Skip to content

Covariance intersection as a visual-update baseline #116

Open
Open
Covariance intersection as a visual-update baseline#116
Labels
enhancementNew feature or requestready-for-humanRequires human implementationresearch
Milestone
Stretch options (need an ADR)

Description

@Telschow
Telschow
opened on Oct 18, 2026

Fusing a relative visual constraint with an unknown correlation to the state is the core of the visual overconfidence problem. Covariance intersection is the standard conservative answer. Compare it with the single anchor and the stochastic clone, behind a flag, defaults unchanged.

Needs an ADR. After the gate issue is decided. Relates to #44.

RICE

Judgment estimates, not measurements. Same scales as contested-nav docs/prioritisation.md: Reach 1 to 10, Impact 0.25 to 3, Confidence 50, 80 or 100%, Effort in ideal working days.

Reach Impact Audit boost Confidence Effort (days) Score
4 1 none 50% 5 0.40

Score = Reach x Impact x Audit boost x Confidence / Effort. Needs an ADR.

Activity

  1. Telschow
    added this to the Stretch options (need an ADR) milestone on Oct 18, 2026
  2. Telschow
    added
    research
    enhancementNew feature or request
    needs-triageMaintainer needs to evaluate this issue
    ready-for-humanRequires human implementation
    and removed
    needs-triageMaintainer needs to evaluate this issue
    on Oct 18, 2026

  3. Telschow commented on Oct 24, 2026

    @Telschow
    Telschow
    on Oct 24, 2026
    OwnerAuthor

    This was generated by AI during triage.

    Triage notes: ready-for-human

    Why a human: it needs an ADR, and the maintainer writes the decision section. It also follows the B1 outcome (PR #187, ADR-0036), and which baseline sits next to the single anchor and the stochastic clone is a judgement call.

    Current state: not implemented. ADR-0030 and ADR-0034 name covariance intersection as a baseline, not a candidate, and ADR-0034 tracks it here.

    Acceptance criteria:

    • Declared in an ADR before the run: the scenes, seeds and columns.
    • Covariance intersection runs behind a flag. Defaults are unchanged and vision_enabled stays False.
    • Results use the same columns as the single anchor and stochastic clone rows.
    • The overconfident single-anchor case stays as a control row.
    • Rows are generated, and each page statement is tied to them by a test.
    • No new runtime dependency (S1).
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or requestready-for-humanRequires human implementationresearch

    Projects

    No projects

      Milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions