ConcordEnter the demo

Case study ยท Interactive demo

Concord

Contradiction review for knowledge bases

Contradictions in a knowledge base quietly become confident wrong answers. A detector flags conflicting passages; Concord is where domain experts settle them, with every decision recorded and measured.

Enter the demo Watch the 40-second walkthrough

Walkthrough

From a flagged passage to a measured decision

Forty seconds, six scenes: detection, the flag record, routing, review, the audit trail and the dashboard. All labels are invented.

Upstream pipeline

Before a flag reaches a reviewer

Before anything reaches a reviewer, an upstream pipeline turns scanned, multilingual public-health guidelines into a searchable knowledge base and surfaces the passages that contradict each other.

Stage 01

Read the scans

Many source documents are scans with no usable text layer, so every page is read as an image.

  • One model transcribes each page into structured markdown.
  • A different model re-reads the same page independently, without seeing the first transcription, so agreement is real corroboration rather than a rubber stamp.
  • Rule-based checks look for dropped figures, broken tables and truncation.
  • Pages that fail get one targeted retry. Pages that still fail are held back for a human instead of entering the knowledge base.
From flag to decision
  1. Raw flagsnoisy by design
  2. Verifiedmost discarded
  3. Human decisionfinal call

  • Two independent readers for every page

    When the same model both transcribes and checks, it makes the same misreads and certifies them. A different, independent reader turns the check into real corroboration.

  • Quarantine over guesswork

    A page that can't be read confidently is held back for a person instead of quietly entering the knowledge base with errors.

  • High recall first, precision second

    Missing a real contradiction is costlier than flagging a false one, so detection casts a wide net and a separate verification pass does the filtering.

Details such as models, prompts and thresholds are intentionally withheld.

Watch a simulated run

How it works

Five steps, one loop

The detector proposes; people decide; the record keeps everyone honest.

  1. 01

    Detect

    An upstream detector flags passages that contradict each other, across two documents or inside one. Each flag carries the cited text, a type, reasoning and an optional confidence.

  2. 02

    Route

    A document-to-reviewer policy decides who sees what. Admins can override per item; anything unresolved waits in the admin inbox.

  3. 03

    Review

    Reviewers read both documents side by side with the citations highlighted, then choose A, B, reconcile (C) or not-a-conflict (D).

  4. 04

    Record

    Every save appends an immutable revision and an audit event. Nothing is overwritten, so changes of mind stay visible.

  5. 05

    Measure

    Coverage, per-reviewer progress, the A/B/C/D mix and a consensus rate from items deliberately reviewed twice.

Product decisions

The calls that shaped it

Each row is a decision, the reason for it and what it cost.

Several reviewers per item, on purpose

The same item can be assigned to more than one reviewer, and "done" is tracked per reviewer.

Why
Agreement between independent experts is the only honest signal of how ambiguous the corpus is. Double review turns that into a consensus metric.
Trade-off
It costs reviewer time. Consensus reads 100% before anything has been double-reviewed (vacuously true), which the dashboard explains rather than hides.

C and D require a comment

Reconcile (C) and not-a-conflict (D) can't be saved without text.

Why
C's comment is the reconciled statement that goes back into the knowledge base. D overrides the detector, so it has to be defensible later.
Trade-off
A little friction on every C/D decision. A and B stay one click.

Append-only revisions and audit

Saves never update in place: each one appends a revision and an audit event.

Why
An edit from A to C means something different from an original C. Reviewers' second thoughts are data.
Trade-off
More rows and a "latest revision" query everywhere. Storage is cheap; lost history isn't recoverable.

An honest highlighter

Citations are fuzzy-matched against messy, OCR-ish document text. When a citation can't be located, the UI says so instead of guessing.

Why
Detectors sometimes cite text that appears in no document. A wrong highlight is worse than none, because reviewers trust what's lit up.
Trade-off
Some items show a "couldn't auto-locate" notice and the reviewer has to read more. Trust is worth the extra minute.

Invite-only, passwordless sign-in

Email one-time codes for a closed list of invited users, with tiered roles (super-admin > admin > reviewer).

Why
A small expert group doesn't need passwords to forget. Invitations are the access-control list.
Trade-off
Depends on email delivery. A dev-only impersonation bypass speeds up local work and throws if it ever runs in production.

Screens

Take the tour

Every screen runs in your browser on fictional data. Changes persist locally until you reset the demo.

Admin screens need an admin persona: switch it from the persona switcher at the right of the demo banner.