Skip to content

fleet-simulation-notes.md

Latest commit

 

History

History
99 lines (82 loc) · 5.73 KB

fleet-simulation-notes.md

File metadata and controls

99 lines (82 loc) · 5.73 KB
Raw

Fleet Simulation Notes — the elephant at fleet scale (a room of rooms)

File: examples/fleet_simulation.py Companion design: docs/fleet-dynamics-design.md Tests: tests/test_fleet_simulation.py

What this demonstrates

The elephant is a room-temperature sense: standalone in a harness (one boat, BoatHarness), or plugged into any system with sensors. The fleet simulation is the proof that the same mechanism scales from a room (one boat) to a room of rooms (the fleet) without changing its nature.

Four fishing boats each run an elephant. Every boat perceives its own rooms (radar = the distribution of the other boats, sounder = the biomass under the keel, nav = course, conversation = the crew). The fleet is the meta-room: boats broadcast numbers only (position, velocity, radar_coherence, fishing_day binned to {-1, 0, +1}), and the meta-room's field is computed over those numbers — a fleet concentration κ over boat positions, plus a warmth over the shared dials. Feeds stay home; only the distilled readings cross the wire.

The 30-day arc is the demonstration: a warm room forms (good fishing), it dissolves (spotty fishing), the fleet feels the deviation from the warm room it acclimated to, and then it feels the warm room return — all through contrast, with catch as the only exogenous label.

The numbers it produces

Per day, the sim reports:

Quantity Meaning Range
kappa fleet concentration — 1 / (1 + weighted mean radial distance from the weighted-median centroid). Clustered on the drag → near 1; scattered searching → near 0. [0, 1]
effective_kappa the damping bell (§4.2): κ·(1 − clamp(κ, 0.2, 0.8)) — kills the herd-panic feedback loop at both ends. [0, ~0.25]
warmth meta-room temperature — 0.5·fishing_day + 0.3·(2·biomass − 1) + 0.2·radar_coherence. [−1, +1]
spread_km the fleet's spatial spread (mean distance to centroid). ≥ 0 km
catch exogenous catch telemetry — the ground truth the anchor is built from (never derived from the dials). [0, 1]
deviation Mahalanobis distance of today's field from the good-week anchor — "does this stretch feel like the good kind?" ≥ 0
fishing_day_per_boat each boat's composite luck dial (local, never leaves the boat). [−1, +1]
fishing_day_binned the signed bin that does cross the wire. {−1, 0, +1}
nudge_mean the fleet-mean nudge prior over 7 modalities (local cognition; only its number is reported). [−1, +1]⁷

The 30-day arc, phase by phase

  • Days 1–7 (good): boats bunch on a slowly-drifting drag → kappa climbs from ~0.35 to ~0.85, warmth rises, catch high. Seven catch-good days become the anchor (biomass_anchor → a Gaussian over [κ, biomass, catch]).
  • Days 8–14 (spotty): boats scatter searching → kappa falls toward ~0.1, warmth goes negative, catch collapses. deviation balloons — the elephant feels "this is not the warm room" without ever being told fishing is bad.
  • Days 15–30 (recovery): boats re-group → kappa climbs back, warmth recovers, and deviation drops back toward the good-week anchor. Day 15+ is the payoff: the elephant recognizes the warm room returning from the shape of the field, before it can be explained.
  • Day 20 (dark-boat charisma, optional): the highest-reputation boat goes dark. The meta-room injects a 3×-weighted virtual point at its last position, so the fleet holds attention toward the hole instead of reading it as thinning — the design's "hot boat went quiet because it's on fish, not because it left" heuristic.

What it proves about the elephant at fleet scale

  1. The math scales. The same vMF/field idea that reads one room reads the fleet. Fleet κ is the room temperature one level up: high κ = boats agreeing on where the fish are (warm); low κ = disagreement (uncertain). No new machinery.
  2. Numbers are enough. The whole fleet runs on broadcast scalars. The meta-room never sees a feed; it feels the distribution. Privacy and bandwidth fall out by construction.
  3. The anchor must have an outside. The review caught the tautology — an anchor built from fishing_day (itself a composite of the dials) would be self-confirming. The sim uses exogenous catch as the anchor's ground truth, so "good day" is defined by what was landed, not by what the dials felt.
  4. Deviation is felt, not reported. The elephant is never handed a "fishing is bad" label. It feels the Mahalanobis gap between today's field and the warm room, and the gap is the whole training signal.
  5. The feedback loop is damped. effective_kappa (the damping bell) shows the hold/scatter loop can't run away — clustered fleets get a weak probe nudge at the top, scattered fleets get a weak hold at the bottom, and the real signal lives in the un-touched middle.

Review pass (Seed-2.0-pro)

Three fixes were adopted from a Seed-2.0-pro critique before implementation:

  • Exogenous catch in the anchor (was fishing_day) — breaks the biomass↔fishing-day collinearity that would have made the Mahalanobis deviation measure shared noise instead of fleet drift.
  • Smooth κ — 1/(1 + weighted mean spread) uses every boat and varies continuously, rather than a median-of-4 MAD that would quantize the concentration into a handful of steps at fleet size.
  • Flagged dark boat — the virtual point is reported explicitly (it is the charisma rule working, not a hidden metric distortion).

Running it

cd /home/eileen/projects/elephant
python3 examples/fleet_simulation.py
python3 tests/test_fleet_simulation.py