Appearance
Deterministic Execution Model
Every workload in Forge Pool is defined by a canonical execution contract.
If the contract is identical, the execution is reproducible. This enables replay, audit, and verification.
Canonical Execution Contract
At minimum:
json
{
"ctx": {
"job_id": "client-supplied-or-empty",
"trace_id": "optional",
"billing": { "mode": "test|billable" }
},
"op": { "name": "mc", "version": 1, "profile": "insurance.v1" },
"seed": { "mode": "explicit|derived", "value": "optional" },
"policy": { "target": "cpu|gpu", "min_agents": 1, "max_agents": 999, "verify": "none|spotcheck|quorum" },
"args": { "iterations": 1000000 }
}args are part of the contract — without them, the computation is not defined.
Execution Lifecycle
All workloads follow the same lifecycle:
contract
→ plan
→ shard
→ execute
→ verify
→ aggregate
→ artifact + replay
→ ledgerPlanning and Sharding
The Hub deterministically plans:
- shard count
- iteration bounds (or split strategy)
- per-shard seed derivation
- verification strategy (policy-dependent)
Shards are the unit of execution.
Execution
Agents execute shards in isolated environments and return:
- result fragments
- metrics (wall time, resource class)
- integrity hashes (when enabled)
Verification
Verification is policy-controlled and workload-dependent:
- redundant execution for a subset or quorum
- statistical consistency checks (probabilistic workloads)
- deterministic hash checks (pure numeric kernels)
Invalid shards are rejected or excluded from aggregation.
Aggregation
The Hub reduces shard outputs deterministically:
- stable aggregation order
- bounded numeric tolerances
- workload-specific reducers (percentiles, HDR merges, etc.)
Aggregation produces:
- final output
- result hash
- execution metadata
Replay
Replay is possible when:
- contract is preserved (including
args) - adapter/kernel version is pinned
- seed semantics are stable
Replay produces the same output within defined numeric tolerances.
Evidence Discipline
Each execution produces immutable evidence:
- job ID and trace references
- shard plan and assignments
- verification outcomes
- aggregation hash
- billing/ledger record (mode-dependent)
