TestedMatches the best single model on 483 unseen questions

AI answers you can actually prove

Several AI models from different companies answer your question, check each other against your documents, and show you which statements hold up, which are disputed, and why.

Start free, no card needed See a checked answer

60 free credits a month
Models from 4 companies
Every statement quoted
Never used for training

Claude Opus 5.5Anthropic
GPT-6.1 SolOpenAI
Grok 4.7xAI
Kimi K3Moonshot AI

Standard check4 models, 4 companiesServices agreement.pdf

Question

Can the supplier share our confidential information with its affiliates without telling us?

Checked statements · select one to see the proof

Checking statements…

Illustrative example. Labels and layout match the product; the agreement is fictional.

Use any top AI model for drafts. Use Danphe when a wrong answer is expensive.

The difference

Same question, two kinds of answer

A chatbot gives you fluent text. Danphe gives you statements you can check, each with its quote and its verdict.

Same question: “Can the supplier share our confidential information with its affiliates without telling us?”

A typical chatbot answer

Yes, the supplier can share your confidential information with its affiliates without notifying you, as long as those affiliates are bound by similar confidentiality terms. However, the supplier must inform you within 30 days after any such disclosure.

No sources. One sentence is invented, and nothing shows which.

Danphe Council

Models agree
Sharing with affiliates needs no prior notice.
“without prior notice to the Disclosing Party” §7.2
Models agree
Affiliates must be bound by equally protective terms.
“obligations no less protective than those set out herein” §7.2(b)
Disputed
The supplier must tell you within 30 days after sharing.
3 of 3 checks found no such clause. Removed from the answer.

Each statement is quoted, checked by other models, and labelled.

Illustrative example; the agreement is fictional.

How it works

How a question is checked

Independent answers first, then every statement is checked by models from other companies against your sources. Nothing is averaged away.

Step 1: Reading sources
Your documents are read and the passages that matter to the question are found.
Step 2: Models answer on their own
Models from different companies answer independently, without seeing each other's answers.
Step 3: Cross-checking every statement
Each statement is checked against your sources by models from other companies, with quotes.
Step 4: Writing the answer
An editor writes the answer from what held up, and marks what didn't.

Smart routing

Why it's fast and affordable

The full council only runs when it's needed. Two models answer first, and only disagreement brings in the rest.

Every statement in both paths is still quoted and labelled.

Fast when models agree
Two models from different companies answer first. If their answers match, you get the answer in about 20 seconds for 3 credits.
The full council when it matters
If they disagree, or your workspace is marked high-risk, all four models answer and check each other: 10 credits.
$0.24 on average
Smart mode's average model cost on the 483-question locked test: $0.236, under the preset limit of half the Standard council ($0.279). Long free-text questions mostly take the full council.

How we measured cost and accuracy

Labels

Every statement gets a verdict

The answer is split into individual statements. Each one carries a label with an icon and a word, never colour alone.

ConfirmedComing soon
Checked and certified
Independent models checked this against your sources, and this workspace has a signed-off calibration, so the error rate of “Confirmed” statements is measured.
Models agree
Backed by independent checks, with a quote (not certified)
Independent models checked this and it is backed by a quote from your sources. It is not marked “Confirmed” because this workspace has no signed-off calibration yet, so the error rate of these statements has not been measured. Check the quote if the decision matters.
Not settled
Backed, but not certified
The independent checks were mixed, or the point is backed but without a quote from your sources. Check the source before relying on it.
Not supported
No model could back this up
Independent checks found no support in your sources, or found evidence against it. It was removed from or flagged in the answer.
Opinion
A judgement, not a checkable fact
Recommendations, interpretations and other judgements. They are shown for context but cannot be proven true or false from the sources.

“Confirmed” vs “Models agree”: a statement is only marked “Confirmed” in a workspace with a signed-off calibration, where the error rate of confirmed statements has been measured. Without one, the same evidence is labelled “Models agree”. No label is a guarantee that a statement is true. No workspace has a signed-off calibration yet, so “Confirmed” is not shown to anyone today.

Dissent

Disagreement is shown, not hidden

When models read your sources differently, you see both readings and the quotes behind them, instead of a confident average.

A single model gives you one voice. A split council is a signal: it tells you exactly where a person should read the source.

Not settledThe checks were split, so this point is not settled.

Do we pay a fee if we end the contract early?

