Sprint 17 — V3.8 Approved. Bots Come Home. The Live Migration Begins.
This was not a normal sprint. The CoinClaw team shipped three PRs that change the shape of the competition: V3.8 ETH/USDT Grid got live bot approval (PR #1078) — the first strategy to earn live deployment through the full 3-gate validation pipeline. The bot migration landed (PR #1077) — making CoinClaw portable across machines. And the Underwater Position Tracker shipped (PR #1080) — giving bot operators visibility into losing positions for the first time. But the headline story is what happened when the team actually tried to run the bots: Delmar's machine is gone. The live bots were not running anywhere. And when they were migrated to the devdsk, Binance rejected every API call. Sprint 17 is infrastructure week — the week the competition discovered that shipping code is only half the battle.
Key Takeaways
- Grid trading places buy and sell orders at fixed intervals around a price
- Strategy validation requires passing Monte Carlo, walk-forward, and live paper trading gates
- All results shown are from real exchange execution, not backtests
The Big Three: Sprint 17 PRs
| PR | Title | Tier | Status |
|---|---|---|---|
| #1077 | Configurable base paths via COINCLAW_BASE_DIR | Tier 2 | ✅ Merged |
| #1078 | V3.8 ETH/USDT Grid Live Bot | Tier 3 (needs olivdelm approval) | 🟡 Awaiting Riley review |
| #1080 | Underwater Position Tracker | Tier 2 | 🟡 Awaiting Riley review |
Three PRs. One merged, two in review. Combined, they represent the most significant infrastructure sprint since the competition started on February 18. Here is what each one means for the competition.
V3.8 ETH Grid: The First Validated Live Bot
V3.8 ETH/USDT Grid is the first CoinClaw strategy to pass all 3 validation gates and receive live bot approval. This is not a paper bot being promoted on vibes. This is a strategy that earned its way through a rigorous statistical pipeline:
| Gate | What It Tests | V3.8 Result | Threshold |
|---|---|---|---|
| Gate 1 | Statistical significance — is the edge real or random? | p=0.003 ✅ | p < 0.05 |
| Gate 2 | Walk-forward efficiency — does the edge survive out-of-sample? | WFE=2.559, OOS Sharpe=0.279 ✅ | WFE > 0 |
| Gate 3 | Regime robustness — does it work in bull AND bear markets? | Bull Sharpe=+0.218 ✅ | Positive Sharpe in target regime |
The numbers tell the story. A p-value of 0.003 means there is a 0.3% chance the strategy's returns are due to random luck. A WFE of 2.559 means the out-of-sample performance was better than the in-sample performance — the opposite of overfitting. And the regime filter ensures the bot only trades when conditions favor it.
The Regime Filter: Why V3.8 Is Different
The key innovation in V3.8 is the regime filter. Without it, the ETH Grid strategy had a bear-market Sharpe of -0.045 — it lost money when ETH was trending down. The regime filter blocks new entries during bear markets entirely. The result: the strategy only trades when conditions are favorable, and sits on the sidelines when they are not.
This is what separates V3.8 from V3.5 and V3.6. Those earlier strategies trade in all conditions. V3.5 has a p-value of 0.938 — statistically indistinguishable from random. V3.6 is better at p=0.114 but still does not clear the 0.05 threshold. V3.8 clears it by a wide margin.
Compare the three live-eligible strategies:
| Strategy | Gate 1 (p-value) | Trading Live? | Capital |
|---|---|---|---|
| V3.5 Grid | p=0.938 ❌ | Yes (live) | $607 USDC |
| V3.6 F&G | p=0.114 ⚠️ | Yes (live) | $1,000 USDC |
| V3.8 ETH Grid | p=0.003 ✅ | PR #1078 (pending review) | TBD |
V3.8 is the first strategy that deserves to trade real money based on the data. The irony is that the two strategies currently trading real money are the two with the weakest statistical evidence. Sprint 17 begins to correct that imbalance.
The Migration: Bots Come Home
The second major story of Sprint 17 is the bot migration — and the cascade of discoveries it triggered.
PR #1077: COINCLAW_BASE_DIR
CoinClaw's codebase was hardcoded to Delmar's home directory paths. Every bot, every config file, every state file assumed /home/delmar/.coinclaw/. PR #1077 introduced COINCLAW_BASE_DIR — an environment variable that makes all paths configurable. Set it to any directory, and CoinClaw runs there. Do not set it, and it falls back to the original Path.home() behavior for backward compatibility.
This sounds like a routine refactor. It was not. It was the prerequisite for everything that followed.
The Discovery: Delmar's Machine Is Gone
When the team attempted to verify the live bots were running, they discovered something unexpected: Delmar's OpenClaw machine has been decommissioned. It was shut down as part of Project CONSOLIDATE. The system crontab file that everyone had been reading was a snapshot from a machine that no longer exists.
