Project Lattice — Disposition Explorer

Prototype scope: package + container registry and CI/CD subsystems of gitlab-org/gitlab (109 MRs sampled)

Trajectories

Stakeholders

Code Areas

Signals

Dispositions

◉

Select a code area

Choose an area from the sidebar to explore its disposition trajectories and stakeholder stances.

◉

Select a stakeholder

Choose a stakeholder from the sidebar to see their dispositions across all code areas.

LLM Constraint Generator

Select a code area to see the exact constraint checklist an AI agent would receive before writing code. This is what makes institutional knowledge computable.

disposition-context.txt

# Enter a code area above and click Generate to see the LLM constraint output. # # This is what an AI agent receives before writing code: # - Active trajectories (what the team moved AWAY from and TOWARD) # - Stakeholder stances (who reviews this area and what they care about) # - Confidence levels (how certain we are about each constraint) # # The agent uses these constraints to produce MRs that align with # the team's established patterns and preferences.

Simulate Agent Context Injection

This visualizes the system prompt that would be injected into an AI agent's context, and the behavioral difference it creates. Select an area and click "Visualize" to see the constraint injection flow. No LLM call is made — the disposition data is live from ClickHouse, the behavior description is illustrative.

System Prompt (being constructed)

system-prompt.txt

# Click "Visualize" to see the system prompt # being constructed from live disposition data.

Agent Behavior Change

Illustrative — not a live LLM call

behavior-diff.txt

# The right panel will show how disposition # context changes the agent's behavior.

Loading graph data...

Disposition Stances

Supportive

Cautious

Opposed

Stakeholder

Code Area

Real Evidence: Before & After

These are actual code outputs from our evaluation study — the same LLM, same task, with and without a disposition constraint. Not mockups.

Model: Claude Sonnet 4, temperature 1.0, n=10 per condition per scenario.

Without disposition context (baseline)

baseline output

With disposition constraint

constrained output