Four records may describe one incident
Same checkpoint, unit, and 36-hour period. No Incident connects them.
Product direction / 2026
From records to collective intelligence.
Bayanat can do more than store completed analysis. It can help a team notice what is missing, remember why decisions were made, and improve the methodology together.
The model finds relevant material. The team makes accountable decisions.
Signals waiting for a teamSynthetic scenario
A methodology change has a history
A revised evidence threshold may affect 183 prior decisions.
Reviewer disagreement reveals ambiguity
The team diverges when location evidence is indirect.
Not automated conclusionsQuestions, precedents, learning
The product shift
Change the unit of teamwork.
TodayA record moves through statuses.
→
ProposedA question gathers evidence and decisions.
01 / Investigation Gap Finder
Work begins with an unanswered question.
Instead of waiting for someone to manually notice connections, Bayanat surfaces possible patterns and contradictions. It never declares an incident. It creates a reviewable question with its evidence attached.
- Questions have owners, status, evidence, and a decision history.
- Contradictory sources stay visible instead of being averaged away.
- Related multilingual passages can be found without identical wording.
Question workspaceSynthetic example
Pattern found
Could four disappearances be one incident?
Four independent records mention the same checkpoint, unit, and 36-hour period.
Agreement
Checkpoint, Unit 18, 12 to 13 May
Conflict
Two witnesses name different transfer destinations
02 / Living Methodology
Every difficult decision can become team memory.
A methodology document explains the general rule. Real cases reveal the hard boundaries. Bayanat can preserve those decisions as precedents, then show their relevance when a similar case or changed definition appears.
- Precedents include evidence, rationale, dissent, and approval.
- A methodology change produces a reviewable impact set.
- Exceptions remain visible and never silently become rules.
Methodology impactDefinition 4.2
Proposed change
Verified detention location
The minimum evidence threshold has changed. Preview the historical effect before publishing the revision.
0records identified for review
0already aligned with the new guidance
Relevant team precedent
Decision 882 used the same evidence pattern and documented one minority view. Reviewed 18 March.
03 / Calibration Lab
Measure ambiguity, not employee productivity.
A small set of records is reviewed independently. Bayanat finds where interpretations diverge and turns the disagreement into a focused methodology discussion. The goal is shared judgment, not staff ranking.
- Blind review prevents early answers from anchoring the group.
- Results are aggregated around methodology boundaries.
- The session ends with a precedent or clarified guidance.
Blind review synthesis5 records / 6 reviewers
Direct evidence
88
Indirect location
51
Named source
77
Boundary requiring discussionArbitrary detention versus enforced disappearance when the location is supported indirectly.
One connected operating model
Each concept improves the next.
The value is not three isolated features. It is a loop where investigations produce decisions, decisions produce precedents, and disagreement improves the methodology.
01Evidence enters Bayanat
02Possible gaps become questions
03Teams make decisions
04Hard decisions become precedents
05Calibration exposes ambiguity
06Methodology improves
Model boundary
Where E5 helps, and where it stops.
Multilingual-e5-small is a retrieval model. That narrow role is useful because every suggestion can point back to existing material.
E5-small
Find semantically similar records, passages, label definitions, and prior decisions across languages. Rank candidates for a person to inspect.
Rules
Enforce workflow requirements, permissions, required evidence, and explicit methodology checks.
Optional small LLM
Summarize retrieved material or draft a question. Its output remains a proposal tied to sources.
Human team
Determine relevance, assess credibility, resolve contradictions, apply methodology, and make the final analytical decision.
Smallest serious prototype
Prove one investigative question end to end.
Do not start by building all three systems. Start with one real analytical pattern and make the evidence trail excellent.
Select one recurring investigative question
For example, records that may describe the same detention event.
Retrieve candidate records with E5
Show the exact passages and metadata that caused each candidate to appear.
Capture the team decision
Record accepted links, rejected candidates, unresolved conflicts, and rationale.
Evaluate usefulness
Measure discoveries, missed connections, reviewer time, and false candidates.