How Evidence Is Chosen — why these entries

The entries in this ledger are not clipped news. Before an event can enter the ledger it has to pass four gates, and no amount of topical heat substitutes for passing them. This page sets out the four gates, why the eight categories are these eight, what is excluded, and where this ledger comes from.

The four gates — selection rules

• (a) Dated and sourced. The item must carry a date and a source that can be named — a filing, a release, a statistical agency, a research institution, or a reputable outlet. Rumour and anonymous relay are excluded.
• (b) Bearing on A versus B. It must be possible to state, in one sentence, which scenario the item bears on and through what mechanism. That is why every entry carries a judgement (Supports / Refutes / Neutral) with its reason.
• (c) Weighted by directness. Material that touches the framework's inputs directly — guidance, filings, official statistics — weighs more than commentary. This is the basis of the 3 · 2 · 1 weights: guidance decisions and filings are 3, notable data is 2, context is 1.
• (d) Falsifiable in principle. Future material must be able to reverse the item's direction. A claim whose interpretation cannot change whatever data arrives is a belief, not evidence, and does not enter the ledger.

The eight categories, and why these eight

Frontier Models

The capability curve itself — the engine of the A/B race. What matters in a release entry is not “a new model appeared” but what moved, and by how much, against the previous generation. One entry in this ledger records a coding benchmark lead alongside a composite-index lag, side by side: a launch has no direction of its own; the composition of its numbers sets the direction.

Benchmarks

The rulers that measure capability. If a ruler is bent, the whole curve reads wrong, so benchmark entries record the score together with the ruler's known failure modes. The METR time-horizon audit is the standing example: the widely cited 50% horizon turned out to be the output of a fitted logistic curve rather than a direct measurement — an audit that did not dispute the exponential trend itself. An InnovationEval entry records self-reported results collapsing from 43% to 2% after correction. Benchmark evidence is therefore read one notch less confidently than the other categories.

Capital Expenditure

The hyperscalers' spending plans and guidance — the only category wired directly into this framework's circuit breaker. Guidance is a company's intention fixed in numbers, so it carries a different weight from commentary and is usually recorded at weight 3. But aggregate totals can move on accounting changes alone (one company's move from $190B to $175B was explained as exactly that), so what this category watches is not the size of the total but the direction and wording of guidance. The representative entries are Meta's guidance floor raised from $130-145B to $135-145B, and the Big Four aggregate held at roughly $732.5B.

Compute

The supply curve of chips and equipment — whether demand is being fixed in actual hardware orders rather than in verbal commitments. Entries here include TSMC raising guidance and holding capital expenditure at $60-64B, and ASML's 2027-28 capacity sold out with prepayment-based discussions for 2029-30 orders. Demand confirmed by sell-outs and prepayments is a different quality of evidence from demand announced in a presentation.

Macro & Rates

The rate regime itself. The daily tells — the TIPS real yield with breakevens, gold decoupling, HY OAS — come out of this category. One data-accuracy rule attaches to it: real yields are taken from the official U.S. Treasury figures, not from widgets. On 12 October 2026 a widget figure (about 2.43%) diverged from the official figure (about 2.90-2.91%) and materially skewed a gauge reading. Every rate figure in this ledger is therefore recorded only after being checked against the official release.

Industry Structure

How the money circulates: who finances whom, and where the chain creaks. Credit stress is mostly recorded here — one neocloud company's credit collapse (2031 notes yielding above 16%, CDS above 1,000bp), the tracking of vendor-financing circularity, and material that verifies a narrative in actual results, such as Anthropic's S-1. The cracks against the thesis usually appear in this structure before they appear in any company's announcements.

Forecasts

Expert forecasts and the movement of consensus. A forecast is a thermometer, not evidence in itself — but the direction and speed at which the median moves are evidence. This desk's consensus (median 2030, range 2029-2031) is where two estimates (2029 and 2031) converged over a three-round debate; it is recorded alongside the outside distribution — a strict-definition prediction-market median in early 2033, a large researcher survey's median of 2047 — so that the spread itself stays on record. Definitions move numbers: one lab leader's “2-3 years” rests on a definition spanning the full range of human cognition, and cannot be compared directly with this ledger's deployment-reliability standard.

Energy & Power

The physical-constraint channel: if power runs short, growth stops even with money available. The category starts from a reverse scenario — a power-price surge passing through into structural inflation, tying the Fed's hands and pulling a credit crunch ahead of the refinancing wall — put at 20-25% probability in the 9 October 2026 debate. One auction-price claim behind that entry was marked “unverified” in the source record itself, and it entered this ledger still carrying that mark. Caveats travel with the item; that is a rule of this ledger.

Where this ledger comes from — an honest note

The 30 seed entries were mined from one research desk's dated records: the operator's daily debate conclusions from 9 to 26 October 2026, and the two frameworks adopted along the way (the gold-scenario framework of 10 October and the financing-watch framework of 14 October). Which record each entry came from is kept in the project's provenance table (data/SEED_NOTES.md).

The seed composition is: by category, Benchmarks 7, Macro & Rates 6, Industry Structure 5, Forecasts 4, Capital Expenditure 3, Compute 2, Frontier Models 2, Energy & Power 1; by judgement, Supports 7, Refutes 11, Neutral 12. That composition is not a target ratio. It is the material that desk actually encountered in that window.

So this ledger is a tracked-evidence sample, not a census. It records what one desk looked at every day, and evidence that desk never saw is not in it. The gauge must be read as a ratio inside this sample, not as a ratio of the world's evidence.

What is excluded

• Unsourced social posts and anonymous relay. No date and named source, no entry.
• Price moves without a mechanism. A stock moving some percent in a day is not an entry. Price material enters only when it has an anchor for interpretation — guidance, a filing, a spread.
• Duplicate tellings of the same event. One event, one entry. Anthropic's $60B financing was folded into the Broadcom bottleneck entry; standing it up separately would count one source twice.
• The day's overall verdict itself. A desk conclusion such as “support dominant today” is not an entry. Entries are events and figures that happened in the world.
• Anything that cannot be tagged. If no one-sentence mechanism can be written for either side, it does not go in.

How entries are maintained

New entries come from the daily debate record. The extraction rules are fixed: one dated event is one entry, figures are copied in the source's own notation, and an entry's address (its slug) is never changed once published.

Tags and weights are never revised silently after the fact. If a judgement changes, a dated correction note is left. This ledger's credibility, as its maintainers see it, comes from not erasing mistaken records.

Filtered this way, the ledger grows slowly. That is the intent: not a fast ledger, but one where, when it is wrong, it is possible to trace where it went wrong.