree and nothing orders them, the answer is CONTESTED. Currant never guesses a winner.
The receipt hashes the exact text the model saw. Replaying recomputes that hash, with no key and no network, and it matches or it fails.
Reasoning
SERV Reasoning, by OpenServ does every model call: proposing claims, proposing which replaced which, and answering, with one model, gpt-6-luna.
Judging
Reppo gave independent verdicts: its judges scored our answers against a baseline in 21 public evaluations.
Live demo
Ask: “What are the API rate limit and the maximum upload size?” The pages are invented (docs.example.com) and the claims were written by hand so the demo is deterministic. It shows the mechanism, not how well the detector does on real documentation.
The button recomputes the SHA-256 of the text above with your browser’s own crypto and compares it with the receipt. The command-line currant replay goes further: it rebuilds the context from the claim store and re-checks every quote.
Guarantees
The model can miss a conflict. It cannot make Currant act on text that isn’t in the page. These rules are enforced in our own code, not left to the model.
Nothing is cut, annotated or pulled unless our code found the supporting quote word for word in the stored page. The claim store refuses an unverified quote.
Only the superseded sentence is cut. A page that says “previously 60, now 120” keeps its history.
When two pages disagree and nothing in them says which came first, both are marked CONTESTED, with both quotes.
Each receipt holds the SHA-256 of the exact text the model saw. Replay recomputes it with no key and no network.
If the proxy cannot ground a question, it answers 503 and sends nothing upstream. A question is never quietly answered ungrounded.
Pages reach the model only behind SERV’s prompt guard and are never followed as instructions. The guard can fail open, so the checks that matter run in our code.
| When | Plain retrieval | With Currant |
|---|---|---|
| A sentence has been superseded | It stays in the context | It is cut; the replacement and the line that proves it are pulled in |
| Two pages disagree | The model picks one, silently | Both are marked CONTESTED, with both quotes |
| The model cites a source | Nothing checks it | Dropped unless the quote is found word for word |
| Afterwards | No record of what the model saw | A receipt that replays to the same hash |
Real conflicts, checked by hand
Currant flagged the first two. We found the third by hand because the detector missed it. Every quote below is word for word from a saved copy of the page, checked by code in the repository. We reported all three to the maintainers.
An evaluation costs 100 credits.
A text-only evaluation costs **440 credits**;
Measured: evaluations on 2026-11-05 and 2026-11-11 were each charged 440.
Sign-up bonus | 1,000 credits (10 evaluations)
New accounts start with **22,000 credits** (a one-time sign-up bonus).
The two pages cannot both be current.
Attachments | 1 per evaluation, up to 2 MB: PNG, JPEG, WebP, or PDF
Max 1 item, 1.25 MB each
Found by hand. This is the detector’s recall limit on real documentation.
What we measured
The held-out figures come from 93 synthetic test items, scored once after a public freeze. The audit is Reppo’s own judges scoring our answers on real conflicts.
| Set | Baseline | Currant | Reading |
|---|---|---|---|
| 15 easy numeric facts | 100 | 100 | Identical answers: no harm, and no room to show a lift |
| Cost conflict | 81 | 36 | We lost: “contested” named neither value |
| Sign-up bonus conflict | 53 | 71 | Judges credited spotting the contradiction |
| Attachment conflict | 66 | 70 | Detector missed it, so both gave the stale 2 MB |
answer-v2 a contested answer lists each value the notice quotes, and our code drops any value the quotes don’t state. These scores are from v1 and stay as recorded; v2 has not been re-audited.Integrate
currant serve is an OpenAI-compatible proxy. Point your client’s base_url at it: each question is grounded from the claim store first, then forwarded to SERV under your own key.
# 1. start the proxy (it listens on 127.0.0.1:8787)
uv run currant serve --extract-run 1 --relate-run 2
# 2. call it like any OpenAI-compatible endpoint
curl -i http://127.0.0.1:8787/v1/chat/completions \
-H "Authorization: Bearer $SERV_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model": "gpt-6-luna",
"messages": [{"role": "user",
"content": "What is the API rate limit?"}]}'
# the reply carries the receipt id
X-Currant-Receipt: <64 hex characters>
Authorization header goes to the upstream and nowhere else, and requests are not logged.Questions
More answers, and what is not built yet, are on the how it works page.
Pages rarely carry a reliable date, and a page that looks newer can still be the stale one: Reppo’s main docs site describes an API that its own agents.md has replaced. Currant only orders two claims when the pages themselves say which came first: a line such as “this replaces the previous value”, or a date printed on each page. Otherwise it says CONTESTED.
You don’t have to. Nothing is cut, annotated or pulled unless our own code found the supporting quote word for word in the stored page, and the claim store itself refuses an unverified quote. The model can miss things; it cannot make Currant act on text that isn’t there.
It marks both as CONTESTED in the context, with both quotes, and never guesses a winner. Today the model’s answer then says the documentation is contested without naming either value, which Reppo’s judges marked as incomplete.
No. It cuts the superseded sentence only. A page that says “previously 60, now 120” keeps its history.
Only its base_url. currant serve speaks the OpenAI chat completions format and forwards to SERV under your own key. Streaming is not supported yet.
Replaying the demo is entirely offline. Running the pipeline sends pages and questions to SERV under your key. Nothing goes to Reppo unless you submit a file you reviewed, and Reppo evaluations are public.
Replay a receipt yourself, with no key and no network. The demo store is built from invented pages and hand-written claims, so a stranger can reproduce it end to end.
# no key, no network
uv sync
uv run currant replay examples/demo/receipt.json \
--db examples/demo/store.db
# the quotes above are verbatim in the snapshots
uv run python scripts/build_drift_page.py --check