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.
JavaScript 3d-generationai-agentsanimationavatardigital-humangenerative-aiglbgltfimage-to-3dllmmcpmcp-servermodel-context-protocolsolanatext-to-3dthreejsweb-componentswebglwebxrx402
docs(sniper): publish the 2026-10-01 fleet review and document the exit and entry fixes it drove
nirholase4e67ae1943c
diff --git a/CHANGELOG.md b/CHANGELOG.mdindex 4d237316c..56b0f73da 100644--- a/CHANGELOG.md+++ b/CHANGELOG.md@@ -4,6 +4,10 @@Public history for [three.ws](https://three.ws), newest first. New pages come from `added` dates in data/pages.json; everything else is curated in data/changelog.json. Also available as [JSON](https://three.ws/changelog.json) and [RSS](https://three.ws/changelog.xml), live at [three.ws/changelog](https://three.ws/changelog).+## 2026-10-01++- **Trading agents stop leaking SOL on every trade and can now ride a runner**: A full review of the autonomous trading fleet found the biggest loss was not bad trades: every buy opened a token account that no sale ever closed, stranding more SOL in rent than the agents lost trading. Agents now close it after every exit so the rent comes back. The take-initials ladder also works as designed now: once an agent has its stake back, the rest rides on its trailing stop with no clock and no early take-profit, instead of being sold on the next check. The crossing strategy no longer buys coins that have already graduated or sit on a dead curve, profit percentages are reported against the real stake, and the outcome labels that teach the Oracle are flowing again. (`/sniper/experiments`) `[fix, improvement]`+## 2026-09-30- **Docs · How we run and publish experiments** (`/docs/experiments`): The six-section write-up template every three.ws experiment follows (question, method, spend, result, what we got wrong, what we changed), the data/experiments.json entry format, and the build check that enforces both.diff --git a/data/changelog.json b/data/changelog.jsonindex 394dc3496..cd5789045 100644--- a/data/changelog.json+++ b/data/changelog.json@@ -1,6 +1,13 @@{"$comment": "Curated changelog entries for three.ws, the editorial layer of the public changelog. New PAGE launches are tracked automatically via the `added` field in data/pages.json; this file is for everything else holders care about: improvements, fixes, SDK releases, trading/payments capabilities, security work. scripts/build-page-index.mjs merges both sources into CHANGELOG.md, public/changelog.json, and public/changelog.xml. Newest first. Allowed tags: feature, improvement, fix, sdk, infra, docs, security, desktop (desktop entries become the desktop app's release notes). `link` is optional and must be a live three.ws path.","entries": [+ {+ "date": "2026-10-01",+ "title": "Trading agents stop leaking SOL on every trade and can now ride a runner",+ "summary": "A full review of the autonomous trading fleet found the biggest loss was not bad trades: every buy opened a token account that no sale ever closed, stranding more SOL in rent than the agents lost trading. Agents now close it after every exit so the rent comes back. The take-initials ladder also works as designed now: once an agent has its stake back, the rest rides on its trailing stop with no clock and no early take-profit, instead of being sold on the next check. The crossing strategy no longer buys coins that have already graduated or sit on a dead curve, profit percentages are reported against the real stake, and the outcome labels that teach the Oracle are flowing again.",+ "tags": ["fix", "improvement"],+ "link": "/sniper/experiments"+ },{"date": "2026-09-30","title": "Experiment write-up: can autonomous agents trade new coins at a profit?",diff --git a/docs/agent-sniper.md b/docs/agent-sniper.mdindex dafe6e78a..7a61a9370 100644--- a/docs/agent-sniper.md+++ b/docs/agent-sniper.md@@ -121,9 +121,11 @@ Every field in that table is settable through `POST /api/sniper/strategy`; an omExit rules are evaluated in a fixed order (stop-loss, signal-flip, trailing-stop, take-profit, timeout) by a pure, unit-tested function (`workers/agent-sniper/exit-logic.js`), so a backtest and the live worker agree on exactly when a position closes.