Skip to content

fleet-operations.md

Latest commit

 

History

History
262 lines (212 loc) · 12.9 KB

fleet-operations.md

File metadata and controls

262 lines (212 loc) · 12.9 KB
Raw

Fleet Operations — running and reading the elephant at scale

This is the interpretation side of the elephant: how to run the fleet simulation, how to read the numbers it produces, and how to run the Tap practice loop that tunes the whole thing. Every number quoted here is from the shipped examples, run as-is (seed 7 for the fleet, seed 42 for the Tap).

Companion reading: docs/fleet-simulation-notes.md (the sim), docs/ deployment-guide.md (how to wire it), docs/tap-night-operations.md (the Tap runbook).

1. Running the fleet simulation

cd /home/eileen/projects/elephant
python3 examples/fleet_simulation.py      # 4 boats, 30 days, seed 7
python3 tests/test_fleet_simulation.py    # the assertions over the arc

The sim drives four boats (EILEEN, PETREL, FULMAR, SHEARWATER, reputations 1.7 / 1.0 / 0.9 / 0.6) through a 30-day fishing arc, one BoatHarness per boat, and computes the fleet field over their broadcast scalars. The anchor feature is [kappa, mean_biomass, catch] where catch is exogenous (drawn from the phase schedule, never a dial).

What each phase means

Days 1–7 — GOOD. Boats bunch on a slowly-tightening drag around the ground point (12, 8) km. The daily line reads, e.g. day 1 κ 0.44, day 7 κ 0.88. The meta-room warms, the sounder runs thick, catch is high. Seven catch-good days become the anchor (a Gaussian over the [κ, biomass, catch] features).

Days 8–14 — SPOTTY. Each boat drifts outward from its day-7 position along its home bearing, searching at 6–8.5 kts over thin marks. κ collapses (day 8 κ 0.26 → day 14 κ 0.05), warmth goes deeply negative, catch dies. deviation balloons — the elephant feels "this is not the warm room" without ever being told fishing is bad.

Days 15–30 — RECOVERY. Boats blend from their day-14 positions back to a tight home formation. κ climbs (day 18 κ 0.77 → day 30 κ 0.93), warmth recovers, and deviation drops back toward the anchor. This is the payoff: the elephant recognizes the warm room returning from the shape of the field, before it can be explained.

How to read the per-phase numbers

The seeded run reports these phase means:

Quantity Good (1–7) Spotty (8–14) Recovery (15–30)
fleet κ (raw / effective) 0.63 / 0.22 0.11 / 0.08 0.76 / 0.17
meta-room warmth +0.66 −0.83 +0.41
exogenous catch 0.81 0.12 0.71
per-boat fishing_day +0.64…+0.69 −0.82…−0.88 +0.40…+0.46
inductive deviation (σ) 1.35 12.17 3.50

How to read each column:

  • κ — fleet tightness, 1/(1+spread_km). Near 1 = boats agreeing on where the fish are (warm); near 0 = disagreement (uncertain). The effective κ applies the damping bell (κ·(1−clip(κ,0.2,0.8))) — it is always smaller and is the number the driving loop actually uses.
  • warmth — the meta-room temperature, `0.5·fishing_day + 0.3·(2·biomass−1)
    • 0.2·radar_coherence`. A warm fleet runs positive; the spotty week dives to −0.83 (the cold plunge).
  • fishing_day — each boat's local composite luck dial. It never leaves the boat as a float; only its bin {-1, 0, +1} crosses the wire.
  • deviation — the Mahalanobis distance of today's field from the good-week anchor, in "good-day sigmas." 1.35 (good) → 12.17 (spotty) → 3.50 (recovery). The report labels a phase "feels like the good kind" when deviation < 0.5 × spotty_deviation.

The 30-day arc in one breath: κ 0.79 (days 1–7) → 0.06 (days 12–14) → 0.94 (days 24–30), deviation spotty 12.2 vs recovery 3.5, anchor mean [0.632, 0.819, 0.813] from 7 good days.

