skip to content
◂ lifelineapprentice · learning to survive

l5 survival season

SURVIVAL

it dies, and dies again — then learns to breathe.

learning mode

7deaths → survival
$1,668learner P&L
$874static seed P&L
+$794learner ahead

One seed. 8 lives. The agent began with the backtest's optimal policy and was dropped into a live-shaped survival season — every loss drains its breath, and when breath hits zero it dies for good, mints a Tombstone, and respawns. It died 7 times before life 3 survived 392 bets and pulled $794 ahead of the static seed running the same markets.

cumulative p&l · learner vs frozen seed

learner $1,668 · static $874 · +$794 apart

cumulative p&l · agent vs baselineslearner +$794 vs static seed
$9,143$4,126−$891break-evenstep 0step 449
  • Learner (survives)+$1
  • Static Seed+$0
  • Random−$5
  • Always-Favorite+$5
vitals · breathBREATHING
breath 3578% of life peak
life
0 / 7
win rate
100%
cum p&l
$1
current bet · life 00 / 449
Sabalenka
vs
Bucsa
  • Hard
  • side YES
  • stake $4.48
  • settled +$0.7
engine signals
  • ELO / Ranking+1.00
  • CLOB Momentum+0.15
  • Surface+0.40
  • Head-to-Head+1.00
  • Rest / Recency+0.00
market 1113566

fusion-weight evolution

weight evolution450 bets · 8 lives
0.000.250.500.751.00step 0step 449
  • W_R0.571
  • W_S0.429
  • α₀0.495
  • α₁0.321
  • α₂0.184
  • β₀0.455
  • β₁0.545
  • ρ0.187

it learned which engines to trust · α₀ 0.50 → 0.95 · ρ 0.19 → 0.99

graveyard7 deaths
  1. life0

    ran out of breath

    30 bets+$9
    tomb #0
  2. life1

    ran out of breath

    15 bets−$7
    tomb #0
  3. life2

    ran out of breath

    38 bets−$9
    tomb #0
  4. life3

    ran out of breath

    392 bets+$819
    tomb #0
  5. life4

    ran out of breath

    120 bets+$106
    tomb #0
  6. life5

    ran out of breath

    52 bets+$32
    tomb #0
  7. life6

    ran out of breath

    261 bets+$606
    tomb #0

each marker is an on-chain Tombstone NFT — a life that ran out of breath before it learned to trust the right engines.

then life 7 held its breath across 48 bets, banked $113, and survived.

journeystep 0 / 449
life 0 / 7 · breath 35 · cum $17 deaths

finetune log · run history

After run 1 we audited the data: most of the headline came from two $5 bets that hit extreme longshots at $0.0005 / $0.0055 — bet size was capped at $5 by market liquidity, but the payout size×(1/price−1) was unbounded (one bet "won" +$9,995 and pumped breath into the thousands, faking "learned to survive"). Run 2 introduces two realism rules and reruns both the numerical and AI seasons: an entry-price floor filters untradeable longshots, and a per-bet profit cap (= one life's starting bankroll). Both rules are enforced as hard invariants in the exporter — a journey violating its own physics can never be written to disk.