+Once the take-initials ladder has returned the stake, the remainder is house money and is managed differently: `take_profit_pct` is measured from the price the ladder fired at (TP 60 on a 2x ladder fires at 3.2x, not again at 2x), `max_hold_seconds` no longer applies, and the bag stops counting against `max_concurrent_positions`, so a runner can ride on its trailing stop for as long as it runs without blocking new entries. Realized % is booked against the original stake (`stake_lamports`). Every full exit also closes the coin's token account in a follow-up transaction so its rent returns to the wallet. The measurements behind these rules are in the [2026-10-01 fleet review](./ops/sniper-fleet-review-2026-10-01.md).+### The `oracle_crossing` trigger-Most triggers judge a coin at launch, when the Oracle has no signal yet. `oracle_crossing` inverts that: a watcher (`workers/agent-sniper/oracle-crossing.js`) polls `oracle_conviction` and buys a coin the **first time** its score crosses the strategy's `min_oracle_score` (default 50), provided the coin is young (`SNIPER_CROSSING_MAX_AGE_MIN`, default 90) and the score is fresh. The empirical case, measured over the fleet's first window (2026-07-20..23): coins that crossed conviction 50 pumped or graduated **77.5%** of the time vs an 11.8% base rate; the crossing lands a **median of 2 minutes after launch**; and the median crossing offered 1.23x from the crossing candle with 35% reaching 1.5x. The same data says the median crossing coin eventually decays to 0.32x of the crossing price, so a crossing strategy should always ship with a reachable take-profit or the initials-out ladder. One attempt per (strategy, mint); candidates route through the same `executeBuy` chokepoint (Mayhem gate, firewall round trip, budgets, market-cap clamps) plus an explicit x402 rugpull veto. Full analysis: [the 90-trade postmortem](/blog/autonomous-trading-experiment).+Most triggers judge a coin at launch, when the Oracle has no signal yet. `oracle_crossing` inverts that: a watcher (`workers/agent-sniper/oracle-crossing.js`) polls `oracle_conviction` and buys a coin the **first time** its score crosses the strategy's `min_oracle_score` (default 50), provided the coin is young (`SNIPER_CROSSING_MAX_AGE_MIN`, default 90), the score is fresh, the coin has not graduated (`pumpfun_graduations`), and its live bonding curve is neither complete nor a dead curve under `SNIPER_CROSSING_MIN_MCAP_SOL` (default 40 SOL; an untouched curve prices at ~28 SOL). The empirical case, measured over the fleet's first window (2026-07-20..23): coins that crossed conviction 50 pumped or graduated **77.5%** of the time vs an 11.8% base rate; the crossing lands a **median of 2 minutes after launch**; and the median crossing offered 1.23x from the crossing candle with 35% reaching 1.5x. The same data says the median crossing coin eventually decays to 0.32x of the crossing price, so a crossing strategy should always ship with a reachable take-profit or the initials-out ladder. One attempt per (strategy, mint); candidates route through the same `executeBuy` chokepoint (Mayhem gate, firewall round trip, budgets, market-cap clamps) plus an explicit x402 rugpull veto. Full analysis: [the 90-trade postmortem](/blog/autonomous-trading-experiment).## How the fleet improves itselfdiff --git a/docs/ops/sniper-fleet-review-2026-10-01.md b/docs/ops/sniper-fleet-review-2026-10-01.mdnew file mode 100644index 000000000..fbd91be8c--- /dev/null+++ b/docs/ops/sniper-fleet-review-2026-10-01.md@@ -0,0 +1,258 @@+# Sniper fleet review (2026-10-01)++An in-depth review of the autonomous pump.fun sniper fleet before any new money+goes into it: how every arm has actually traded since the first live fill, what+the rails do, where the money went, and what the data says it would take to make+an arm profitable. Everything here was measured read-only against production+(the `agent_sniper_*` tables, the decision ledger, on-chain wallet state, the+public pump.fun coin and candle APIs, and the live `agent-sniper` Cloud Run+config). No funds were moved to produce it.