Reading A · 2 models

The early-termination fee applies whenever you end the contract for convenience.

If Customer terminates for convenience, Customer shall pay the Early Termination Fee.Services Agreement, §12.3

Reading B · 2 models

The fee no longer applies after the first contract year.

Section 12.3 shall not apply after the first anniversary of the Effective Date.Services Agreement, §12.4

What the answer tells you: both clauses exist and point different ways depending on timing. Read §12.3 and §12.4 together before deciding. (Illustrative example.)

Accuracy

Measured, not promised

We test on questions the system has never seen, with the rules written down before the run. On short questions Danphe matches the best single model; the gain is that every statement is checked and quoted.

Accuracy on 483 unseen questions

Five domains, locked test, November 2026. Dot = accuracy; bar = 95% confidence interval. Paired tests are against Danphe Smart mode.

Danphe Smart mode

Reference

0.0%

Claude Opus 5.5 alone

p = 0.13 · statistical tie

0.0%

GPT-6.1 Sol alone

p = 0.002 · Danphe significantly better

0.0%

Preregistered on 5 November 2026; exact McNemar paired tests; failures count as wrong. Average model cost $0.24 per question. How we measured

$0.24average model cost per question: Smart mode on the locked test.
100%quotes on the right passage, contracts: ContractNLI, n = 100.
82.1%quotes on the right passage, science papers: QASPER, n = 100.

Danphe or one model on its own?

Including the rows where one model alone is the better choice.

Danphe Smart mode compared with one top model on its own, including where the single model is better
What you get	Danphe Smart mode	One top model alone	Who comes out ahead
Accuracy, short questions (483 unseen)	85.1%	83.4% Claude Opus 5.5 81.0% GPT-6.1 Sol	Ties the best (p = 0.13); beats GPT-6.1 Sol (p = 0.002)
Precision of long answers (100 unseen)	93.0% of statements supported	95.1% (GPT-6.1 Sol)	A single careful writer did better
Every statement checked against your sources	Yes, with the exact quote	No	Danphe
Disagreement between models	Shown, with both readings	Not visible: one voice	Danphe
Audit record	Every answer, hash-chained	No	Danphe
Model cost per question	$0.24 on average	Lower	Single model
Time to answer	Seconds to minutes	Faster	Single model

Locked tests, November 2026. The long-form test was preregistered and Danphe did not pass it. Read where Danphe is not better

Use cases

Built for work that gets checked

Evidence

An audit record for every answer

Download an evidence pack for any answer, as JSON for your systems or as a printable page for a reviewer. Records are hash-chained and anchored with a signing key, so a later change to a record can be detected.

{
  "statement": "Affiliates may receive data without notice",
  "label": "AGREED",
  "quote": { "doc": "MSA.pdf", "section": "7.2" },
  "checks": ["gpt-6.1-sol: support", "grok-4.7: support"],
  "prev_hash": "9f2c…e41a",
  "hash": "c07b…5d19"
}

Every model's answer
Who answered, in what order, with which settings.
Sources and quotes
The exact passages each statement relies on.
Every check
Each checker's verdict, including the ones that disagreed.
Hash chain
Tamper-evident links between records, verifiable offline.

For developers

The same checks, from your own systems

Send a question and your files over the API and get the answer back with every statement's label, quote and audit record.

Request

curl -X POST https://app.danphecouncil.com/api/v1/deliberations \
  -H "Authorization: Bearer $DANPHE_KEY" \
  -H "Idempotency-Key: $(uuidgen)" \
  -H "Content-Type: application/json" \
  -d '{"workspace_id": "ws_…", "mode": "auto",
       "question": "Can they share our data?",
       "documents": [{"file_id": "f_…"}]}'

API keys are part of Pro Max; create one per integration and revoke it any time.
REST under /v1, an OpenAPI 3.1 contract, RFC 9457 errors and cursor pagination.
A retried write with the same Idempotency-Key never runs twice.

Read the API guide

Response (shortened)

{  "id": "4c1f…",1  "status": "completed",2  "path": "COUNCIL",3  "answer": "Yes. Sharing with affiliates needs…",  "claims": [{    "id": "c1",    "text": "Sharing with affiliates needs no prior notice.",    "label": "CONTESTED",4    "agreement": "agreed",5    "evidence": [{ "doc_id": "d1", "pages": [7],      "quote": "without prior notice to the Disclosing Party" }]6  }],  "guarantee": { "valid": false },7  "audit": { "record_id": "dar_…", "record_sha256": "c07b…" }8}

the run; follow it on /events
queued, running, completed
Smart routing: FAST or the full COUNCIL
not certified without a calibration
shown as “Models agree”
the exact quote
true only with a signed-off calibration
the signed audit record

Security

Your documents stay yours

We hold no security certifications yet; SOC 2 is planned. Here is what is in place today.

