docs: add DECISIONS.md and rewrite PROJECT.md #8

Merged
mathias merged 2 commits from chore/project-context-and-decisions into main 2026-06-22 18:10:32 +00:00
Owner

What

Two documentation files establishing the research foundation for jepa-fx-risk.

DECISIONS.md (new)

Five Architecture Decision Records:

  • ADR-001 — Architecture: TS-JEPA + SIGReg as Phase 1 backbone. Supersedes the informal TS-JEPA-only decision. Records the systematic four-variant evaluation (TS-JEPA, LeJEPA, MTS-JEPA, Var-JEPA) and rationale for the hybrid.
  • ADR-002 — Data: DUKASCopy hourly G10 FX, 2008–2022 train / 2023 val / 2024 test.
  • ADR-003 — Research roadmap: Phase 0 (MAE gate) → Phase 1 (autoresearch sweep) → Phase 2 (MTS-JEPA multi-resolution) → Phase 3 (internal data).
  • ADR-004 — Evaluation separation: Go harness owns all metrics, Python owns training only.
  • ADR-005 — Compute: Blackwell GPU on koala, PyTorch cu130 wheel.

.context/PROJECT.md (rewritten)

Replaces the template-go-web hostexecutor placeholder with the actual project context. Covers purpose, architecture decision summary, stack, repo layout, phase structure, data, conventions, eval metrics, benchmarks, and agent guidance.

After merging, run task context:sync to regenerate CLAUDE.md, AGENTS.md, and .cursorrules from the updated PROJECT.md.

Related issues

Addresses the identity mismatch noted in Issue #6 (PROJECT.md / CLAUDE.md still said "hostexecutor").
Provides the DECISIONS.md referenced in Issue #3 ("Architecture: TS-JEPA as starting implementation").

## What Two documentation files establishing the research foundation for jepa-fx-risk. ### `DECISIONS.md` (new) Five Architecture Decision Records: - **ADR-001** — Architecture: TS-JEPA + SIGReg as Phase 1 backbone. Supersedes the informal TS-JEPA-only decision. Records the systematic four-variant evaluation (TS-JEPA, LeJEPA, MTS-JEPA, Var-JEPA) and rationale for the hybrid. - **ADR-002** — Data: DUKASCopy hourly G10 FX, 2008–2022 train / 2023 val / 2024 test. - **ADR-003** — Research roadmap: Phase 0 (MAE gate) → Phase 1 (autoresearch sweep) → Phase 2 (MTS-JEPA multi-resolution) → Phase 3 (internal data). - **ADR-004** — Evaluation separation: Go harness owns all metrics, Python owns training only. - **ADR-005** — Compute: Blackwell GPU on koala, PyTorch cu130 wheel. ### `.context/PROJECT.md` (rewritten) Replaces the `template-go-web` hostexecutor placeholder with the actual project context. Covers purpose, architecture decision summary, stack, repo layout, phase structure, data, conventions, eval metrics, benchmarks, and agent guidance. After merging, run `task context:sync` to regenerate CLAUDE.md, AGENTS.md, and .cursorrules from the updated PROJECT.md. ## Related issues Addresses the identity mismatch noted in Issue #6 (PROJECT.md / CLAUDE.md still said "hostexecutor"). Provides the `DECISIONS.md` referenced in Issue #3 ("Architecture: TS-JEPA as starting implementation").
mathias added 2 commits 2026-05-28 05:42:44 +00:00
docs: rewrite PROJECT.md with actual jepa-fx-risk project context
CD / Lint / Test / Vet (pull_request) Failing after 4s
CD / Build & Import (pull_request) Has been skipped
CD / Deploy via GitOps (pull_request) Has been skipped
50635cd80e
mathias merged commit e616575979 into main 2026-06-22 18:10:32 +00:00
Sign in to join this conversation.
No Reviewers
No labels
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: mathias/jepa-fx-risk#8