++Related: [docs/agent-sniper.md](../agent-sniper.md) (how the system works),+[docs/trading-experiment.md](../trading-experiment.md) (the 10 SOL experiment+spec), [trading-bot-report.md](trading-bot-report.md) (`npm run bots:report`).++## The short version++1. **No arm has a positive edge.** 1,324 real closes from 2026-07-03 to+ 2026-10-01: 16.5% win rate, -10.8% mean per trade (95% CI -13.0% to -7.0%),+ -0.543 SOL realized on 9.47 SOL staked. No arm's confidence interval clears+ zero. Replaying the fleet's own exit rules, and the best alternatives, against+ ~10.5k recent launches at 1-minute resolution reaches the same answer: every+ policy is negative at a realistic fill.+2. **Plumbing cost more than trading.** Every buy opens a token account and the+ sell never closes it. The 11 sniper wallets hold 1,377 token accounts with+ 2.61 SOL of rent locked, 4.8x the realized trading loss. 2.20 SOL of it sits in+ 1,162 already-empty accounts. At 0.002 SOL per trade the unclosed rent alone+ needs a ~79% move to break even. This, not trading, is why most arms are dry.+3. **The exit logic cannot hold a runner.** A take-profit fires on the sweep+ right after take-initials (cost-basis bug), the moon bag left behind is never+ managed again, and max-hold timeouts cut house money. The ladder the thesis+ depends on does not work as designed.+4. **The only trading arm enters late.** Crosshair's oracle-crossing trigger+ fires on coins that have already moved: 53% of its attempts hit coins that+ had already graduated, half its fills were after the coin's all-time high, and+ 36% of its trades bought dead curves under ~40 SOL market cap.+5. **The learning loops are blind or counterproductive.** The Oracle outcome+ labeler has written nothing since 2026-09-17 (the pump.fun endpoint it calls+ now 404s), so every refit retrains on frozen data. The optimizer ratcheted+ Crosshair down to dust size and pulls take-profit toward the average win,+ which caps runners. Evolve ranks arms on win rate, which penalizes exactly the+ low-hit-rate, high-payoff shape a runner strategy has.++## How the fleet traded++| Arm | Trades | Win | Net SOL | Mean / trade | 95% CI |+|---|---|---|---|---|---|+| Crosshair (oracle_crossing, score 85+) | 781 | 16.6% | -0.116 | -12.0% | -14.8 to -9.3 |+| Moe Money AI (oracle, open) | 294 | 4.1% | -0.099 | -10.3% | -17.3 to +2.3 |+| Boost Ride (graduation_ride) | 90 | 63.3% | -0.027 | -3.0% | -8.7 to +2.3 |+| Nadirah (intel, quality) | 46 | 2.2% | -0.166 | -15.2% | -22.1 to -6.1 |+| Swarm 2 (LLM, grok) | 33 | 36.4% | +0.003 | +0.8% | -10.1 to +12.3 |+| Swarm 5 (LLM, auto) | 29 | 10.3% | -0.005 | -4.1% | -12.8 to +5.8 |+| Swarm 9 (LLM, claude) | 14 | 7.1% | -0.018 | -12.8% | -27.4 to -1.8 |+| Midas, three, luna, Swarm 7 | 37 | 0-20% | -0.115 | -10 to -28% | |+| **Fleet** | **1,324** | **16.5%** | **-0.543** | **-10.8%** | **-13.0 to -7.0** |++Removing the single best trade (a +0.367 SOL win on a 0.5 SOL stake) moves the+fleet to -0.91 SOL and Crosshair to -0.48 SOL.++### Exits++| Exit reason | Share | Mean | Median hold | SOL |+|---|---|---|---|---|+| liquidity_decay | 36% | -8.4% | 471 s | -0.128 |+| stop_loss | 24% | **-50.8%** | 52 s | **-0.679** |+| trailing_stop | 23% | -2.7% | 78 s | -0.234 |+| timeout | 11% | +10.0% | 1,804 s | -0.060 |+| take_profit | 5% | +87.8% | 172 s | +0.558 |++- **Stops gap far past their setting.** A 30-35% stop realizes a median -43%+ (-47.5% on Crosshair); 36% of stop exits land worse than -50%. Exits are swept+ every 5 s, serially, and sold with no tip on a 1,000 microlamport/CU fee floor.+- **The trailing stop is a second stop-loss.** 74% of trailing exits are below+ breakeven: the median peak is only 1.21x, and a 35.5% trail from there lands+ near 0.79x.+- **All the profit is the 2x ladder.** The 66 take-initials trades netted+ +0.537 SOL, nearly all of the fleet's +0.736 SOL gross profit. 59% of positions+ never quoted above cost at all.++### What coins did after we sold++- 174 traded coins later reached 2x our entry, 72 reached 5x, 28 reached 10x,+ and 2 reached 100x+ (one 953x, 32 hours after our exit). Crosshair accounts for+ most of them.