Skip to main content

​
The problem this solves

If you keep ingesting documents and transcripts into a POT, sooner or later two facts will contradict each other. Typical examples:
  • An old proposal says “volume discount starts at 500 units” — a recent email says “starts at 300 units”
  • A meeting mentions “the single point of contact is Carlos” — a contract says “single point of contact is Laura”
  • A quarterly report says “growing 12% YoY” — the investor deck says “8% YoY”
Without management this becomes noise — kb2b would answer with low confidence or contradict itself depending on which fact got retrieved. Contradiction resolution is the formal flow for the team to decide which one stands.

​
How they’re detected

kb2b detects contradictions automatically as it ingests a new fact:
  1. For each new fact, it looks for semantically similar facts already in the POT.
  2. An LLM evaluates whether the relationship is: supports, refines, derives_from, or contradicts.
  3. If contradicts, an edge is created between the two facts with that relationship type.
  4. Each edge has a weight (0.0–1.0) indicating how strong the contradiction is (concept overlap + score divergence).

​
Where they appear

Contradictions show up in two places:
  • Persistent banner at the top of the entire Knowledge section (/dashboard/knowledge and sub-routes) — shows total count and per-severity. If there are no contradictions, the banner hides.
  • Dedicated page at /dashboard/knowledge/constitution — the full list, expandable, with AI Triage if there are ≥ 3.

​
Severities

kb2b classifies each contradiction by its weight: Severity is computed automatically from the edge weight — not manually assigned. Contradictions sort descending by weight, so critical ones always appear first.

​
AI Triage

When there are 3 or more active contradictions, kb2b shows an automated Triage panel with Claude-powered recommendations: it groups related contradictions, suggests where to start, and summarizes the conflicts. It does not resolve for you — it only helps you prioritize.

​
The two-phase workflow

Resolving a contradiction has two phases, intentionally separated:
The split is deliberate: stances accumulate asynchronously across the team, and the final decision happens once there’s enough evidence.

​
Phase 1 — Stances (all editors)

Each contradiction shows the two fact cards side by side. Under each is a set of three buttons:
  • “Support this fact” (confirm_a or confirm_b depending on the side) — you vote in favor of the fact.
  • “Not a contradiction” (dismiss) — you mark the edge as a false positive.
Each stance can include optional notes explaining why. Each one also triggers a score delta that propagates to the involved facts:
  • Confirming a fact raises its POT Score proportional to your expertise level.
  • The opposing fact drops proportionally.
Counts are shown live: “3 support A · 1 supports B · 0 dismiss”. If you change your mind, your stance is replaced (not duplicated).

​
Phase 2 — Resolution (editor / admin / curator)

Once there’s enough evidence, someone with permissions makes the final call. There are 5 options: Each resolution can include notes explaining the rationale. When confirmed, kb2b mutates the edge in SciPot: contradicts → resolved_contradiction, marks who resolved and when, stores the notes, and applies the final score deltas.

​
Reverting a resolution

If you made a mistake or new evidence appears, the original resolver can revert their resolution. Click “Revert” on the resolved card:
  • The edge goes back to contradicts (active state).
  • Previous stances are kept as history.
  • The revert is recorded in the audit trail.
Only the person who resolved can revert their own resolution. Other editors can take a new stance if they disagree, but can’t overwrite directly.

​
Constitutional axioms and contradictions

When a Constitutional fact (POT Score 1.0) participates in a contradiction:
  • kb2b shows an explicit amber warning in the UI: “This fact contradicts a constitutional axiom. Resolving against the axiom requires administrator intervention.”
  • It does NOT block the resolution technically — the call stays with the responsible person.
  • The warning is there so you think twice: if you’re about to tear down an axiom, the right move is to go to POT Constitution and downgrade or remove it formally with score re-propagation.
The SciPot whitepaper describes axioms as “immune to stance penalties”. In kb2b the current implementation is warning-aware, not hard-block: the UI warns you but lets the action through. If your organization needs real blocking (regulatory audit, compliance), confirm with your admin before operating on contradictions that touch axioms.

​
What gets logged (audit trail)

Every action on a contradiction generates an activity event logged for audit: Historical stances appear in the “Team positions” panel inside each contradiction — who voted what, when, with what notes, and what score delta it caused. Resolved contradictions are kept in a separate panel: “My resolutions” (yours) and “Resolved by others” (the team’s) — dimmed but accessible for review.

​
Notifications

In this version kb2b does NOT send push notifications when:
  • A new contradiction is detected
  • Someone takes a stance on a contradiction that involves you
  • A contradiction you originated or voted on is resolved
All changes are in the activity feed and visible when you visit the Knowledge section. Push notifications are on the roadmap.

​
When NOT to use this flow

  • If you find a fact extracted incorrectly (it contradicts nothing, it’s just wrong), don’t open a contradiction — use 👎 curation from the fact inspector (see Explore and discuss facts).
  • If you want to discuss a fact but it doesn’t contradict another, use the Discussion tab, not this flow.
  • If two facts appear to contradict but are actually about different things, resolve with dismissed or both_valid — don’t force a binary decision where it doesn’t apply.

​
Permissions

Page in progress Phase 1 — banner, stance workflow, and resolution modal screenshots coming in Phase 2.