Purpose
This is the focused discussion issue for RFC-0002: Causal Fabric Evidence Protocol (CFEP).
CFEP is classified as Direction 3 — research bet: no current sprint dependency, no semver commitment. The RFC is preserved because the AEP signing contract and causal chain primitives already form the exact foundation CFEP needs; the incremental investment to extend downward is smaller than building from scratch.
The one decision that matters right now
Should anyone commit 1–2 weeks of part-time engineering to Phase 0 (toy simulator)?
Phase 0 scope is narrow: a pure-software, abstract-event-level demo of the Checkpoint → Record → Seal → Reconstruct lifecycle plus a Wasm overhead benchmark. No real CXL registers, no hardware, no new simulator written from scratch (builds on CXL-DMSim or QEMU CXL).
The honest success criterion: (a) the lifecycle runs end-to-end; (b) the Wasm overhead question is answered with a number; (c) the biggest Phase 1 risk is named.
Four open questions from the RFC
These are worth discussing before Phase 0 starts — not all need answers upfront, but they should be on the table:
1. Wasm overhead ceiling
At what simulated transaction frequency does observe-transaction become infeasible without AOT or batching? Is a coalescing strategy (batch N transactions, one policy check) safe?
2. Causal graph granularity across layers
When a FabricActivityRecord and an AEP ActionEvidence share a causal_chain_id, is that the right entity boundary? A CXL page migration and an LLM tool call are very different abstraction levels. Should CFEP use a separate fabric_chain_id that AEP references by foreign key instead?
3. Simulator choice: CXL-DMSim vs QEMU CXL vs both via SimBricks
CXL-DMSim has epoch concepts that map naturally to CFEP epochs. QEMU CXL 2.0 has device emulation. Do they compose via SimBricks for Phase 0, or do we pick one and accept its fidelity limits?
4. Re-evaluation triggers
The RFC lists four triggers for moving CFEP up in priority (industry partner co-investment, Direction 1/2 validating layered evidence in production, Wasmtime sub-microsecond AOT, CXL-DMSim/SimBricks maturity). Are these the right triggers? Anything missing?
What this discussion is NOT for
- AEP v0.3 schema questions → wasmagent-js#7
- erp-agent architecture → .github#94 / RFC-0001
Purpose
This is the focused discussion issue for RFC-0002: Causal Fabric Evidence Protocol (CFEP).
CFEP is classified as Direction 3 — research bet: no current sprint dependency, no semver commitment. The RFC is preserved because the AEP signing contract and causal chain primitives already form the exact foundation CFEP needs; the incremental investment to extend downward is smaller than building from scratch.
The one decision that matters right now
Should anyone commit 1–2 weeks of part-time engineering to Phase 0 (toy simulator)?
Phase 0 scope is narrow: a pure-software, abstract-event-level demo of the Checkpoint → Record → Seal → Reconstruct lifecycle plus a Wasm overhead benchmark. No real CXL registers, no hardware, no new simulator written from scratch (builds on CXL-DMSim or QEMU CXL).
The honest success criterion: (a) the lifecycle runs end-to-end; (b) the Wasm overhead question is answered with a number; (c) the biggest Phase 1 risk is named.
Four open questions from the RFC
These are worth discussing before Phase 0 starts — not all need answers upfront, but they should be on the table:
1. Wasm overhead ceiling
At what simulated transaction frequency does
observe-transactionbecome infeasible without AOT or batching? Is a coalescing strategy (batch N transactions, one policy check) safe?2. Causal graph granularity across layers
When a
FabricActivityRecordand an AEPActionEvidenceshare acausal_chain_id, is that the right entity boundary? A CXL page migration and an LLM tool call are very different abstraction levels. Should CFEP use a separatefabric_chain_idthat AEP references by foreign key instead?3. Simulator choice: CXL-DMSim vs QEMU CXL vs both via SimBricks
CXL-DMSim has epoch concepts that map naturally to CFEP epochs. QEMU CXL 2.0 has device emulation. Do they compose via SimBricks for Phase 0, or do we pick one and accept its fidelity limits?
4. Re-evaluation triggers
The RFC lists four triggers for moving CFEP up in priority (industry partner co-investment, Direction 1/2 validating layered evidence in production, Wasmtime sub-microsecond AOT, CXL-DMSim/SimBricks maturity). Are these the right triggers? Anything missing?
What this discussion is NOT for