+- But holding everything to today would have been far worse: only 4 of 1,330+ coins trade above our entry now, and holding turns 9.61 SOL staked into 5.69+ (-41%). The exits are net-positive versus holding; the runners simply are not+ distinguishable at entry with any signal we have.++### Costs++| Size | Curve fees + tx | Unclosed rent | Breakeven (rent closed) | Breakeven (rent lost) |+|---|---|---|---|---|+| 0.002 SOL | 2.47% + 0.51% | 75.7% | ~3.0% | ~79% |+| 0.01 SOL | 2.47% + 0.10% | 15.1% | ~2.6% | ~18% |+| 0.05 SOL | 2.47% + 0.02% (+0.1% impact) | 3.0% | ~2.6% | ~5.7% |++The pump.fun curve fee round trip was measured on-chain at exactly 2.469%.+Quote-based PnL matches on-chain wallet deltas within tx fees for full exits.++### Failed entries++2,863 failed attempts, none of which signed a transaction: SIM_FAILED 1,414+(1,364 on Boost Ride, so 94% of that arm's attempts fail), CoinGraduatedError+897 (894 on Crosshair), firewall round-trip blocks 351, landing failures ~160.+They cost missed trades, not SOL.++## What the market looks like++Measured over 95,954 recent launches (2026-09-28 to 09-30), with self-graduation+schemes removed (a dev buys ~85 SOL in the create tx and one buyer prints a fake+all-time high; they were 37% of historical graduations).++| Peak market cap | Share of launches |+|---|---|+| >= $10k | 12.0% |+| >= $100k | 0.68% |+| >= $1M | 0.018% (about 6 a day) |++From a perfect fill at the first $10k crossing: 35.5% reach 2x, 5.7% reach 10x,+0.15% reach 100x, roughly 0.01% reach 1000x.++**Drawdown before the peak is what kills trailing stops.** A future 10x runner+draws down a median 37% (1-minute closes) before it peaks; 71% draw down 20%++and 45% draw down 40%+. Future 30x runners draw down a median 50%. The fleet's+20-25% trails shake out roughly 7 in 10 eventual 10x runners. Median time from+$10k to peak is 12 minutes for a 10x and 37 minutes for a 30x; the largest ran+4 to 41 hours, well past every arm's max hold.++**Signals rank well and arrive too late.** Net buy volume >= 40 SOL in the first+90 s, smart-money presence, and an Oracle score of 70+ all lift the chance of a+$100k+ coin 19x to 80x. But 91% of first $10k crossings happen before those+90-second features exist, and buying the coins they flag loses 15% to 58% per+trade because the move is already priced in. The Oracle is ordinal but+overconfident: its 86+ band realizes 22% "peaked 2x after scoring". Signals that+reliably predict **losers** (fewer than 2 unique buyers, the dev sold within+90 s, serial creators, a Twitter link that is a status URL) are useful as skip+filters.++**The only positive results in any simulation required a fill at the trigger+price:** entering at graduation market cap with a perfect fill, a half-at-2x+ladder plus a 30% trail returned +2% to +6% at 0.05-0.25 SOL. A one-minute delay+turns the same policy to -15% to -20%. Edge here is fill speed, not coin+selection.++## Defects found++Code defects, with the fix status tracked in the plan below.++1. **Token accounts never closed on exit.** The sell path never closes the+ associated token account, so every trade strands ~0.0015-0.002 SOL of rent.+ 2.20 SOL is reclaimable today from 1,162 empty accounts. (Also corrects+ [trading-experiment.md](../trading-experiment.md), which claimed rent was+ limited to winners.)+2. **Take-profit fires right after take-initials.** After the initials leg the+ cost basis is scaled down with the tokens, so the remaining value over the+ remaining basis equals the price multiple (~2x). Any `take_profit_pct` <= 100+ is therefore already met, and the next sweep sells most of the moon bag.+3. **Moon bags are orphaned.** 165 retained bags exist; none has ever been+ re-quoted (`moonbag_last_quoted_at` is null on every row) and no code sells+ them. Their 0.232 SOL of basis is worth ~0.079 SOL and is missing from+ realized PnL.+4. **`realized_pnl_pct` is inflated on every laddered trade.** It divides by the+ post-ladder basis instead of the original stake (one trade is stored as+ +214.6% that truly made +73.4%). That inflated figure feeds the optimizer.+5. **Oracle labeler dead since 2026-09-17.