Skip to content

Combine instrument contributions without double-counting coverage - #61

Merged
satyalyadav merged 6 commits into
mainfrom
feat/mixed-instrument-coverage
Nov 6, 2026
Merged

satyalyadav merged 6 commits into
mainfrom
feat/mixed-instrument-coverage

Conversation

@satyalyadav

Copy link
Copy Markdown
Collaborator

Summary

Closes #52. Researchers can accept one study with up to six independently configured instruments across six spacecraft, then retrieve each contribution and combined geometric access.

prepare exact configurations -> accept setup revision -> run TAT-C per configuration
  -> retain simultaneous contributors
  -> combine daily area by visited cell ID union
  -> fetch combined or configuration events, daily rows, cells and spacing

Preserves revisions, evidence, orbit association, per-instrument illumination and full-window applicability. Applies configuration/grid/work limits before calculation with one pool of at most four workers. Shared accounting, task construction and illumination validation replace duplicated logic. Combined envelopes are explicitly contributor hulls and retain each contributor's bounds.

Evidence

  • Before: No mixed workflow; review probes exposed missing horizon warnings, unlabelled calculation failures, zero-event integration fixtures and false source attribution for a daylight assumption.
    After: Public MCP regressions cover acceptance and all result scopes, known-event cell unions, half-open UTC boundaries, exact simultaneous epochs, independent area within 0.001 km², nonzero real TAT-C mixed contributions and independent daylight filtering. Horizon warnings and configuration/revision error identity persist.
  • Python 3.12 and TAT-C 3.6.0: full server suite 404 passed, one existing Authlib deprecation warning; coastline preload and git diff --check passed.
  • Code-review skill completed Standards and Spec axes; addressed their defects and duplication findings. Kiro Opus 5.5, medium independently reviewed, then approved corrected code at 43eede5 with no blocking defects. The following documentation-only commit qualifies an unverified historical eval turn cap.
  • Real tool-calling evals: legacy boundary/short repeat-2 rerun 20/20 passed. Corrected 24-hour mixed run completed ten conversations: four scored passes, six infrastructure errors consisting of a context-limit error and five provider timeouts. All five stream probes passed. The final payload compaction was verified through MCP tests, without another provider rerun.
  • Full baseline comparison on 65 paired scored rows was 58 baseline passes versus 56 issue-run passes, including boundary 5/5 to 3/5 and short 5/5 to 4/5. The full issue run stopped at 83/85 rows. Rescoring older scoped mixed traces changed 6/6 to 4/6 because two replies falsely sourced a study assumption; corrected fixtures remove that claim. See evaluation details. Provider coverage remains incomplete.

Merge Danger

Door: two-way

Blast Radius: server

Legacy and single-instrument regression checks pass, but shared accounting and server instructions also affect those workflows. Merging to main rebuilds/deploys the MCP image and restarts LibreChat.

Bloat review removed duplicate headers, cell aliases, repeated revisit metrics and pagination checks, and shared repeated logic. Full accepted snapshots remain in finish to preserve source evidence and rationales: about 74 KB for finish, 72 KB for preparation, and 24 KB for the largest measured fetch. This can exhaust model context. Finish is the only full-snapshot view; a client truncating it cannot retrieve the complete snapshot elsewhere. This limitation and the provider eval gaps remain documented. No new dependencies or persistent storage.

@satyalyadav
satyalyadav merged commit c9f6a18 into main Nov 6, 2026
1 check passed
@satyalyadav
satyalyadav deleted the feat/mixed-instrument-coverage branch November 6, 2026 19:14
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Combine instrument contributions without double-counting coverage

1 participant