The per-day nudge prior (7 modalities) tells the same story: the good week averages radar=+0.69, sounder=+0.82, conversation=+1.00; the spotty week flips to radar=−0.90, conversation=+0.30. The elephant stops telling the vision model to look at the radar cluster, and starts saying there is nothing there worth comparing.

What the dark-boat event demonstrates

On day 20, EILEEN (reputation 1.7, the best skipper) stops broadcasting at (12.3, 8.3) km. The fleet field then uses the 3 active boats' positions + one virtual point at her last position, weighted 3 × 1.7 = 5.1. The remaining boats' recovery target becomes that last position — they re-group on her mark, not the old ground point.

What it proves: charisma is field displacement. A strong presence that goes quiet should still pull the room toward where it was — not read as a thinning of the meta-room. The virtual point is reported explicitly (it is the charisma rule working, not a hidden metric distortion). Set dark_boat_event=False and the rule never fires; all four boats broadcast all 30 days.

How to extend the sim

  • More boats — FleetSimulation(n_boats=6, seed=7); home geometry generalizes (_home_dir places boat i at angle 2πi/n), and boat names / reputations wrap by index modulo.
  • Real tracks — the clean seam is _simulate_day: replace the phase-scheduled pos/heading/speed/biomass draws with your own AIS / log data, keep the harness feeding (3 sensor reads at 0h/8h/16h + one conversation line/day), and keep catch exogenous. The fleet field, anchor, and deviation machinery are unchanged.
  • A different anchor window — run(days=30) fits the anchor from features[:min(7, len(features))]. Change that slice to match your real good-week horizon; the covariance uses OAS shrinkage (elephant/fleetmath.py::_oas_shrinkage) so it stays well-conditioned for N ≲ d (a week of good days).

2. Interpreting the elephant in production

A warm room vs a cold one

The field is the ensemble of all seven dials. In the single-boat demo (examples/fleet_harness_demo.py), the good hour reads RoomField(warmth=+0.50, κ=2.09) with mood=1.0, joke_landing=1.0, cynicism=0.0, fishing_day=+0.83; the spotty hour reads RoomField(warmth=−0.55, κ=3.62) with mood=−1.0, cynicism=1.0, fishing_day=−0.29. The two rooms are the sauna and the cold plunge — the same dials, inverted.

Operationally: a warm room feels mood and joke_landing up, cynicism and panic down, warmth positive; a cold room is the mirror. The demo spaces (examples/demo_spaces.py) give three worked contrasts in one run:

  • warm MUD bar — warmth +0.45, κ 1.91, mood=+1.0, joke_landing=+1.0
  • heated chat thread — warmth −0.62, κ 3.68, mood=−1.0, cynicism=+1.0
  • quiet sensor deck — warmth −0.05, κ 2.02, all seven dials near rest

When to trust κ vs warmth

They measure different things and disagree usefully.

  • κ (concentration()) is tightness — how many ways there are to be in the room. A cold room is one way (high κ, rigid); a warm room is many ways (low κ, loose). Watch κ fall as regulars settle a vibe: at the Tap it drifts from ~1.7 down toward ~1.5 on the good nights.
  • warmth is valence — whether that one-or-many ways is good or bad.

They can diverge: a room can be tight and warm (a crew locked onto a hot drag — high κ, positive warmth) or loose and cold (a scattered fleet searching alone — low κ, negative warmth). Trust warmth for "is this good right now," trust κ for "how rigid/settled is this room." The single-boat good hour is κ 2.09 and warm +0.50 — tight, and good. Do not treat κ alone as health.

How acclimation and charisma show up in the numbers

Both are dynamics, and both are visible:

  • Acclimation (field.py::acclimation_curve) — an agent relaxes toward the room at rate = their acclimation_rate (1/τ). A high-rate agent is indistinguishable from the room within an evening; a low-rate agent keeps their own shape. In the Tap the captain (rate 0.40) warms to the room fastest; the engineer (rate 0.15) stays the dry voice longest.
  • Charisma (field.py::charisma_pull) — a strong presence pulls the room toward them over interactions. Measure it as the gap between room_field() (effective, charisma-displaced) and raw_field() (dial-only): that gap is the sum of the room's strong presences bending the field toward themselves. In the Tap the captain (charisma 0.30) is the one whose dial "pulled the field" on all 14 nights.