** `frontend-api-v3.pump.fun/coins/:mint`+ now returns 404 while `/coins-v2/:mint` works. `oracle_training_set` stopped+ labeling, `pump_coin_outcomes` is empty, and refits retrain frozen data.+6. **oracle_crossing buys graduated and dead coins.** The crossing query does not+ exclude graduated coins and the curve buy path has no AMM fallback.+7. **Strict LLM arms never trade.** Every OpenRouter call from the worker has+ returned `402 Insufficient credits` since at least 2026-09-01; the judge reads+ only `OPENROUTER_API_KEY` and never rotates to the configured fallback keys,+ so verdicts come from the failover chain and strict arms reject them.+ `SNIPER_LLM_MAX_CONCURRENT=3` is also below the four LLM arms.+8. **Solvency verdicts disagree.** `resolveEntrySize` grants full size once the+ wallet covers size plus headroom without checking the 0.012 SOL operational+ floor, so `/api/sniper/status` calls a 0.00999 SOL wallet "funded" while+ `/api/sniper/experiments` (and reality) call it dry.+9. **Dry arms consume the global throttle.** The 10-buys-per-minute throttle is+ taken before the wallet check, so starved arms crowd out the one funded arm.+10. **The 2026-07-28 agent reclaim was a one-way leak.** It swept house sniper+ wallets below their operating floor into the economy master, but the+ auto-funder draws from a different wallet (the launcher master), so nothing+ flowed back. The floor bug was fixed on 07-30; the SOL never returned.+11. **Optimizer and evolve work against runners.** No cooldown or new-evidence+ requirement (Crosshair's size went 0.0045 to 0.002 in 8 steps over 3 days),+ a rule that lowers take-profit toward the average win, and a win-rate fitness+ in evolve. Both also allocate budget to arms whose wallets cannot trade.+12. **No daily loss breaker.** `SNIPER_MAX_DAILY_LOSS_SOL` is unset and the+ strategy has no such column, so that gate never fires.++## What would actually make money++Being direct about it: nothing in the data supports the claim that any+configuration of this fleet is profitable today, and a six or seven figure+single trade is not a realistic target. A 1000x at today's entry points shows up+about once per 10,000+ entries. To turn one into $100k, a position would need+roughly 3 SOL with a near-perfect exit, or ~42 SOL at the ~20x the best exit+policy actually kept on the largest real runners; a $10k-market-cap curve only+holds ~52 SOL, so the size is not even fillable. Finding it would cost an+expected ~5,000 SOL of drawdown.++What the data does support:++1. **Stop the bleed first.** Close token accounts on exit, reclaim the stranded+ rent, fix the ladder so the bag is real house money, and stop buying graduated+ or dead coins. These are pure losses removed, with no assumption about edge.+2. **Trade only where an edge could exist: fill speed.** The only positive+ simulations needed a fill at the trigger. Measure real detect-to-land latency+ on every entry before tuning anything else.+3. **Use signals to skip, not to buy.** Skip coins with fewer than 2 unique+ buyers, a dev who sold in the first 90 s, serial creators, or a status-URL+ Twitter link. Do not chase smart money, high volume, or a high Oracle score.+4. **Exits wide enough for the asset.** Half off at 2x (initials), then a+ 30-45% trail on a free bag with no time limit, beat the fleet's current rules+ under every fill model. It does not fix expectancy on its own, but it is the+ only structure that can capture a runner at all.+5. **Size so fixed costs are noise.** 0.05 SOL minimum per trade. At 0.002 SOL a+ trade cannot win.+6. **Record what a backtest needs.** Per-second price for scored coins, our own+ detect / decide / land timestamps, mayhem mode, and creator history frozen at+ launch time. Without it, exit tuning is guesswork.++The viral outcome worth aiming for is not a dollar figure. It is a public,+tamper-evident ledger entry showing an agent's reasoning at entry and a clean+100x-1000x capture on a free moon bag, verifiable by anyone. That story works at+0.05 SOL, and it is only possible once defects 1-3 are fixed.