FMZ 2014 Depth-Reach Quasi-Market-Maker (by 小草, fmz.com/strategy/1088), Faithful Port: Long-Only Full-Inventory Maker Quotes ±10bp Around the Touch, Cancel-and-Replace Every Bar, No Stop (SOLUSDT.BINANCE USD-M, 1-Minute Bars, Post-Only Maker Orders)

Port of 小草's 2014 'simple Bitcoin high-frequency bot' (FMZ strategy 1088). The strategy is by that author. Classification: (a) a concrete, tradeable strategy. Original rules, in our words: every 3.5 seconds the bot reads 20 levels of order-book depth. Its buy price is one tick above the bid level where cumulative bid size first passes 8 coins. Its sell price is one tick below the ask level where cumulative ask size first passes 8 coins. If the sell price minus the buy price is 1.5 price units or less, it moves both quotes 10 units further out. It then cancels every resting order and posts two limit orders: a sell of all coins held and a buy using all cash minus a small buffer, each only if it is larger than 0.02 coins. There is no stop-loss and no cost-basis logic. Port decisions: (1) Data. The catalog has no historical L2 depth: the order_book_delta and quote_tick folders hold only one or two snapshots from 2026-04/05. On a modern liquid perp, 8 coins of size sit at or next to the touch, so the original code's own 1.5-unit branch fires on essentially every tick. The quotes therefore become touch -10 and touch +10 units. At the ~10,000 BTC price in the author's example, 10 units = 10bp and 1.5 units = 1.5bp. We keep those defaults as relative values, using the last 1-minute close as the touch. (2) Cadence. 3.5 seconds becomes one update per 1-minute bar close. This is a representation compromise. (3) Venue. Binance spot charges 0.10% maker, which is more than the whole 20bp quote width, so spot cannot work. We use Binance USD-M (0.02% maker, 0.04% round trip), long-only at 1x, to keep the original cash/coin inventory behaviour. Hyperliquid's 1m history is only about 3 days, too short to validate on. (4) Instrument. SOL instead of BTC, because BTC is the most over-used series in the corpus. (5) Sizing. Full position, as in the original. Flat state: all cash with a resting bid, expected about 45-60% of the time. Honest prior: this is sub-hour passive mean reversion on a liquid major, which is the fee-graveyard class.

verification_loop: Verification failed (Layer 4 — QA review) [class=code_defect]: - [edge_concern] Even with sizing fixed, a 20bp-wide quote pair on 1m SOL has ~0.04% maker round trip and relies on the backtest's touch-fill model filling passive quotes. Adverse selection is not modelled, so any positive result is likely flattering. The hypothesis itself calls this the fee-graveyard class. - [critical] Gross exposure goes over equity at leverage 1.0, and the account was liquidated. In the sandbox, avg_position_pct is 252% with config leverage 1.0, and the account hit liquidated=true on 2026-07-27 with total_return -100%. `cash = equity*lev*(1-buffer) - coins*close` is recomputed after every fill from the venue balance. Nothing in it caps gross notional independently: no check against notional already committed or resting, and no hard ceiling of gross <= equity x leverage. A long-only 1x book with this design cannot be liquidated unless exposure has gone above 100%. That makes this a sizing/accounting defect, not an edge outcome. (line 131) — fix: Cap the BUY quantity so (existing long qty + resting buy qty) x price <= equity x leverage x (1 - buffer), using mark-to-market equity rather than the raw wallet balance. Assert the cap before every submit. Before posting the replacement bid, confirm the cancel of the previous bid was processed (no open BUY in cache.orders_open) so two bids cannot both fill. - [critical] The impact/capacity limit is blown by sizing. total_impact_usd is $624,541 across 2,311 trades, about $270 per trade. That is almost exactly the average loss of $298.58, and it is why the win rate is exactly 0% on a design that sells 20bp above where it buys. Every quote commits 100% (in practice ~250%) of equity to a single 1-minute SOL fill. The modelled market impact of that clip is far larger than the 20bp quote width, so each round trip is structurally negative before any signal question comes up. Per the fail_class rules, a capacity limit blown by position sizing is a code_defect. (line 142) — fix: Clip each quote to a capacity-aware fraction of equity. Size it from recent 1m volume, e.g. <= a small % of bar notional volume, so modelled impact stays well under the 10bp half-width. Report the resulting per-trade impact in the rationale. - [warning] The trade