The chemistry of agents — why two agents get along one day and not the next

Two agents can read warm together one evening and cold the next with no change in either's settings, because engagement is peer-relative and the room field is what moves. A single evening's field is the sum of everyone's vibe, displaced by charisma and shaped by the works read aloud; that aggregate swings night to night.

The Tap demo shows it in the raw record: night 2 warmth +0.13, κ 1.53 (the room warmed), night 3 warmth −0.04, κ 1.74 (the room cooled), night 8 −0.10, κ 1.95, night 13 −0.14, κ 2.13. The same six people, the same priors — and the room moves. The married-couple off day is not a bug in either agent; it is the field they are jointly generating. If you want stability, look at the inputs (what was read, what reactions landed) and the long-run divergence (below), not at any one night's warmth. Do not hand-tune a participant because one night read cold — the whole design is that the settings are discovered by running cycles, not designed top-down.

3. The practice loop — running Tap nights

The elephant becomes a production tool only through many cycles. Each evening: feed the works → read the field → self-tune once per participant → persist. examples/tapnight_cycles.py runs 14 evenings with 6 personalities (writer, poet, essayist, engineer, critic, captain) and prints the diverged taste table. The loop is the one from docs/tap-night-operations.md §4:

for evening in range(1, N + 1):
    session.start_session()
    for author, text, reactions in load_evening(evening):
        session.speak(author, text, reactions=reactions)
    field = session.room_field()
    for name in session.participants:
        session.tune_participant(name)     # self-fine-tune dial_weights
    save_settings(session.settings())      # persist between evenings

Reading the diverged taste table

After 14 nights the table reads (top dial + weight):

personality top dial (weight) accl charisma
writer mood 0.53, joke_landing 0.44 0.35 0.20
poet volume 0.60, mood 0.29 0.25 0.15
essayist earnestness 0.74 0.30 0.10
engineer cynicism 0.69 0.15 0.25
critic cynicism 0.60, joke_landing 0.35 0.20 0.18
captain mood 0.73, presence 0.21 0.40 0.30

The tell is the mean pairwise distance of dial_weights: initial 0.389 → final 0.859 (tastes diverged). Each personality settles onto a different (or at most lightly shared) dial — the captain's "different guitarists" made visible. Note the interesting overlap: engineer and critic both land on cynicism, but the engineer's top weight (0.69) is sharper than the critic's (0.60), and their second dials differ (earnestness vs joke_landing) — divergence is about the whole 7-vector, not a single argmax.

Folding insights back into tuning

  • Watch the trend, not a night. learning_rate=0.15 is an exponential moving average: each evening moves a dial a little, and structure (who is systematically distinctive on which dial) is what survives the noise. Judge after dozens of evenings, not two or three.
  • If a taste looks "wrong" after many cycles, fix the inputs, not the knobs. The self-tuning is driven by felt_engagement() — peer-relative (vibe − cast_mean_vibe) amplified by reaction heat (😂→joke_landing, ❤️→mood, 🙄→cynicism…). If a participant isn't diverging, the works or the reactions aren't giving them a distinctive signal.
  • Feed reactions in. joke_landing and the self-tuning both lean on reaction heat. Without reactions the system still works (engagement rests on the peer-relative vibe alone), but "did the joke land" becomes real only when the crowd's hands are in the log.
  • Keep regulars registered. Auto-registration (neutral defaults) is for drop-ins; the regulars' settings must persist to accumulate learning.
  • Seed for reproducibility. The demo seeds np.random.default_rng(42); seed your own run to reproduce the table above.
  • Know the v0 caveats. The dials are hand-crafted and keyword-matching: they can saturate (a room of earnest writers reads earnestness ≈ 1.0) and cannot detect sarcasm ("great." reads warm). The peer-relative signal is designed so divergence survives that naivety; v1 trained dials will sharpen it.

Run the demos before trusting any of this — the elephant is only real when you walk into a different room and it is a very different elephant.