diff --git a/docs/integrations/cave-integration-spike.md b/docs/integrations/cave-integration-spike.md
new file mode 100644
index 0000000..21e0dcd
--- /dev/null
+++ b/docs/integrations/cave-integration-spike.md
@@ -0,0 +1,120 @@
+# CAVE Integration Spike (Non-invasive)
+
+## Objective
+Define a **CAVE-compatible data flow** for PyTC Client that can be implemented incrementally without changing current runtime behavior.
+
+This spike is documentation-first and establishes a contract for future work.
+
+---
+
+## Target use case in this app
+Primary use case: allow a user to run existing PyTC workflows (training, inference, proofreading, and visualization handoff) while optionally publishing workflow artifacts to a CAVE-aligned backend for downstream analysis.
+
+In practical terms:
+- Users continue using current PyTC routes and UI exactly as they do today.
+- A future, explicit integration path can transform local workflow outputs (predictions, masks, metadata, provenance) into CAVE-style entities.
+- Teams that rely on CAVE tooling can consume PyTC results without changing PyTC core execution paths.
+
+Non-goal for this spike:
+- No live auth, network calls, or deployment wiring.
+- No routing changes in existing server endpoints.
+
+---
+
+## Expected inputs and outputs
+
+### Inputs (from current PyTC workflow artifacts)
+1. **Run context**
+   - run ID / job ID
+   - timestamp
+   - model/config reference
+2. **Dataset references**
+   - source image path(s)
+   - optional chunk/volume coordinates
+   - resolution metadata
+3. **Inference / proofreading outputs**
+   - segmentation or mask outputs
+   - object-level summaries/statistics where available
+4. **Provenance and annotations**
+   - operator/user reference (if available)
+   - comments or QC notes (if available)
+
+### Outputs (CAVE-aligned payload expectations)
+1. **Dataset/collection descriptor**
+   - canonical dataset key
+   - spatial metadata
+2. **Versioned result artifact record**
+   - result identifier tied to run context
+   - linkage to upstream input dataset references
+3. **Optional annotation payload(s)**
+   - typed annotation entries linked to coordinates/objects
+4. **Status envelope**
+   - accepted/rejected/queued semantics (for future async transport)
+
+---
+
+## Mapping: PyTC workflow artifacts → CAVE concepts
+
+| PyTC artifact / concept | CAVE-aligned concept | Mapping notes |
+|---|---|---|
+| Training/inference run ID | Version / operation reference | Preserve immutable run identifier for lineage. |
+| Input image volume path + scale | Dataset / image source descriptor | Include voxel resolution and bounds when available. |
+| Output mask / segmentation file | Segmentation payload reference | Publish as versioned result; avoid overwriting by default. |
+| Detected objects / stats | Table/annotation-like records | Map each object to stable ID and geometric summary. |
+| Proofreading edits | Annotation update set | Track editor and timestamp in provenance fields. |
+| Config YAML path + hash | Metadata / provenance attachment | Enables reproducibility checks in CAVE consumers. |
+| Runtime status/result message | Ingest status envelope | Normalize to queued/succeeded/failed taxonomy. |
+
+---
+
+## Suggested integration contract (future)
+Introduce an **explicit opt-in adapter path** that is only invoked by future routes or background jobs, for example:
+- `prepare_cave_payload(run_artifacts) -> CavePayload`
+- `validate_cave_payload(payload) -> ValidationResult`
+- `publish_to_cave(payload, transport) -> PublishResult`
+
+Design constraints:
+- Adapter must be import-safe and side-effect free.
+- No network or auth dependency at import time.
+- Existing endpoints continue untouched unless a new flag/route explicitly calls the adapter.
+
+---
+
+## Non-invasive rollout plan
+1. **Phase 0 (this spike):** document contract and mappings (this file).
+2. **Phase 1:** add inert adapter scaffolding with interface stubs only.
+3. **Phase 2:** add offline payload generation and validation tests.
+4. **Phase 3:** add optional transport implementation behind explicit feature flag.
+5. **Phase 4:** wire opt-in route or async task that calls adapter.
+
+---
+
+## TODOs (auth/network/deployment assumptions)
+
+### Auth TODOs
+- Define service-to-service auth model (token issuer, rotation policy, scopes).
+- Decide whether end-user identity is propagated or service account is used.
+- Specify failure behavior for auth expiration (retry vs fail-fast).
+
+### Network TODOs
+- Define allowed egress targets and timeout/retry policy.
+- Decide sync vs async delivery semantics for large artifacts.
+- Define payload size limits and chunking/compression behavior.
+
+### Deployment TODOs
+- Define environment variables/secrets required for CAVE endpoints.
+- Choose deployment topology (in-process transport vs worker queue).
+- Add observability requirements (structured logs, metrics, tracing IDs).
+
+---
+
+## Actionable next steps
+1. Create `server_api/workflow/cave_adapter.py` with pure interface stubs and typed payload models.
+2. Add `tests/test_cave_adapter_contract.py` to lock down:
+   - no side effects on import,
+   - deterministic payload mapping from fixture artifacts,
+   - explicit invocation requirement (not called in default app path).
+3. Add a feature flag proposal (`PYTC_ENABLE_CAVE_ADAPTER=false` default) for a future opt-in route.
+4. Review mapping table with CAVE stakeholders and finalize required fields.
+
+This keeps current behavior unchanged while preparing a low-risk implementation path.