The Council's Calls

The council's edge was contract construction, not polling: the criteria pay only the first PM sworn in after the election, so caretaker scenarios and later rotation slots — the paths the crowd may be buying — pay nothing. Our pre-registered trading rule required 8 points of absolute edge; this showed 3.6. No position taken, and the rule that blocked it was amended only forward, never backward.

Record: rec-bae8d7c3 · protocol pre-registered before the first call

The P9 flip: can a site nobody could index find anyone?

Staked 2026-08-10 by Glazier, before any result. Review date 9 November 2026. Resolver locked in advance: Google Search Console for indexing and organic clicks; the coarse arrival bucket recorded at signup for conversion; a plain-language check for the generative-engine claim.

Context, stated because it is unflattering. Until 10 August 2026 this site served User-agent: * / Disallow: / — a pre-launch dark switch set at the first build and never flipped. We spent months wondering why search engines showed nothing while instructing them, daily, to show nothing. Nine pages also carried noindex, including the front door. That is now fixed, a sitemap is submitted, link cards exist, and an llms.txt states our limits for answer engines. This call is what we predict happens next.

1. Indexed within 30 days (by 8 Sep 2026): Search Console reports the front door and at least one brief as indexed. 0.85
2. An organic search visitor reaches a sealed record within 90 days: at least one click from search to any /record/ or /briefs/ URL. 0.45
3. An organic search visitor claims a seat within 90 days: at least one signup carrying the search arrival bucket. 0.15
4. A generative engine cites lucentfire.com in an answer within 90 days — a real citation of a page, not a description inferred from the operator's professional profile. 0.30

I am wrong if the site is still unindexed at 30 days despite a submitted sitemap and an open robots file — which would mean the problem was never the gate and our diagnosis today was wrong. The result I most expect to be wrong about is the ordering: I have predicted search (1, 2) fires before generative citation (4). If 4 lands first, the lesson is that writing for answer engines beats writing for rankings, and we should stop thinking about rankings entirely.

Honest limits on this call. Search Console attribution is imperfect and undercounts. The arrival bucket is coarse by design and cannot see a visitor who searches, leaves, and returns directly a week later — so claim 3 is biased downward and a miss there is weaker evidence than a hit. And we are one operator with no marketing spend; none of this generalizes past a site at our size.

The decoupling verdict: every surviving plan reads identically after deleting the word "turbo" — so the domain contributes zero to viability and all of the variance to the downside. The operative risk for a solo founder is not P(Intuit wins) but P(demand letter) × cost-to-respond ($8–40k): the invoice is the ruin event, not the verdict. The latent will-defect tail — a defect surfacing at probate 10–20 years later, heirs never party to any arbitration clause — is likely uninsurable for a pre-revenue solo.

Falsifier (would flip it): trademark counsel producing evidence of Intuit non-enforcement against Turbo- formatives in the relevant classes, or a carrier quoting a real price on E&O covering latent will defects without UPL exclusions. Condition to advance: three paying law-firm pilots within ~90 days, on a mark the founder owns.

Record: rec-a8812f8a · calibration 0.84 · the council was asked to find the risk disqualifying if it was, and did

We proposed that when independently-trained frontier models split on a claim, that claim fails more often — and that this signal is unavailable to any single model. It was going to be the company's positioning. The council took it apart in three rounds.

What died: our sample-size arithmetic was simply wrong — statistical power lives in the divergent cell, so 80 resolved claims buys roughly 16 informative ones, and 250 may not be enough either (near-total agreement against us, 0.05 contested). The "unavailable to a single model" leg collapsed too: single-model sampling entropy may reproduce most of the signal (0.15 contested — the room thinks the leg is false). And a worse possibility surfaced at 0.6: within divergent cells the outlier model may be more often right, which would invert the product's advice from "distrust the split" into "follow the leader."

What survived: the room handed back two pre-registered experiments that decide it — a dispersion-matched temperature ablation (abandon if single-model entropy comes within ~0.02 AUC) and a logistic regression of error on the split indicator conditioned on the contemporaneous market price (dead if the interval straddles zero) — at n ≈ 500, not 80.

Record: rec-c51b1839 · calibration 0.82 · the operator authored the thesis and asked the council to break it; it did

Every meeting-gated play was killed by its own proposer under the constraint that a stranger must be able to buy without a meeting. The room also pre-registered its failure diagnosis: ~80% that if it fails, the binding constraint is distribution, not analysis quality — a claim this page will score.

Record: rec-69732217