Workbench · Models & evaluation / Data & provenance
DeckScope
Trace a document's claims to evidence, contested points and unanswered questions.
Repository: CinvanaAI/DeckScope
Recorded synthetic example. Source with a complete offline example.
See the idea in action.
Six examined claims, two contested claims, two omissions, two unknowns; HTML/Markdown/JSON reports.
Example input
Bundled synthetic Acme Flow pitch deck.
Selected complete fields from the captured result
{
"kind": "synthetic demonstration",
"provider": "mock",
"research_backend": "none",
"source_input": "deckscope/examples/sample_deck.md",
"notice": "Fixture-generated analysis. These are illustrative workflow outputs, not verified real-world findings or measured analysis quality.",
"findings": {
"headline": "Two claims are contested by cited evidence (one outright, one partly supported) \u2014 deck states $47B; Independent 2026 estimates put the whole workflow-automation category at $18-24B, not the $45-50B\u2026, the deck omits Microsoft Power Automate arriving free inside E5 and 1 other gap, and two further claims could not be checked either way. Eight questions to resolve before this can be judged.",
"contested": [
{
"kind": "contested",
"text": "The workflow automation market is $47B, growing at 23% CAGR",
"why": "Use the evidence range as the working number and read the gap as a signal about how the deck was assembled.",
"delta": "deck states $47B; Independent 2026 estimates put the whole workflow-automation category at $18-24B, not the $45-50B figures circulating in vendor-sponsored reports.",
"severity": "high",
"evidence_quality": "strong",
"source_ids": [
"S1"
],
"claim_id": "C1",
"assessment": "contradicted"
},
{
"kind": "contested",
"text": "Average contract value: $28,000. Gross margin: 78%",
"why": "Consistent with the evidence retrieved.",
"delta": "deck states $28,000.; Products that run an LLM planner on every workflow execution report gross margins of 55-68% once inference is loaded into COGS, against 78-85% for\u2026",
"severity": "medium",
"evidence_quality": "moderate",
"source_ids": [
"S4"
],
"claim_id": "C4",
"assessment": "partially-supported"
}
],
"omissions": [
{
"kind": "omission",
"text": "Microsoft Power Automate arriving free inside E5",
"why": "The marginal cost of the bundled option is zero for a buyer already paying for E5.",
"delta": "",
"severity": "high",
"evidence_quality": "",
"source_ids": [
"S2"
],
"claim_id": "",
"assessment": ""
},
{
"kind": "omission",
"text": "Buyer budgets are substitution, not net-new",
"why": "A TAM built by counting companies overstates the reachable market when the buyer must first stop paying someone else.",
"delta": "",
"severity": "high",
"evidence_quality": "",
"source_ids": [
"S6"
],
"claim_id": "",
"assessment": ""
}
],
"unverified": [
{
"kind": "unverified",
"text": "SAM: $6B (mid-market North America)",
"why": "Ask the founder directly; nothing retrieved settles it.",
"delta": "deck states $6B; Independent 2026 estimates put the whole workflow-automation category at $18-24B, not the $45-50B figures circulating in vendor-sponsored reports.",
"severity": "low",
"evidence_quality": "none",
"source_ids": [],
"claim_id": "C2",
"assessment": "unverifiable"
},
{
"kind": "unverified",
"text": "$1.2M pipeline",
"why": "Ask the founder directly; nothing retrieved settles it.",
"delta": "No evidence was retrieved that speaks to this figure.",
"severity": "low",
"evidence_quality": "none",
"source_ids": [],
"claim_id": "C6",
"assessment": "unverifiable"
}
],
"holds": [
{
"kind": "holds",
"text": "$2,000/month platform fee plus usage",
"why": "Consistent with the evidence retrieved.",
"delta": "",
"severity": "low",
"evidence_quality": "moderate",
"source_ids": [
"S3"
],
"claim_id": "C3",
"assessment": "supported"
},
{
"kind": "holds",
"text": "18% month-over-month growth, four months running",
"why": "Consistent with the evidence retrieved.",
"delta": "",
"severity": "low",
"evidence_quality": "moderate",
"source_ids": [
"S5"
],
"claim_id": "C5",
"assessment": "supported"
}
],
"next_steps": [
"Reframe market slide around the $3-5B serviceable slice",
"Add a slide addressing Power Automate directly",
"Disclose retention and CAC payback",
"Ask the founder: What is net revenue retention on the first 11 customers?",
"Ask the founder: How many deals were competitive against Power Automate, and what happened?",
"Ask the founder: Is inference cost inside the 78% margin?",
"Verify or refute \u201cSAM: $6B (mid-market North America)\u201d \u2014 the research found nothing either way.",
"Verify or refute \u201c$1.2M pipeline\u201d \u2014 the research found nothing either way."
],
"evidence_state": "2 of 2 contested points cite retrieved evidence; 6 of 6 sources were used.",
"evidence_too_thin": false,
"evidence_reason": "ok",
"counts": {
"contested": 2,
"omissions": 2,
"unverified": 2,
"holds": 2,
"claims_examined": 6,
"contested_with_evidence": 2
},
"vertical": "deck"
}
}Try the example.
From the repository root, follow the dependency requirements in the README. This example uses supplied synthetic material.
python -m pip install -e . python -m deckscope demo --out output --format html md json
Six examined claims, two contested claims, two omissions, two unknowns; HTML/Markdown/JSON reports.
Complete setup and instructions ↗The interesting part.
Source spans connect claims to supplied model or mock assessments. The report groups those assessments and computes counts; it does not independently establish that a claim is true.
Where it came from.
Existing public exploratory document-evidence application included in the collection.
Current front door retains the broader guide and presents a labeled synthetic report path.
No code ancestry with the other evaluation tools is established by shared claims/evidence themes.