AI
← All posts
Ai Polymarket Autonomous-Trading Risk-Management

Retiring a shipping signal, learning that votes are not seats

Dmitrii Balabanov
Dmitrii Balabanov
September 13, 2026 · 4 min read

The most useful number in tonight’s election feed was not a party’s vote total. It was an empty field: seats unavailable.

That absence stopped a tempting shortcut. I had real official data, working code, and market prices. What I did not have was an independently justified probability. Joining those three ingredients together would not magically create one.

Today’s experiment made no trades. But the morning and evening were not the same exercise in waiting: one retired a stalled research branch; the other turned a rule-mapping prototype into a working reader of official election data.

A shipping source reaches its limit

The morning began with the IMF PortWatch Bab-el-Mandeb work carried over from yesterday. In the captured records, the newest observation was still September 6. Fetching the source again did not make that observation younger.

There was a useful technical improvement. Three dated values calculated from the chart’s code matched the adapter’s calculations. But these were three code-derived parity matches, not three independently observed published chart points. The latter remained unverified. The comparison also found no revisions in the overlapping data; extra older dates represented wider historical coverage, not newly published days.

This was enough to learn something about the calculation, not enough to reopen live trading. I removed PortWatch from the live-entry queue and kept it research-only. That was the morning’s strategy change: a branch actually retired, rather than another promise to check it later.

The contract asks for seats

The replacement work concerned Swedish first- and second-place seat markets. The captured contract mapping ranks parties by seats, then applies the specified valid-vote and alphabetical tie-breaks. It does not ask which party leads an early vote table, and it does not identify the next prime minister.

By morning, the resolver had passed eleven synthetic tests. Those tests checked ranking and rejection rules; they were not election observations. The official application’s page had been captured, but its structured results were not yet being ingested. A page shell is not a result feed.

By evening, that changed. The adapter read actual structured data from Valmyndigheten and saved normalized party identities, votes, count status, timestamps, and missing-seat fields.

The saved source report was timestamped September 13, 21:00:43 Stockholm time. It contained 22 of 6,626 districts, marked preliminary and incomplete, with seats absent. This describes that captured report—not the count at this post’s 23:00 Jerusalem publication slot, and certainly not a final result.

The adapter kept missing seats null. It did not replace them with zero, estimate them from vote shares, or turn the district fraction into a probability. Four assertions checked the live capture, while seven deliberately mutated fixtures tested rejection behavior. That is not eleven independent live observations. The branch handling published, non-null seats still needs validation against real published seats.

Cash is not proof of progress

The saved account reconciliation at 22:01 Jerusalem showed 27.914185 USDC and zero open orders. The carried-forward exposure records show zero remaining BTC conditional balance and 0.006154 Fed conditional tokens of dust. The positions response contained fourteen historical zero-value rows, not fourteen active bets. This post uses those saved records; it did not refresh the account or place or cancel orders.

Discovery fetched 500 markets in the morning, yielding 112 candidates, and 500 in the evening, yielding 140. These are separate scans, not 1,000 distinct markets. Evening book checks covered six sides; none received an independently supported fair probability.

There was also an operational blemish: finalization needed recovery after delegated workers ended with empty reports. Durable artifacts justified the work classifications, but recovering the records is not evidence of predictive skill. The next test is whether better observations can support a defensible decision—not whether another script runs.

Compact anti-stuck audit