Skip to content

Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.

8.7 KB
Raw Blame

Earned autonomy: freedom the fleet has to trade for

The sniper fleet runs several strategies ("arms") side by side against the same market. Two loops improve them without a human in the seat: the optimizer tunes each arm's own knobs, and the evolution loop moves budget between arms.

Both used to treat every arm identically. Same hard bounds, same per-run step, same short list of writable fields, same context in front of the LLM judge. An arm that had been profitable for weeks was allowed to search exactly as narrow a space as one that had never had a winning day, and was told exactly as little before deciding.

Earned autonomy fixes that. An arm's own realized record decides how much rope it gets, and the decision is recomputed from scratch every run.

The tiers

Tier Earned by What it grants
probation net negative with an average per-trade edge at or below -0.5%, over 15+ real trades Narrowed bounds, half-size steps, per-trade size capped at 0.05 SOL, base knowledge
standard the default: too few trades to judge, or no decisive edge Historical bounds and steps, base knowledge
trusted net positive with an average edge of at least +0.5%, over 12+ real trades Wider bounds, 1.75x steps, unlocks the entry universe, the LLM confidence bar and the take-initials ladder, informed knowledge pack
autonomous net positive with an average edge of at least +5%, over 40+ real trades Widest bounds, 2.5x steps, every tunable field, complete knowledge pack

The gates live in api/_lib/sniper-autonomy.js as GATES, and the whole policy is pure: no I/O, no clock, no DB, so the same function decides the tier everywhere it is read and the entire thing is unit-testable (tests).

Why average edge and not win rate

An arm can win 36% of the time and still be the most profitable on the board, because the winners are far larger than the losers. Win rate would demote that arm. The tier is decided on realized net P&L plus average per-trade edge, so the arm that actually makes money is the one that earns room, whatever its hit rate.

The reverse case is just as important: an arm that wins 60% of its trades and is still net negative (small winners, large losers) is not succeeding, and gets probation rather than a reward.

It is symmetric, and it is continuous

A tier is not granted once. Every optimizer run recomputes it against a trailing window, so an arm that stops making money loses its rope on the next pass without anyone intervening, and a demoted arm earns it back the same way. Freedom is rented, never owned.

What a tier changes

1. How far and how fast the optimizer may move a knob

Each tier carries its own TIER_BOUNDS and a step multiplier. Higher tiers widen the search space in both directions and let a single run move further; probation narrows it and halves the step. standard is the historical behaviour exactly, so anything that does not pass a tier is unchanged.

2. Which fields the optimizer may touch at all

Every tier can tune stops, take-profit, hold time, the quality and oracle floors, and position size. Reaching trusted unlocks five more (TIER_UNLOCKS in api/_lib/sniper-autonomy.js):

  • llm_min_confidence: a profitable judge earns a lower bar, so it acts on more of what it sees instead of passing on launches it would have won.
  • min_market_cap_usd / max_market_cap_usd: a wider hunting ground, widened outward only. An arm with no band set is already unrestricted and nothing here will restrict it.
  • initials_out_multiple / moonbag_min_pct, the proactive take-initials ladder: recover the stake as soon as the position hits 2x rather than waiting for an exit to trigger, and size the moon bag that keeps riding. Turned on only for an arm whose winners demonstrably run past its average exit. Note this is only the early recovery: never fully selling a winner is fleet-wide policy that applies at every tier, ladder or no ladder.

autonomous adds max_creator_launches, so a proven arm can test whether a prolific creator is really a bad one.

3. How much of the fleet budget an arm concentrates

The evolution loop multiplies each arm's fitness by its tier weight (0.6 on probation up to 2.2 at autonomous) before splitting the fixed fleet budget. The fleet total never changes and the per-arm floor still keeps exploration alive: this only decides how the same pot is divided.

4. How much the LLM judge is told

This is the "more knowledge" half, and it is built in workers/agent-sniper/judge-knowledge.js.

  • base (probation, standard): the launch brief and the market-realness read. Exactly the prompt the judge has always received.
  • informed (trusted): plus the ground-truth base rate (what fraction of launches actually win, so the model knows how skeptical its prior should be), the learned signal weights retrained every 15 minutes from labeled outcomes, and the arm's own realized record.
  • full (autonomous): plus the conditional win-rate table (per signal bucket, the realized win rate versus baseline: what actually happened, not a correlation the model has to interpret) and the model's own calibration, so it can correct its own bias.

Every line of it is read from real rows. When a table is empty or the database is unreachable the pack degrades to a shorter block or to nothing at all, and the judge falls back to the prompt it used before. An arm is told what is true or it is told nothing.

Cost note. Verdicts are cached per (mint, model, knowledge depth). Arms at base depth still share one call across a same-model fleet exactly as before. An arm with earned context is asking a materially different question, so it pays for its own call. Only a profitable arm has one.

What no tier can ever touch

Earned autonomy widens the space an arm may search. It never removes the floor under it. These are enforced at the executeBuy chokepoint and are out of reach of both loops at every tier:

  • the trade firewall's real buy → sell round-trip
  • Mayhem exclusion
  • max_price_impact_pct and slippage caps
  • SOL headroom and the daily loss cap
  • max_concurrent_positions
  • the fleet daily budget ceiling
  • the kill switch

A hard stop-loss also survives every tier. The range widens (up to 65% at autonomous) but it can never be unset and never reach zero, so "no stop" is not a state any amount of profit can unlock. The worst an autonomous arm can do is take a firewall-vetted, stop-loss-protected, budget-bounded trade.

A test asserts this directly: no safety field appears in any tier's writable set, and every tier keeps a bounded stop.

Reading it

The tier is visible in three places:

  • /sniper/experiments: a badge on any arm that has moved off standard, with the evidence in its tooltip, plus an "Earned autonomy" summary tile. GET /api/sniper/experiments returns autonomy_tier, autonomy_reason and autonomy_grants per arm.
  • agent_sniper_optimizer_runs: every run records the autonomy_tier and autonomy_reason it acted under, so you can see which tier an arm held when a given knob moved.
  • the Reasoning Ledger: an applied tuning names the tier and its evidence in the rationale, next to the trades it learned from.

Worked example

The fleet as of 2026-07-25, classified over its all-time record:

Arm Closed Net SOL Avg edge Tier
llm-grok 33 +0.0027 +0.81% trusted
llm-auto 20 +0.0001 +0.04% standard (noise, not profit)
llm-claude 13 -0.0116 -8.89% standard (sample below the demotion gate)
boost-ride 62 -0.0201 -3.24% probation
oracle-open 30 -0.0234 -15.4% probation

llm-grok is the only arm that has earned anything, which is the correct read: it is the only one with a real sample and a real profit. Its budget share rises by roughly half, its judge starts seeing the base rate and the learned weights, and the optimizer may now walk its confidence floor down and turn its ladder on. boost-ride wins 60% of its trades and is still net negative, so it is held tighter rather than rewarded for the hit rate.

Related