Skip to content

CMC_REPLAY.md

Latest commit

 

History

History
185 lines (128 loc) · 3.07 KB

CMC_REPLAY.md

File metadata and controls

185 lines (128 loc) · 3.07 KB
Raw

CMC Replay Surface

Status: early replay and admissibility sketch.

This document defines the first replay-oriented surface for the Causal Memory Controller (CMC).

The goal is not full deterministic replay yet. The goal is to establish a machine-readable causal trace boundary.

Core idea

CMC does not only emit logs.

It emits:

causal transition evidence

The replay surface allows a reviewer or future runtime to reconstruct:

what transition was attempted
why it was accepted or rejected
which cause authorized it
whether the effect was committed

Replay pipeline

operation
 -> decision
 -> TraceEvent
 -> JSONL export
 -> replay parser
 -> deterministic reconstruction

Current replay format

CMC currently exports replay-oriented traces as JSONL.

Example:

{"seq":1,"kind":"WRITE","decision":"REJECT_MISSING_CAUSE","address":4096,"effect_id":null,"cause_id":null,"message":"memory write requires an explicit cause"}

Each line represents a single causal decision boundary.

Why JSONL

JSONL is intentionally simple:

  • append-friendly
  • stream-friendly
  • replay-friendly
  • diff-friendly
  • deterministic ordering
  • easy to feed into external tooling

Future layers may add:

  • hash chains
  • signatures
  • compressed binary formats
  • T-Trace compatibility
  • LTP admissibility metadata

Minimal replay guarantees

Current replay guarantees:

same operation sequence
 -> same decisions
 -> same trace ordering

This is enough for:

  • CI verification
  • fixture comparison
  • blocked-transition proofs
  • causal audit demos
  • reproducible research artifacts

Relation to LTP and T-Trace

CMC TraceEvent
 -> local causal decision

T-Trace
 -> integrity + verification layer

LTP
 -> replay/admissibility transport layer

CMC replay is intentionally minimal. It focuses on causal decision boundaries near memory/effect operations.

Admissibility direction

Future replay layers may support:

proof that an effect
was causally legitimate
at execution time

Potential future checks:

  • missing cause detection
  • effect-before-commit detection
  • replay divergence detection
  • forbidden transition verification
  • trace completeness verification

Example replay questions

A future reviewer should be able to ask:

Which transition failed?
Which cause authorized it?
Was the cause committed?
Did the effect execute anyway?
Did replay diverge from the original trace?

Non-claims

CMC replay currently does NOT provide:

  • cryptographic integrity
  • Byzantine guarantees
  • distributed consensus
  • complete deterministic execution replay
  • production audit admissibility
  • hardware trace persistence

It only establishes the first machine-readable causal replay surface.

Long-term direction

software trace
 -> replay surface
 -> admissibility checks
 -> embedded trace stream
 -> FPGA TraceOut
 -> hardware causal replay

The long-term thesis is:

agentic systems may eventually require
not only replayable execution,
but replayable legitimacy.