Roadmap-of-record for the founder performance packet (2026-08-31), with the 80/20 tranche executed and measured.
Executed this tranche (receipts committed)
- P0.1 quota fix: done before the packet landed - ninfer#27 merged, CI green, production quota sized (24576 MiB).
- Fanout probe (
scripts/fleet_probe.py, docs/measurements/2026-08-31-fanout-probe-main.json): on released v0.4.0 at a 67.7K base - cold prefill 25.4 s; continue-from-latest 0.57 s (settled) / 14.6 s (immediately after a turn - post-turn capture stall, datapoint on ninfer#23); branch-from-older-id always root, ~40 s ; restore-from-disk 16 s wall then 0.57 s ttft.
- Anthropic cache_control fanout probe: zero shared-prefix hits as shipped - ninfer#29.
- Text-only arm (
docs/measurements/2026-08-31-text-only-arm-5090.json): frees 0.92 GiB on the 5090 but the KV ceiling is already maxed; 4090 lane already ships vision-off. No production churn; fold into the next profile train.
- Client budgets right-sized (fleet models config): main 16K out, scout 4K out.
Dogfood pattern decided from the data
Same-lane fanout does not reuse on the released engine, and all lanes are C1 (serialized) anyway. Therefore: distribute roles across lanes (each lane pays its base once, then stays warm at ~0.5-1 s), continue sequentially within a lane, use checkpoint restore for recovery/restart (16 s at 5 GB). The packet's four-subagents-one-base criterion becomes achievable same-lane only after the engine work below.
Deferred to the v0.4.x engine train (with the packet as spec)
Standing guards adopted from the packet
No NAS import until ninfer#21/#24 close; live checkpoints stay on local NVMe; secret-bearing sessions never replicate to unencrypted shares; catalogs stay modest until #121-class regression lands; no cross-fingerprint restores.
Packet source: founder-provided omp-ninfer-v0.4-performance-roadmap.md (private).
Roadmap-of-record for the founder performance packet (2026-08-31), with the 80/20 tranche executed and measured.
Executed this tranche (receipts committed)
scripts/fleet_probe.py,docs/measurements/2026-08-31-fanout-probe-main.json): on released v0.4.0 at a 67.7K base - cold prefill 25.4 s; continue-from-latest 0.57 s (settled) / 14.6 s (immediately after a turn - post-turn capture stall, datapoint on ninfer#23); branch-from-older-id always root, ~40 s ; restore-from-disk 16 s wall then 0.57 s ttft.docs/measurements/2026-08-31-text-only-arm-5090.json): frees 0.92 GiB on the 5090 but the KV ceiling is already maxed; 4090 lane already ships vision-off. No production churn; fold into the next profile train.Dogfood pattern decided from the data
Same-lane fanout does not reuse on the released engine, and all lanes are C1 (serialized) anyway. Therefore: distribute roles across lanes (each lane pays its base once, then stays warm at ~0.5-1 s), continue sequentially within a lane, use checkpoint restore for recovery/restart (16 s at 5 GB). The packet's four-subagents-one-base criterion becomes achievable same-lane only after the engine work below.
Deferred to the v0.4.x engine train (with the packet as spec)
Standing guards adopted from the packet
No NAS import until ninfer#21/#24 close; live checkpoints stay on local NVMe; secret-bearing sessions never replicate to unencrypted shares; catalogs stay modest until #121-class regression lands; no cross-fingerprint restores.
Packet source: founder-provided
omp-ninfer-v0.4-performance-roadmap.md(private).