Contradictions.
When someone says one thing today and the opposite next week, Bonfiyah notices — and tells you about it before the next meeting starts.
The "wait, didn't you say..." moments.
Two weeks ago, the client said the launch had to be by June 30. Today, they're saying mid-July is fine and "we never committed to June." You think you remember it differently. You're pretty sure you do. But you're not certain enough to push back, and the recording is buried somewhere in your library.
This happens constantly — to lawyers preparing depositions, to sales reps tracking deal terms, to managers running through promised launches, to therapists noticing a client's narrative drift, to journalists vetting a source.
The information is in your recordings. You just can't reach it in time.
Contradictions is the feature that catches these moments automatically — surfaced before the next conversation, not discovered after.
How it works
- 1
Statement extraction
As each recording transcribes, Bonfiyah extracts the factual claims — what someone said happened, what they say will happen, what they assert about a price, a date, a quantity, a relationship. Each claim is anchored to a speaker, a recording, and a timestamp.
- 2
NLI-verified cross-recording reasoning
A model on Bonfiyah's backend (Anthropic's Claude, under a no-training commercial contract) compares each new claim against prior claims by the same speaker (and, where relevant, claims about the same topic from different speakers). It runs natural-language inference — does statement B logically contradict statement A? — not just keyword overlap. "We're shipping by June 30" contradicts "we never committed to June" even though the words are different, and only a verified entailment relationship is surfaced, which keeps false alarms down.
- 3
Surfaced as evidentiary before/after pairs
A verified contradiction is presented as a before/after pair: the earlier quote and the later quote, side by side, each with its speaker, date, timestamp, and a tap to play. It appears in two places — the Truth Inbox (a chronological feed of flagged pairs) and the Pre-Brief for your next meeting with that person — so you see the issue before you walk in, with both sides of it already in hand rather than reconstructed from memory.
What "contradiction" means here (and doesn't)
Contradictions is not a lie detector. It is not a credibility-scoring system. It does not assign motives. It surfaces statements that are logically inconsistent with prior statements by the same person, and lets you decide what to do with that information.
Sometimes the inconsistency is innocent — circumstances changed, the person learned something new, two people heard the same thing differently. Contradictions surfaces the conflict and gives you the recordings; you bring the judgment.
It is also tunable. In Pro AI settings, you can set thresholds: only flag contradictions involving dates, prices, deadlines, or named third parties. Or flag everything and curate the inbox manually. The default is conservative — it's better to miss a borderline case than to be the friend who keeps yelling "you said X two weeks ago!" at every dinner.
The constructive twin: Consistencies
Contradictions has two tabs. The first catches contradictions. The second — Consistencies — does the opposite: it surfaces the moments a person reinforced the same commitment across different meetings. When someone says "I'll own the rollout" in week one and independently restates it in week four, that's a pattern worth seeing too — a track record, not a trap.
It's the same cross-recording engine pointed at agreement instead of conflict: same-speaker statement-pairs from different conversations that reinforce each other, each with both quotes and the recordings behind them. You review with one tap — Save to keep it, Hide to dismiss — exactly like the contradictions inbox.
Consistencies appear once you've recorded the same person across more than one conversation. Like contradictions, they're computed on Bonfiyah's backend from your transcript text — never your raw audio, which is deleted from our servers within 7 days.
A follow-through score for the people you work with
Once the same person has made — and kept, or missed — enough commitments across your recordings, Bonfiyah computes a follow-through score for them: how often do they actually do what they said they'd do? It's drawn from Promise Tracker's kept-versus-missed record, not from a vibe, and it lives on each person's People Memory profile.
It's the same idea as the contradiction inbox, pulled forward into a single number: a reliable picture of someone's word, built from evidence you already have. Used well, it's not a scarlet letter — it's calibration. You learn whose "I'll have it Friday" means Friday, and whose needs a reminder on Thursday, without keeping a private grudge ledger in your head.
FAQ
Does this work across speakers, or only within one person's statements?
By default, only within one speaker — Contradictions flags inconsistencies in what one person said over time. A "third-party mode" extends this to claims made about a third party — useful for journalists fact-checking sources or for legal prep.
Can I delete or undo flagged contradictions?
Yes. Each flag has a "Dismiss" action that removes it from your inbox and adds it to the model's "non-contradiction" set so similar pairs aren't re-flagged. You can also delete the underlying recording, which removes all associated flags.
Does the Contradictions scan leave my device?
Contradiction detection is processed on Bonfiyah's backend by Anthropic's Claude under a no-training commercial contract. Your transcript text is analyzed for the comparison and is never used to train any model; your raw audio is deleted from our servers within 7 days, after which your recordings live only in your private iCloud.
Can I export contradictions for legal use?
Yes. Each flagged contradiction can be exported as a PDF with both quotes, timestamps, audio clips, and a chain-of-custody hash. The export format is designed for use in deposition prep and litigation; talk to your attorney before relying on any specific export for evidentiary purposes.
Used in these workflows
Where this surface earns its weight.
See the full nineteen-workflow catalog →
Want a real example?
We'll send a 90-second walkthrough of Contradictions running over a real (anonymized) sales conversation, including the moment it caught a price contradiction the rep had missed.
No spam. We use ConvertKit. See our privacy policy.