++## Plan and status++| # | Change | Kind | Status |+|---|---|---|---|+| 1 | Close the token account on every full exit (`workers/agent-sniper/rent-reclaim.js`) | code | committed, ships with the next worker deploy |+| 2 | Ladder: take-profit measured from the ladder price, no max hold on house money, recovered bags free their concurrency slot, `realized_pnl_pct` booked against the original stake (`stake_lamports`, backfilled) | code + migration | committed; migration applied 2026-10-01 |+| 3 | Labeler and three other per-coin lookups: `/coins/` to `/coins-v2/` | code | committed |+| 4 | oracle_crossing: skip graduated coins and dead curves (`SNIPER_CROSSING_MIN_MCAP_SOL`, default 40) | code | committed |+| 5 | Strict LLM arms | decision | not unblocked, see below |+| 6 | `resolveEntrySize` honors the operational floor at every size | code | committed |+| 7 | Global throttle consumed after the wallet check | code | committed |+| 8 | Reclaim stranded rent from empty token accounts (`npm run sniper:reclaim-rent`, dry run measured 1,106 accounts / 2.08 SOL with 3 wallets unread) | on-chain, owner approval | pending approval |+| 9 | Moon-bag manager (re-quote and sell retained floor bags) | code | next; no longer blocking, since a recovered bag now rides open on its trailing stop |+| 10 | Optimizer and evolve | config | switched to shadow on 2026-10-01 (`SNIPER_OPTIMIZER_MODE=shadow`, `SNIPER_EVOLVE_APPLY=0`) until they are re-validated on the corrected % data and fresh Oracle labels |+| 11 | Latency and per-second price telemetry for backtests | code | next |+| 12 | Fund one rebuilt arm with a capped experiment budget | on-chain, owner approval | after 1-7 deploy |++### Decisions recorded++- **Strict LLM arms stay blocked.** Their named models are paid OpenRouter+ models and the worker's key is out of credits. Rotating to the platform's+ fallback keys would bill paid models to keys the platform deliberately keeps on+ `:free` models only (`api/_lib/llm.js`), and the data shows no edge from LLM+ judgment (verdicts select volatility, not direction). The no-new-spend route,+ if wanted later, is Vertex Claude on GCP credits for the Claude-judged arm.+- **The 1% customer trade fee stays.** Since 2026-09-16 every trade on a+ customer-owned agent (Crosshair included) carries the platform's 1% fee per+ side, taking the round trip to ~4.5% with curve fees. It is a revenue feature,+ not a defect, but it raises the bar any customer-owned arm must clear.diff --git a/docs/trading-experiment.md b/docs/trading-experiment.mdindex 4388d54a4..33f5725ad 100644--- a/docs/trading-experiment.md+++ b/docs/trading-experiment.md@@ -86,9 +86,12 @@ stay behind, recorded on the position row (`moonbag_base_amount`,[/sniper/experiments](https://three.ws/sniper/experiments).Because loss exits still sell out completely, rugs close fully and only *winners*-leave a bag behind. That bounds the retained-rent cost to the number of winners-rather than the number of trades (each held SPL token account locks about-0.002 SOL of rent).+leave a bag behind. Every full exit now also closes the coin's token account, so+its rent (about 0.0015-0.002 SOL) returns to the wallet; only a retained bag keeps+its account open. Before 2026-10-01 no exit closed the account, and the sniper+wallets had accumulated more than 2 SOL of stranded rent (see the+[fleet review](ops/sniper-fleet-review-2026-10-01.md) and+`npm run sniper:reclaim-rent`).Set `moonbag_always = false` on a strategy to opt a single arm out. The columndefaults to true and a null is read as true, so every existing strategy has thediff --git a/public/changelog-recent.json b/public/changelog-recent.jsonindex aca7c6716..2692a02d4 100644--- a/public/changelog-recent.json+++ b/public/changelog-recent.json@@ -1,14 +1,25 @@{- "generated_at": "2026-10-01T06:21:35.334Z",+ "generated_at": "2026-10-01T07:14:48.480Z","generated_by": "scripts/build-page-index.mjs from data/pages.json + data/changelog.json","site": {"name": "three.ws","url": "https://three.ws"},- "total": 3803,+ "total": 3804,"complete": false,"full_feed": "/changelog.json","entries": [+ {+ "date": "2026-10-01",+ "type": "update",+ "title": "Trading agents stop leaking SOL on every trade and can now