ledger does not reconcile. avg_trade_return_pct is -0.0000333% while avg_trade_pnl_usd is -$298.58 and total_return is -100% (lessons L140/L116). Per-trade economics cannot be compared to the 0.04% maker round trip until the denominator is the notional actually deployed. Also, trades_count (6,040) differs from total_trades (2,311). (line 139) — fix: After fixing sizing, check that the ledger books each fill against its real notional before any fee-floor comparison. - [warning] The requote cadence is calendar-anchored (bar ts minute % 5), which is correct. But drift requotes need >=10bp of movement and happen only every 5 minutes. The original cancels and replaces every loop, so quotes can sit up to one full quote-width stale. This is a documented compromise for wall-clock reasons, so it is only a warning. (line 97) — fix: Once sizing is fixed and the order count drops, consider a tighter tolerance (e.g. 5bp) on the calendar grid. - [warning] min_bars_required override shadows the base method. It is harmless here because there are no indicators, but it is flagged by static analysis. (line 51) — fix: Keep it only if the base contract allows it.

Verification failed (Layer 4 — QA review) [class=code_defect]: - [edge_concern] Even with sizing fixed, a 20bp-wide quote pair on 1m SOL has ~0.04% maker round trip and relies on the backtest's touch-fill model filling passive quotes. Adverse selection is not modelled, so any positive result is likely flattering. The hypothesis itself calls this the fee-graveyard class. - [critical] Gross exposure goes over equity at leverage 1.0, and the account was liquidated. In the sandbox, avg_position_pct is 252% with config leverage 1.0, and the account hit liquidated=true on 2026-07-27 with total_return -100%. `cash = equity*lev*(1-buffer) - coins*close` is recomputed after every fill from the venue balance. Nothing in it caps gross notional independently: no check against notional already committed or resting, and no hard ceiling of gross <= equity x leverage. A long-only 1x book with this design cannot be liquidated unless exposure has gone above 100%. That makes this a sizing/accounting defect, not an edge outcome. (line 131) — fix: Cap the BUY quantity so (existing long qty + resting buy qty) x price <= equity x leverage x (1 - buffer), using mark-to-market equity rather than the raw wallet balance. Assert the cap before every submit. Before posting the replacement bid, confirm the cancel of the previous bid was processed (no open BUY in cache.orders_open) so two bids cannot both fill. - [critical] The impact/capacity limit is blown by sizing. total_impact_usd is $624,541 across 2,311 trades, about $270 per trade. That is almost exactly the average loss of $298.58, and it is why the win rate is exactly 0% on a design that sells 20bp above where it buys. Every quote commits 100% (in practice ~250%) of equity to a single 1-minute SOL fill. The modelled market impact of that clip is far larger than the 20bp quote width, so each round trip is structurally negative before any signal question comes up. Per the fail_class rules, a capacity limit blown by position sizing is a code_defect. (line 142) — fix: Clip each quote to a capacity-aware fraction of equity. Size it from recent 1m volume, e.g. <= a small % of bar notional volume, so modelled impact stays well under the 10bp half-width. Report the resulting per-trade impact in the rationale. - [warning] The trade ledger does not reconcile. avg_trade_return_pct is -0.0000333% while avg_trade_pnl_usd is -$298.58 and total_return is -100% (lessons L140/L116). Per-trade economics cannot be compared to the 0.04% maker round trip until the denominator is the notional actually deployed. Also, trades_count (6,040) differs from total_trades (2,311). (line 139) — fix: After fixing sizing, check that the ledger books each fill against its real notional before any fee-floor comparison. - [warning] The requote cadence is calendar-anchored (bar ts minute % 5), which is correct. But drift requotes need >=10bp of movement and happen only every 5 minutes. The original cancels and replaces every loop, so quotes can sit up to one full quote-width stale. This is a documented compromise for wall-clock reasons, so it is only a warning. (line 97) — fix: Once sizing is fixed and the order count drops, consider a tighter tolerance (e.g. 5bp) on the calendar grid. - [warning] min_bars_required override shadows the base method. It is harmless here because there are no indicators, but it is flagged by static analysis. (line 51) — fix: Keep it only if the base contract allows it.

