strawberry/deploy/validation
trvon f4adc12a74
gate: judge the model against a calibrated rarity comparator
The rarity-greedy comparator acted on every evaluated window, so it spent
each hour's budget on the first ~15 windows and scored 0 hits in every
cell. Criteria 2 and 3 of the Stage 4 gate were void against it.

rarity_calibrated keeps the same rarity value, chooser, costs and budget,
and adds a per-node threshold frozen before the evaluated span: the m-th
largest value over the trailing 24 h of label-free history between the
warm-up and the --eval-after bound, with m set so expected spend matches
the hourly budget. Without that history the comparator is uncalibrated
and the gate fails by name; it never falls back to another rule.

rarity_greedy (the old first-come rule) and rarity_hindsight (per-hour
ordering, not causal) are reported as diagnostics and never judged.
fleet gate pools counts, keeps thresholds per node, and fails a pooled
cell when any node is uncalibrated or lacks the comparator.

Contract change: in `fleet gate --json`, rarity_hits and
rarity_cpu_seconds now describe the calibrated comparator; the old
numbers are rarity_greedy_hits and rarity_greedy_cpu_seconds.

Pre-registration: docs/evaluation/stage4-rarity-baseline-preregistration.md,
sha256 4916c95cd2581e3749acfe46c3ccf827aa5ce98c06150b49b978fc70da013f95,
indexed in YAMS at 2026-11-20T18:55:48Z, before the code. This commit is
dated after the registration time in that document (2026-11-21T00:00Z);
the YAMS record is what predates it.

Checked: meson suite 58/58, scripts/fleet_e2e.sh, clang-tidy and the
strict analyzer on the changed TUs. Not run: fold on the changed TUs,
make fuzz, the live Docker tests and the perf KPI. The comparator has
not been run on validation-range data.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
2026-11-20 19:20:55 -07:00
..
range gate: judge the model against a calibrated rarity comparator 2026-11-20 19:20:55 -07:00
README.md Stage 6: local multi-node fleet-loop validation (gap #4) 2026-11-08 09:32:32 -07:00
run-fleet-loop.sh Stage 6: local multi-node fleet-loop validation (gap #4) 2026-11-08 09:32:32 -07:00

Strawberry distributed-learning validation

run-fleet-loop.sh validates the full hub-and-spoke loop end to end with the real binaries, isolated per-node dirs/configs, and file-based transport — no mocks.

What it proves

seed N nodes (store + rarity.state)
  -> round-1 export         : bundles carry rarity sketches, NO fleet feature
  -> central fleet train    : merges N distinct nodes, trains the shadow model
  -> central fleet overlay  : sums per-node rarity sketches -> non-empty overlay
  -> distribute overlay
  -> round-2 export          : feature vectors now carry the fleet-wide rarity

Hard assertions (exit non-zero on any failure):

  • each round-1 bundle has rarity sketches (rarity>0) and fleet_feat=0;
  • central sees exactly N distinct nodes (no node_uid collision);
  • the overlay is non-empty (sketches merged);
  • the round-2 export carries the fleet feature (fleet_feat>0).

A passing run prints: PASS: fleet loop validated (round-1 fleet_feat=0 -> round-2 fleet_feat=<N> ...).

Run it (local, no Docker)

meson compile -C builddir strawberry strawberry-fleet-seed strawberry-fleet-inspect
bash deploy/validation/run-fleet-loop.sh          # NODES=2 EXECS=240 by default

Override binaries/params via env: SB, SEED, INS, NODES, EXECS, WORK.

Tools (build-only, not installed)

  • strawberry-fleet-seed <store> --node <n> --execs <N> [--seed S] [--model-dir <d>] — writes a deterministic synthetic telemetry store (and, with --model-dir, a rarity.state, simulating a node that has been running the daemon).
  • strawberry-fleet-inspect <bundle> — prints node_uid windows features labels rarity fleet_feat for an export bundle (fleet_feat = feature vectors carrying a non-zero fleet slot).

Scope / honesty

This validates the distributed DATA + LEARNING path (export → merge → train → overlay → distribute → re-export) across isolated nodes. It uses a seeder instead of the live eBPF daemon, so it does NOT re-validate live telemetry capture (the single-node daemon path covers that). The Docker-swarm harness (when built) runs this same script body inside containers for host-isolation realism; the logic proven here is identical.

Integration facts this surfaced (true of real deployments)

  • Nodes must export with a distinct node_id (config) or they collide to one node_uid on a shared identity; in containers, distinct machine-ids also suffice.
  • The fleet overlay only forms once nodes have a rarity.state (accumulated by the running daemon) — a cold node's first export carries no rarity sketches, so fleet_feat stays 0 until a later round.
  • fleet train's max_node_share cap is for multi-node fleets; a single node is cap-degenerate.