Read the security overview

Encrypted per tenant
Encrypted in transit and at rest, with a separate key for each organisation.
Isolated by design
Database row-level security keeps every organisation's data apart.
Never used for training
Your files and questions are not used to train models.
Tamper-evident audit
Every answer's record is hash-chained and signed.

Who's behind this

Who's behind Danphe Council

Danphe Council comes from Danphe Intelligence Lab, whose research is on calibration and uncertainty: knowing how far to trust an AI answer.

Built by Danphe Intelligence Lab

Danphe Intelligence Lab is an AI research company founded in Nepal in 2024. Its research is on making machine-learning systems say how sure they are, and be right about it. That is the question behind every label in Danphe Council.

Kathmandu, Nepal
Danphe Intelligence Lab Pvt. Ltd. (Nepal) and Danphe Intelligence Lab LLC (USA)

danphelab.com GitHub [email protected]

Research behind the labels

Calibration
Measuring whether a model's confidence matches how often it is right (ECE, temperature scaling).
Uncertainty and abstention
Knowing when a system should not answer, and saying so, instead of guessing.
LLM evaluation
Preregistered tests on unseen data, with the rules written down before the run.

Design partner programme

3 places for legal, audit or compliance teams

3 months of Pro Max free, in exchange for feedback while we shape the product with you.

Apply as a design partner

Pricing

Start free, upgrade when it matters

Every plan uses several models from different companies that check each other.

Free

Try the council

$0/ month

60 credits a month, no extra usage

Start free

Pro

Pro Max

Your own councils and long documents

$199/ month

800 credits a month, more on request

Start with Pro Max

Compare plans and credits

FAQ

Questions people ask first

Is Danphe more accurate than the best AI model?

Not on short questions, and we don't claim it is. On 483 unseen questions Smart mode scored 85.1% against 83.4% for Claude Opus 5.5 alone, a statistical tie (p = 0.13), and was significantly better than GPT-6.1 Sol alone (81.0%, p = 0.002). On long answers a single model was more precise. What Danphe adds is a check of every statement against your sources, visible disagreement and an audit record. See the full results.

What is the difference between “Confirmed” and “Models agree”?

“Confirmed” is only used in a workspace with a signed-off calibration, where the error rate of confirmed statements has been measured. Without one, the same evidence is labelled “Models agree”. No label is a guarantee that a statement is true.

Which models are in the council?

Free uses a fixed council of four models from different companies. Pro and Pro Max use the flagship council (Claude Opus, GPT-6.1 Sol, Grok 4.7 and Kimi K3), and you can choose the model for each seat. A model never checks a statement written by its own model family.

Are my documents used to train models?

No. Your files, questions and answers are not used to train models, by us or by the model providers we use under their enterprise terms.

Do you have SOC 2 or ISO 27001?

Not yet. We hold no security certifications today; SOC 2 is planned. The security page lists what is in place now.

When should I not use Danphe?

For drafts, brainstorming and quick questions where a wrong answer is cheap, one model on its own is faster and costs less. Use Danphe when you need to show why an answer holds up.

Ask your first question in two minutes.

Free plan, 60 credits a month, no card needed.

Start free Book a demo

Legal, audit or compliance team? A 30-minute call is the fastest way to see if it fits.

AI answers you can actually prove

Same question, two kinds of answer

How a question is checked

Step 1: Reading sources

Step 2: Models answer on their own

Step 3: Cross-checking every statement

Step 4: Writing the answer

Why it's fast and affordable

Every statement gets a verdict

Disagreement is shown, not hidden

Measured, not promised

Danphe or one model on its own?

Built for work that gets checked

An audit record for every answer

The same checks, from your own systems

Your documents stay yours

Who's behind Danphe Council

Built by Danphe Intelligence Lab

3 places for legal, audit or compliance teams

Start free, upgrade when it matters

Free

Pro

Pro Max

Questions people ask first

Ask your first question in two minutes.