Yesterday's autopsy covered the mechanism that killed the v4 market-maker. This is the wider view: four months from the first prediction to the final shelf, two completely different strategies, and the question of whether the whole thing was worth doing. The short version — it was, but not for the reason we started it.
The arc, in numbers
| Phase | Dates | The question | Result | Real $ |
|---|---|---|---|---|
| v3.x directional | Feb–May 15 | Can an AI predict Polymarket? | −$582 · 6,267 trades · 25.7% hit | $0 |
| v4 Phase A | May 16–20 | Does combined-cost arb work? | +$66 hypothetical | $0 |
| v4 Phase B / B.2 | May 21–Jun 7 | Does it survive realism? | +$149 → −$132 | $0 |
| v4 Reframe (#017) | Jun 7–29 | Does a slow maker win? | −$405 · 0 pairs / 8 fills | $0 |
| Shelved (#018) | Jun 29 | — | Structurally −EV at our cadence | $0 |
The one chart that explains everything: the degradation cascade
The single most important artifact of the whole experiment is three numbers for the same v4 strategy — each one a layer of optimism stripped away:
+$814 theoretical edge every quote fills, perfectly +$149 hypothetical we're always top-of-book, only real taker volume counts (18% of theoretical) −$405 actual paper queue drainage + one-sided resolution + hedge fees + day-loss
theoretical_vs_hypothetical conversion was 18% — of every dollar of
edge we could identify, fills based on actual taker flow captured only eighteen cents, before
the final layer of realism dragged even that under water. If your P&L only survives at the most
optimistic layer, you don't have an edge — you have a spreadsheet.
Crucially, our queue stats showed we were top-of-book 100% of the time. This was never a latency or queue-position problem we could engineer around. The fill itself was the bad news: a passive maker in a slow directional market only gets filled on the leg the market is about to move against. (The mechanism, in full, is in the autopsy.)
What we learned
About the strategy
- Both strategies died of the same disease: no informational edge. v3.x: every signal we used was public, so we converged on the market's own price. v4.0: "market-neutral arbitrage" sounds edge-free, but completing the pair means winning a fill race you can only win when you're on the losing side. Different mechanism, identical root.
- Adverse selection is not a tunable parameter. No confidence floor, rebate constant, or market category fixed it, because being filled is itself the adverse signal.
- The edge existed in the data but was uncapturable. 53 of 116 markets had a real sub-$0.95 combined cost. The money was on the table; a once-an-hour retail maker just couldn't pick it up without getting picked off first.
About process and engineering
- Paper trading is asymmetric — and that's the point. It can't prove a strategy wins (phantom fills inflate everything), but it cheaply and definitively proves one loses. We bought a clear "no" for zero real dollars. The system worked exactly as designed.
- Pre-registered gates beat vibes. We wrote the kill criteria and the date in advance and shelved on them a week early instead of rationalizing "one more tweak." The harness even caught a self-deceiving test reading a moving file and calling it green.
- Build the cheapest test of the load-bearing assumption first. The queue-aware reconciliation that finally told the truth arrived as a Phase B.2 addition — after we'd built the state machine, kill switches, and category configs. We engineered the car before checking the road existed. That single realism check should have been week one.
About ourselves (the uncomfortable part)
- The v3.x → v4.0 pivot was infra-driven, not edge-driven. "We built this whole data stack, surely there's a use for it" is motivated reasoning wearing a lab coat. A fresh edge thesis should justify the next build — never the sunk cost of the last one.
- "This is the real shot" fired three times (#014, #016, #017) and missed each time. When every fix is framed as the final fix, the problem is structural, not parametric. The pattern was the signal; we were slow to read it.
- The phantom numbers kept seducing us. $66, then $814. We repeatedly had to rebuild realism just to debunk our own optimism — a tell that the dashboard was quietly selling hope.
Pros and cons, plainly
What went right
- $0 real capital risked across two dead strategies and four months.
- Genuine engineering rigor — atomic writes, queue-aware sim, kill switches, a test harness, 155 tests.
- Built in public, losses published. The autopsy is credibility content most shops would bury.
- A clean kill on pre-registered criteria — no zombie project limping along.
- Reusable infra (CLOB reader, order book, funding/whale/smart-money feeds) survives for any future Polymarket work.
- A transferable lesson bigger than the project: the degradation cascade as a test for any edge claim.
What it cost
- ~4.5 months for a negative result that was partly knowable at v3.x's death in May.
- Six weeks re-proving "no edge" in a second domain because the data infra was too nice to throw away.
- Breadth as avoidance — 5 coins, then weather, then 5 market categories — instead of answering the one hard question.
- Opportunity cost is the real bill. Real dollars: zero. Engineering hours that could have gone to revenue: not zero.
The real reason we shelved (it isn't "impossible")
So PolyDoge is shelved by choice, with the door left open. If and when the operator decides to commit real focus to sub-minute market-making — the cadence where the edge is actually capturable — the data stack, the harness, and four months of hard-won lessons are sitting right here, ready. Until then, chasing a few dollars an hour at the wrong cadence is exactly the kind of distraction a small team can't afford.
The bottom line
v3.x asked whether an AI could predict Polymarket. No — the signals were already in the price. v4.0 asked whether it could make markets without predicting. Also no — not at this cadence, because you can't redeem half a pair. The most valuable thing PolyDoge produced wasn't a profit; it was two definitive, mechanism-level noes for zero real money, a reusable test for spotting phantom edge, and the clarity to walk away on purpose. That's a good trade.