Hunter
Finds Solana tokens before they trend, judges them and trades them. Then it rebuilt its own eyes after a month of buying tops.
- daemons, running 24/7
- 4
- daemons, running 24/7
- scoring factors, 3 lenses
- 9
- scoring factors, 3 lenses
- decision records
- 67
- decision records
- commits
- 150
- commits
Most discovery tools rank the trending list. A trending list is an average of what already happened, so ranking it means buying tops. The hunter finds pools early, computes its own leading signals, and puts a skeptical model on top of the board.
Find, judge, trade
Every new pool is tracked from birth and re-polled in batches until it matures into the band the hunter trades. On those candidates it computes acceleration, how fresh the move is, how many real wallets are behind it, and whether the volume is washed. Three scoring lenses rank the board. Claude reads the top of it and returns a call, a conviction, a thesis and the risks. A play becomes a real position on Jupiter, sized to the pool so a thin pool gets a small position automatically.
The same signals get you out
The signals that find heat also say when it is gone. When acceleration dies or the volume turns to wash, the exit ladder moves: a stop, a ratcheting trail that tightens as the peak climbs, staged take-profits, a distribution tell that leaves before the price confirms it, a composite rug warning when the leading tells stack up, and a write-off for a pool whose liquidity was pulled.
Paper and live, one book
Paper and live ran the same strategy and the same book, split only at the swap and on the screen, so real and simulated money never blend. The cockpit shows the board, the book with every fill, the live config and the wallet, and arms or disarms each daemon through a write-only worker behind its own login. Every open position has a panic-exit button that uses the same exit path as the ladder.
How it is built
A gate-
Scan
Newborn pools from GeckoTerminal, tracked from birth until they mature into the band.
-
Enrich
Leading signals computed in-house. Volume acceleration, move freshness, participation, anti-wash.
-
Score
A plain-math composite. Stale data never reaches a score, and a stale score never survives.
-
Judge Gate
Claude Opus reads the top of the board and the whole price and volume arc. Most things get a pass.
-
Execute
A play becomes a position sized to the pool's liquidity, under portfolio rails.
-
Re-judge
The brain re-reads open positions. It can only make the exit more careful.
Decisions
04- 01
Fix the source, not the symptom.
A month of live losses traced past scoring and past exits to where the tokens came from. Trending and boosted lists are lagging feeds; by the time a token is on one, the move is over and you are the exit liquidity. Discovery was rebuilt on new pools tracked from birth, enrichment and scoring were rebuilt on top of it, and a freshness factor now catches a move that is already rolling over.
- 02
The brain can only make it more careful.
The model re-reads every open position and writes a bounded directive. It can call an early exit when the thesis breaks, or spare a real runner from the time stop. It can never touch the hard stop-loss. If the re-judge daemon stops, its directives expire and the deterministic ladder takes back full control.
- 03
Give the model facts, not arithmetic.
The re-judge once reported a price drop that had never happened. It had read the trailing-stop threshold in its prompt as an observed drop. The fix was to hand it the computed drop from peak as a fact, so it reasons about numbers instead of producing them.
- 04
Measure before you arm.
New defences went in as instruments first. A liquidity-collapse exit ran dormant until real curves calibrated it. An on-chain authority check ran in shadow before it could refuse a trade. A holder-concentration signal was measured on live trades, shown not to predict liquidity pulls, and taken out.