38 lines
2.7 KiB
Markdown
38 lines
2.7 KiB
Markdown
# Architecture rollout — 2026-09-09
|
|
|
|
Authority: user approved implementation; architecture and plan-changing surprises are owned by
|
|
`openai/gpt-6-astra` (the active lead). The user's correction supersedes SOL/MYTHOS routing.
|
|
|
|
## Decisions
|
|
|
|
- `sots-re` is the only campaign state authority. Contracts, checkpoints, decisions, evidence,
|
|
leases and accepted results live here. Chat memory and engine docs are not campaign state.
|
|
- Primary lead and resolver: GPT-6 Astra. Analysts, implementers and independent verifiers:
|
|
GPT-5.6 Terra. Lab operator: GPT-5.5. Compaction: GPT-5.5. No silent model fallback.
|
|
- One pilot, then at most two independent implementation slices. No worker subdelegation.
|
|
- Workers checkpoint every 20 tool calls or 15 minutes, whichever comes first, and before
|
|
compaction, handoff, experiments or stopping. Bounded sessions resume from repository state.
|
|
- New slices use paired worktrees. This bootstrap alone uses exclusive FILE ownership in the
|
|
clean canonical trees: the lead owns architecture/state/config/dashboard; gate-worker owns
|
|
gate/evidence/reporter tooling; engine-worker owns accounting and engine tests. No shared-file
|
|
writes, staging, commits, pushes, remote deployment or Ghidra mutations are authorized.
|
|
- Replace obsolete live workflow documents and agents; history is in Git. Retain scientific
|
|
evidence and domain findings. Historical observations do not become current acceptance.
|
|
- Local gate builds a fresh source snapshot, hashes actual source and inputs, runs CTest once,
|
|
enforces expected identities and required execution, and records limitations. Full acceptance
|
|
additionally requires assets, shim build and slice-specific oracle criteria. No remote shared
|
|
staging overlay; operators run the same gate on a leased lab host with explicit paths.
|
|
- Reporter requires explicit binary and provenance, errors on missing/failed pairs or coverage,
|
|
separates byte/inflated/state equality, and supports measurement versus acceptance explicitly.
|
|
|
|
## Bootstrap task ownership
|
|
|
|
| Worker | Model | Files | State |
|
|
|---|---|---|---|
|
|
| lead | openai/gpt-6-astra | architecture, AGENTS, configuration, campaign tool/state, dashboard | implementing |
|
|
| gate-worker | openai/gpt-5.6-terra | tools/{gate.sh,gate.py,evidence.py,standalone_report.py}, verify/tooling/ | assigned |
|
|
| engine-worker | openai/gpt-5.6-terra | src/app accounting, RNG support as necessary, tests/, CMake test properties | assigned |
|
|
|
|
Surprises must be recorded in the worker checkpoint with evidence and the proposed discriminating
|
|
check, then escalated to the lead before changing architecture or scope. Ordinary implementation
|
|
errors within a contract may be fixed without escalation.
|