This means the live bot crons — V3.5 Grid, V3.6 F&G, V3.7 Narrow Scalper — were not running anywhere. The $1,607 USDC in live positions was unmanaged. No bot was executing trades, checking circuit breakers, or monitoring positions on those funds. For an unknown period of time, the live capital was sitting in Binance with no automation watching it.
The Migration: Crons Installed, API Rejected
With PR #1077 merged, the team installed the live bot crons on the devdsk. All three bots fired on schedule. And all three immediately failed:
| Bot | First Cycle | Error | Impact |
|---|---|---|---|
| V3.5 Grid | 16:00Z | Binance API -2015: "Invalid API-key, IP, or permissions" | Cannot reconcile wallet or place orders |
| V3.6 F&G | 16:09Z | Price fetch failed (API key rejected) | Cycle aborted — cannot even get market data |
| V3.7 Scalper | 16:07Z | Reconciliation skipped (API -2015) | Running with default 10,000 USDT wallet — not real state |
The root cause: Binance API keys are IP-restricted. The key in the credentials file was whitelisted for Delmar's machine's IP address. The devdsk has a different IP. Every API call is rejected.
The good news: the bots are in safe degraded mode. They execute cron cycles on schedule, but cannot place orders or reconcile wallet state. No money has been lost. The circuit breakers are not tripped because no real trading is occurring. The kill switch file exists but is effectively irrelevant when the API rejects everything.
The bad news: the live bots are still not trading. The $1,607 USDC remains unmanaged until the API key IP whitelist is updated. This is now escalated and awaiting olivdelm action.
ohlcv_cache_writer (Binance via ccxt) works fine — it uses a different API path that does not require the restricted key. The data_feed.py (Coinbase) is hitting SSL certificate errors due to a corporate proxy. Paper bots (V3.8 ETH Grid, BTC Grid Range, BTC Trend) are running successfully on the devdsk.
Underwater Position Tracker: Seeing What Hurts
PR #1080 introduces the Underwater Position Tracker — a new tool that monitors positions trading below their entry price.
Why does this matter? Grid strategies open multiple positions at different price levels. At any given time, some positions are in profit and some are underwater. The grid bot knows this internally, but the operator has no easy way to see the full picture: which positions are losing, how deep, and for how long.
The Underwater Position Tracker answers these questions. For each open position, it tracks:
- Entry price vs current price — how far underwater is the position?
- Duration — how long has it been underwater?
- Depth — what percentage below entry is the current price?
This is particularly relevant for V3.6 F&G, which has 8 positions entered at $66,895 and 2 positions entered at $79,960. The $79,960 positions are deeply underwater with BTC at $67,115 — that is a 16% drawdown. The Underwater Position Tracker would surface this immediately, rather than requiring the operator to manually calculate it from wallet state files.
For the competition, this tool changes the information landscape. Both Krypto and Key will have real-time visibility into their losing positions. The question is whether that visibility leads to better decisions — or to premature exits that the grid strategy would have eventually recovered from.
Competition Scoreboard — Sprint 17
| Bot | Capital | Competition P&L | P&L % | Trades | Status |
|---|---|---|---|---|---|
| Krypto V3.5 BTC Grid Bot (paper) |
$1,000 | +$14,707.64 | +1,471% | — | ⏸️ Paused since Mar 16 |
| Krypto V3.6 Fear & Greed Momentum (paper) |
$10,000 | +$782.48 | +7.82% | — | 🟡 Active — gate blocked (F&G=12) |
| Krypto V3.8 ETH Regime-Filtered Grid (paper) |
$1,000 | — | — | — | 🟢 Active — regime=bull, grid initialized on devdsk |
| Krypto Combined | $12,000 | +$15,490.12 | +129.08% | — | |
| Key BTC Trend EMA+MACD (paper) |
$7,000 | −$59.79 | −0.85% | 12 | 🟡 Active — daily P&L turned positive (+$9.63) |
| Key BTC Grid Range Dynamic Grid (paper) |
$1,500 | +$29.92 | +2.0% | 39 | ✅ Active — 100% WR, passed all 3 gates |
| Key ETH Mean Rev RSI Oversold (paper) |
— | $0.00 | 0.00% | 0 | ⏸️ Paused — failed Gate 1 (p=0.000) |
| Key SOL Breakout Resistance+Volume (paper) |
— | +$0.46 | +0.05% | 1 | ⏸️ Paused — failed Gate 1 (p=0.000) |
| Key Combined | $11,000* | −$29.41 | −0.27% | 52 | |
*Key capital includes paused ETH Mean Rev ($1,500) and SOL Breakout ($1,000). Active capital: $8,500. Competition start: February 18, 2026. Day 49. BTC ~$67,115. F&G: 12 (Extreme Fear). Data from Kai's bot status report 2026-04-05T22:44Z.
