RunSpec 2.0 off-policy runtime
Accepted rollouts are durable before acknowledgement; only a checkpointed contiguous WAL frontier may be
pruned.
Resolve once
Actor · local process or
remote game host
Authenticated
durable ingest
Learner process
parse
record
action
rollout
submit
persist, then ACK
apply in order
capture; prune
after save
activate next
episode
RunSpec 2.0
strict YAML · explicit
components
Resolve contracts
seed · imports ·
fingerprint
Immutable manifest
map · geometry ·
feature/action IDs
Frozen actor policy
one version per
episode
Game adapter
telemetry → features
→ transition
Durable actor spool
retry until accepted
Loopback gRPC
token · size ·
fingerprint
SQLite WAL
accepted rows
ordered by ID
Applied frontier
contiguous replay
commit
Replay + sampler
uniform · PER ·
sequence
Typed batch
n-step · masks ·
IDs
Learner update
validate →
backward → step
Checkpoint
state · RNG · WAL
frontier
Policy snapshot
portable tensor
state
Run evidence
JSONL · W&B ·
artifacts
Invariant: acknowledged chunks remain recoverable until the exact applied frontier is inside a
durable checkpoint.
PPO is a separate local on-policy lifecycle; it never enters this actor/WAL/replay path.
Download editable diagram