Free tool · works now
Evidence-to-Claim Workbench
Audit meeting notes and executive summaries for unsubstantiated claims, vague quantifiers, and missing proof points.
Runs 100% client-side. Surface ungrounded assertions, vague claims, and missing metrics in meeting notes before sending memos to leadership.
826 characters · 13 lines
# Evidence-to-Claim Audit Report **Overall Evidence Confidence:** 38% (3 High, 0 Moderate, 6 Ungrounded) ## Audited Statements & Claims | Line | Claim | Category | Confidence | Evidence Grounding | Gaps & Flags | | :--- | :--- | :--- | :--- | :--- | :--- | | 1 | Executive Briefing — Infrastructure Migration | assertion | 35% (ungrounded) | None | Unanchored assertion: no metric, date, or named owner detected | | 5 | 1. Database migration was a huge success and everyone is happy with latency. | assertion | 10% (ungrounded) | None | Vague quantifier: "huge" lacks empirical grounding; Unanchored assertion: no metric, date, or named owner detected | | 6 | 2. SQLite FTS5 query latency dropped by 74% from 42ms to 11ms on production benchmarks. | metric | 75% (high) | Cites verifiable metric or measurement | None | | 7 | 3. We will probably deprecate the legacy cloud sync service sometime next quarter. | assertion | 10% (ungrounded) | None | Vague quantifier: "sometime" lacks empirical grounding; Tentative assertion: "probably" denotes unconfirmed status; Unanchored assertion: no metric, date, or named owner detected | | 8 | 4. Marcus committed to deploying the ScreenCaptureKit loopback driver to staging by 2026-10-18. | action | 80% (high) | Named owner or attributed stakeholder; Pinned ISO date or concrete milestone | None | | 9 | 5. The team saw massive improvements in battery consumption during live transcription. | assertion | 10% (ungrounded) | None | Vague quantifier: "massive" lacks empirical grounding; Unanchored assertion: no metric, date, or named owner detected | | 10 | 6. Local Whisper Base model memory footprint measured at 148 MB on Apple M3 Max hardware. | metric | 75% (high) | Cites verifiable metric or measurement | None | | 11 | 7. As discussed in previous syncs, customers definitely prefer on-device storage. | assertion | 10% (ungrounded) | None | Vague quantifier: "as discussed" lacks empirical grounding; Unanchored assertion: no metric, date, or named owner detected | | 12 | 8. Security team approved air-gapped deployment architecture with zero external egress. | decision | 35% (ungrounded) | None | Unanchored assertion: no metric, date, or named owner detected |
Why meeting summaries suffer from hallucinated consensus
In modern knowledge work, meeting summaries are frequently distributed as gospel truth to stakeholders who were not in the room. When an automated AI notetaker or hurried meeting host condenses a forty-five-minute debate into five bullet points, a dangerous transformation often takes place: tentative ideas become firm decisions, partial support becomes universal agreement, and unverified estimates become committed roadmaps.
This phenomenon—hallucinated consensus—is one of the primary drivers of organizational misalignment. A summary that proclaims "The engineering team agreed to migrate the database by next month" can trigger cascading commitments across marketing, sales, and executive leadership, even if the lead architect merely said "We might look into it if sprint velocity allows."
The anatomy of an evidence-grounded claim
To protect project integrity and leadership trust, internal documentation must distinguish between verified evidence and unanchored assertion. This workbench analyzes text against three empirical markers:
1. Metric Grounding
Statements that include concrete numbers, percentages, benchmark times, or financial quantities carry verifiable evidentiary weight. Writing that "SQLite FTS5 query latency dropped from 42ms to 11ms" provides an auditable claim that can be verified against monitoring dashboards. Conversely, writing that "search became significantly faster" is an unanchored assertion that prevents objective evaluation.
2. Attribution Grounding
Every committed decision or action item requires an explicit owner. Using passive voice ("It was decided that...") or diffuse pronouns ("The team will handle...") creates accountability voids where tasks languish uncompleted. When an individual (@marcus, Elena Vance, Lead Architect) is directly attached to a commitment, accountability is preserved.
3. Temporal Grounding
Vague time horizons like "soon", "next quarter", or "as discussed previously" introduce schedule ambiguity. High-confidence notes pin commitments to unambiguous calendar dates (e.g., 2026-10-18) or formally tracked sprint milestones.
How the workbench audits your notes
When you paste notes into this tool, the deterministic engine scans each sentence for linguistic signals:
- Vague quantifier detection: Flags exaggerations like huge, massive, dramatic, vast majority, and everyone. These words often mask an absence of actual data.
- Tentative assertion detection: Identifies qualifiers like probably, maybe, hoping, and assuming that signal uncertain consensus.
- Formula-safe CSV export: When exporting audit logs for project reviews, cells starting with arithmetic operators (
=,+,-,@) are automatically guarded with leading apostrophes to protect spreadsheet users against CSV injection attacks.
Verifying meetings with source audio
The most reliable safeguard against hallucinated summaries is retaining access to verbatim source audio. While cloud meeting bots discard raw audio or store sensitive customer conversations on external servers, Oyma records meetings locally via macOS loopback audio. You can always cross-reference any generated summary against the synchronized local Whisper transcript. Learn more about on-device AI summaries or use our Meeting Record Builder to format structured decision logs.
Frequently asked questions
Questions
Does this tool send my internal meeting notes to an external AI model?
No. All statement extraction, grounding checks, and confidence scoring execute deterministically in your browser tab. Zero network requests are made, and your text is never transmitted or stored.
Why are vague quantifiers like "huge" or "everyone" penalized?
Vague quantifiers obscure reality. In executive meetings, saying "everyone agreed" often hides unresolved objections, while "massive performance gain" prevents teams from verifying whether latency dropped by 5% or 75%. Grounding requires verifiable metrics.
What makes a statement high-confidence?
A high-confidence claim anchors its assertion to objective evidence: a measured metric (e.g. 74% reduction, 11ms), a specific named owner, or a calendar milestone (e.g. 2026-10-18).
Can this replace human review of meeting records?
No. This workbench acts as a preflight linter for meeting records and executive summaries. It surfaces ambiguous language and unanchored assertions so team leads can corroborate details before distributing memos.
How does the Mac app approach meeting accuracy?
The desktop app records system audio directly via macOS ScreenCaptureKit and performs on-device Whisper transcription. It allows you to query the original verbatim transcript to verify every summarized decision.
Tired of inaccurate AI meeting summaries?
Run private on-device summaries on your Mac with verbatim transcript cross-referencing.