verification_loop: Verification failed (Layer 4 — QA review) [class=code_defect]: - [edge_concern] Even with sizing fixed, a 20bp-wide quote pair on 1m SOL has ~0.04% maker round trip and relies on the backtest's touch-fill model filling passive quotes. Adverse selection is not modelled, so any positive result is likely flattering. The hypothesis itself calls this the fee-graveyard class. - [critical] Gross exposure goes over equity at leverage 1.0, and the account was liquidated. In the sandbox, avg_position_pct is 252% with config leverage 1.0, and the account hit liquidated=true on 2026-07-27 with total_return -100%. `cash = equity*lev*(1-buffer) - coins*close` is recomputed after every fill from the venue balance. Nothing in it caps gross notional independently: no check against notional already committed or resting, and no hard ceiling of gross <= equity x leverage. A long-only 1x book with this design cannot be liquidated unless exposure has gone above 100%. That makes this a sizing/accounting defect, not an edge outcome. (line 131) — fix: Cap the BUY quantity so (existing long qty + resting buy qty) x price <= equity x leverage x (1 - buffer), using mark-to-market equity rather than the raw wallet balance. Assert the cap before every submit. Before posting the replacement bid, confirm the cancel of the previous bid was processed (no open BUY in cache.orders_open) so two bids cannot both fill. - [critical] The impact/capacity limit is blown by sizing. total_impact_usd is $624,541 across 2,311 trades, about $270 per trade. That is almost exactly the average loss of $298.58, and it is why the win rate is exactly 0% on a design that sells 20bp above where it buys. Every quote commits 100% (in practice ~250%) of equity to a single 1-minute SOL fill. The modelled market impact of that clip is far larger than the 20bp quote width, so each round trip is structurally negative before any signal question comes up. Per the fail_class rules, a capacity limit blown by position sizing is a code_defect. (line 142) — fix: Clip each quote to a capacity-aware fraction of equity. Size it from recent 1m volume, e.g. <= a small % of bar notional volume, so modelled impact stays well under the 10bp half-width. Report the resulting per-trade impact in the rationale. - [warning] The trade ledger does not reconcile. avg_trade_return_pct is -0.0000333% while avg_trade_pnl_usd is -$298.58 and total_return is -100% (lessons L140/L116). Per-trade economics cannot be compared to the 0.04% maker round trip until the denominator is the notional actually deployed. Also, trades_count (6,040) differs from total_trades (2,311). (line 139) — fix: After fixing sizing, check that the ledger books each fill against its real notional before any fee-floor comparison. - [warning] The requote cadence is calendar-anchored (bar ts minute % 5), which is correct. But drift requotes need >=10bp of movement and happen only every 5 minutes. The original cancels and replaces every loop, so quotes can sit up to one full quote-width stale. This is a documented compromise for wall-clock reasons, so it is only a warning. (line 97) — fix: Once sizing is fixed and the order count drops, consider a tighter tolerance (e.g. 5bp) on the calendar grid. - [warning] min_bars_required override shadows the base method. It is harmless here because there are no indicators, but it is flagged by static analysis. (line 51) — fix: Keep it only if the base contract allows it.