Record
The scoreboard first, with the number of runs behind every figure; then every failure, fallback and withdrawn claim.
PAPER n = 4 decisions: far too early to judge anything. No broker is connected: every number here is the paper book, after the declared costs.
The market part is the return before costs; the cost part is every declared fee and spread the paper book has paid since its start (building the book from cash paid most of it).
The paper book is compared with three controls: the SQ-8 mechanical stock rule (on paper), the index held, and a matched index (the same trades' sector and beta). All are base 100 at the start and after the same declared costs.
| Day | Paper book (PAPER) | SQ-8 mechanical rule (PAPER) | Matched index (beta x sector) | Index held |
|---|---|---|---|---|
| 2 Oct 2026 | 98.1 | — | — | — |
| 1 Oct 2026 | 97.5 | — | — | — |
2 of 10 daily points: a chart replaces this table at 10. Not recorded yet: SQ-8 mechanical rule (PAPER), Matched index (beta x sector), Index held. The controls start after enough closed sessions.
PAPER Every swing idea pitched across 4 paper decisions (n = 16 ideas), and how far it really got. These were traced runs: every idea was also pushed through every stage to show the whole flow, but the funnel counts only where each idea really stopped.
Where each idea stopped, coloured by the seat that stopped it:
no_factsno_catalyst_idday2_unconfirmedno_catalyst_idskeptic_failedskeptic_rejectcatalyst_misreadskeptic_rejectskeptic_waitskeptic_waitpm_passpm_passpm_passPAPER Every idea the swing council saw (n = 15 across 4 paper decisions), priced as if it had been bought at its slot, whatever stopped it. Same paper rules for every idea: exits from the next completed session, a gap through the stop books at the open, a session touching both stop and target books the stop, the time stop books the close. Returns are net of the declared 1.25% cost on each leg; open ideas are marked to the close of 1 Oct 2026.
| Group | n | Average net | Median net | Hit rate | Open · resolved |
|---|---|---|---|---|---|
| Skeptic wait, debated (Manager passed) | n = 1 | avg −2.5% | median −2.5% | hit 0% | 3 open · 0 resolved |
| Skeptic said wait | n = 0 | avg — | median — | hit — | 2 open · 0 resolved |
| Skeptic rejected | n = 4 | avg −0.2% | median +1.1% | hit 75% | 4 open · 0 resolved |
| Dropped by code (eligible names) | n = 2 | avg −2.7% | median −2.7% | hit 0% | 5 open · 0 resolved |
| Paper-only setups | n = 0 | avg — | median — | hit — | 0 open · 0 resolved · 1 excluded |
Hit rate: the share of priced ideas above zero after costs. Open ideas count at their mark.
Left out (corporate action): CTVA long · decision #1.
PAPER What the Skeptic said on the ideas it judged (n = 11), and how often its reply could not be used. Small numbers: one idea moves a share by 9 points.
Unusable replies: 1 of 11 Skeptic calls (9%). An unusable reply counts as no verdict: the idea stops there.
| Incident | Opened | Severity | Status | What happened |
|---|---|---|---|---|
| INC-0004 | 30 Sep 2026, 14:40 UTC | medium | resolved | Data and model outage, then two slots not run At the 30 Sep 2026 14:40 UTC paper slot the model provider, the price feed and the SEC feed were all unreachable: the Scout could not run, the movers screen was missing, and the cycle closed with nothing proposed. The 30 Sep 18:40 UTC slot was then not run at all, and the 1 Oct 14:40 UTC slot ran 83 minutes late, because the paper scheduler lived inside an interactive session that was busy at those times. A standalone scheduled job for paper runs now exists (operator-loaded); until it is loaded, a busy session can still miss a slot. |
| INC-0003 | 29 Sep 2026, 14:40 UTC | medium | resolved | Model-provider billing outage: one run without replies, one run late On 29 Sep 2026 the model provider refused every call over a billing problem on the account (seen at about 18:05 UTC). A paper run of the 14:40 UTC slot got no model replies: no idea was reviewed and nothing was decided. The account was fixed by about 19:50 UTC and the 18:40 UTC slot then ran 69 minutes late; no other slot was affected. Since 1 Oct a refused call of this kind stops the run at once with a billing-error flag and alerts the operator, instead of waiting through the outage retries. |
| INC-0002 | 28 Sep 2026, 18:40 UTC | medium | resolved | The Skeptic's model swapped from GLM to DeepSeek The Skeptic ran on GLM, a different model family from the Scout, so that it would not share the Scout's blind spots. On the first paper test runs (replays of the 28 Sep 18:40 UTC slot, made early on 29 Sep) GLM returned a usable reply on only 1 of 3 Skeptic calls (none on the run before): it ignored the structured-output and no-reasoning settings and spent its budget writing prose. On 29 Sep 2026 the Skeptic moved to the Scout's model family (DeepSeek). It still judges blind (claim, catalyst and fact card, never the Scout's thesis), but its independence now rests on its inputs, not on a different model; every run carries a 'same model' flag. |
| INC-0001 | 25 Sep 2026, 14:40 UTC | low | resolved | The council's model was retired by its provider Found during the 25 Sep 2026 rehearsal work (the exact hour of the retirement is not known): the model provider had retired the council's model version, every call was refused, and the outage guard fell back to the mechanical reference, as designed. The council moved to the official successor of the same model family. Agreement and parse rates are measured again from scratch; nothing measured on the old version is carried over. |
Read straight from the published runs: late starts, model replies that could not be used, fallbacks and inputs that were missing. Minor on their own, listed so none is hidden.
Claims from the research that preceded this project which were later withdrawn. They are kept here so they are not repeated.
Most of the result came from a mechanical drawdown gate rather than from the model's judgement, and the test window lay inside the model's training data, so the model may already have known how that period ended. The result is not evidence of forecasting skill.
The interim figures were wrong: funding and trading costs were counted incorrectly. The scoreboard was withdrawn and replaced by the corrected result.
A note recorded shorting as validated; later tests contradicted it. Shorts are therefore allowed only through the council, in downtrends, and only when a leg passes the net-of-cost gate.