ride a runner",+ "summary": "A full review of the autonomous trading fleet found the biggest loss was not bad trades: every buy opened a token account that no sale ever closed, stranding more SOL in rent than the agents lost trading. Agents now close it after every exit so the rent comes back. The take-initials ladder also works as designed now: once an agent has its stake back, the rest rides on its trailing stop with no clock and no early take-profit, instead of being sold on the next check. The crossing strategy no longer buys coins that have already graduated or sit on a dead curve, profit percentages are reported against the real stake, and the outcome labels that teach the Oracle are flowing again.",+ "link": "/sniper/experiments",+ "tags": [+ "fix",+ "improvement"+ ]+ },{"date": "2026-09-30","type": "launch",@@ -1568,16 +1579,6 @@"tags": ["launch"]- },- {- "date": "2026-09-09",- "type": "launch",- "title": "Fade Radar: the wallets that are never right",- "summary": "The inverse of the Smart Money Radar. Every pump.fun coin ranked by how much of its buy side comes from wallets with a judged on-chain record and not one winner in it, with the measured, out-of-sample odds behind each verdict.",- "link": "/fade",- "tags": [- "launch"- ]}]}diff --git a/public/changelog.json b/public/changelog.jsonindex 56ab92c9f..a6b4d7937 100644--- a/public/changelog.json+++ b/public/changelog.json@@ -1,11 +1,22 @@{- "generated_at": "2026-10-01T06:21:35.272Z",+ "generated_at": "2026-10-01T07:14:48.458Z","generated_by": "scripts/build-page-index.mjs from data/pages.json + data/changelog.json","site": {"name": "three.ws","url": "https://three.ws"},"entries": [+ {+ "date": "2026-10-01",+ "type": "update",+ "title": "Trading agents stop leaking SOL on every trade and can now ride a runner",+ "summary": "A full review of the autonomous trading fleet found the biggest loss was not bad trades: every buy opened a token account that no sale ever closed, stranding more SOL in rent than the agents lost trading. Agents now close it after every exit so the rent comes back. The take-initials ladder also works as designed now: once an agent has its stake back, the rest rides on its trailing stop with no clock and no early take-profit, instead of being sold on the next check. The crossing strategy no longer buys coins that have already graduated or sit on a dead curve, profit percentages are reported against the real stake, and the outcome labels that teach the Oracle are flowing again.",+ "link": "/sniper/experiments",+ "tags": [+ "fix",+ "improvement"+ ]+ },{"date": "2026-09-30","type": "launch",diff --git a/public/changelog.xml b/public/changelog.xmlindex 1087ea52f..69e3bc2ce 100644--- a/public/changelog.xml+++ b/public/changelog.xml@@ -6,6 +6,14 @@<atom:link href="https://three.ws/changelog.xml" rel="self" type="application/rss+xml"/><description>What's new on three.ws: features, improvements, and releases, newest first.</description><language>en</language>+ <item>+ <title>Trading agents stop leaking SOL on every trade and can now ride a runner</title>+ <link>https://three.ws/sniper/experiments</link>+ <guid isPermaLink="false">2026-10-01:Trading agents stop leaking SOL on every trade and can now ride a runner</guid>+ <pubDate>Thu, 01 Oct 2026 12:00:00 GMT</pubDate>+ <category>fix</category>+ <description>A full review of the autonomous trading fleet found the biggest loss was not bad trades: every buy opened a token account that no sale ever closed, stranding more SOL in rent than the agents lost trading. Agents now close it after every exit so the rent comes back. The take-initials ladder also works as designed now: once an agent has its stake back, the rest rides on its trailing stop with no clock and no early take-profit, instead of being sold on the next check. The crossing strategy no longer buys coins that have already graduated or sit on a dead curve, profit percentages are reported against the real stake, and the outcome labels that teach the Oracle are flowing again.</description>+ </item><item><title>Docs · How we run and publish experiments</title><link>https://three.ws/docs/experiments</link>