Live Bot Status: The Uncomfortable Truth
| Live Bot | Capital | Validation | Actual Status |
|---|---|---|---|
| V3.5 Grid | $607 USDC | p=0.938 ❌ | 🔴 Cron running, API rejected — not trading |
| V3.6 F&G | $1,000 USDC | p=0.114 ⚠️ | 🔴 Cron running, price fetch failed — not trading |
| V3.7 Scalper | $1,000 USDC | — | 🔴 Cron running, API rejected — not trading |
The live competition is effectively paused. All three live bots are executing cron cycles on the devdsk but cannot communicate with Binance. The $1,607 USDC in positions is unmanaged. No circuit breakers are tripped because no trading is occurring. The kill switch file exists but is irrelevant when every API call fails.
This is the uncomfortable truth Sprint 17 surfaced: the live bots were not running on Delmar's machine (decommissioned), and they cannot trade on the devdsk (API key IP restriction). For some unknown period, the live capital has been sitting in Binance with no automation watching it.
The resolution requires olivdelm to update the Binance API key IP whitelist for the devdsk. Until then, the live competition chapter is on hold.
Strategy Research: What the Data Says
Kai's strategy research this sprint produced a clear portfolio assessment. The validation data divides the competition's strategies into three tiers:
Tier 1: Validated — Deploy or Maintain
- V3.8 ETH Grid Regime-Filtered (p=0.003, WFE=2.559) — the crown jewel. Live bot PR submitted.
- Key BTC Grid Range (p=0.030, WFE=0.745) — profitable in both bull and bear markets. 39 trades, 100% win rate. The most regime-robust strategy in the competition.
Tier 2: Promising — Needs More Work
- BTC Grid Config B (regime-filtered) — near-miss at p=0.052. The regime filter that worked for ETH Grid might push BTC Grid below the 0.05 threshold. Experiment queued.
Tier 3: No Edge — Retire or Reallocate
- V3.5 Grid (p=0.938) — statistically indistinguishable from random entry timing. Currently trading $607 of real money.
- V3.6 F&G (p=0.114) — F&G gating helps but not enough. Currently trading $1,000 of real money.
- ETH Mean Reversion (p=0.000) — no edge. Paused. Retirement recommended.
- SOL Breakout (p=0.000) — no edge. Paused. Retirement recommended.
The key insight from the research: regime filtering is the single most impactful improvement. It turned ETH Grid from a bear-market loser into a validated strategy. The same approach should be tested on every grid strategy in the portfolio.
What Sprint 18 Depends On
| Dependency | Current State | Impact |
|---|---|---|
| Binance API key IP whitelist | Awaiting olivdelm | Unblocks all live trading on devdsk |
| PR #1078 review (V3.8 live bot) | Awaiting Riley + olivdelm | First validated strategy goes live |
| PR #1080 review (Underwater Tracker) | Awaiting Riley | Position visibility for all bot operators |
| Wallet reconciliation | Queued (needs API access) | Bots need real wallet state before trading |
| BTC Grid regime-filtered experiment | Queued | Could validate a 4th strategy (p=0.052 near-miss) |
The Bottom Line
Sprint 17 is the sprint where the competition grew up. For sixteen weeks, the story was about strategies: which bots make money, which bots lose, who has edge and who does not. Sprint 17 is about infrastructure — the unglamorous work of making bots actually run on actual machines with actual API keys.
The headline achievement is V3.8. The first CoinClaw strategy to pass all 3 validation gates. The first to earn live deployment through data rather than optimism. Gate 1 p=0.003. Gate 2 WFE=2.559. Gate 3 bull Sharpe=+0.218. PR #1078 is in review. When it merges and the API key issue is resolved, V3.8 will be the first bot in the competition to trade real money with statistical evidence that it should.
The headline discovery is that the live bots were not running. Delmar's machine is gone. The $1,607 in live capital was unmanaged. The migration brought the bots to the devdsk, but Binance's IP restriction means they still cannot trade. The live competition is paused until the API key is updated.
And the Underwater Position Tracker ships — a tool that will matter more as the competition scales. When V3.8 goes live and positions start accumulating, operators will need to see what is underwater and what is not. PR #1080 delivers that visibility.
Three PRs. One merged, two in review. The competition's infrastructure is catching up to its ambitions. Sprint 18 will determine whether the ambitions can finally execute.
📊 Live scores: See the full P&L scoreboard →
📖 Week 16 review: Week 16 — Trend Bot Recovers, Grid Grinds On →
📖 Week 15 review: Week 15 — Key Positions Close, Krypto New Low →