How the scenarios are authored

The cascade describes possibilities, not predictions. Putting a precise number — “Taiwan flashpoint: 25%” — on the public surface implies a calibration claim we cannot honestly back with twelve scenarios per year of data.

We do still record what we believe internally. Every scenario carries a dated, hash-chained probability commitment in our forecast ledger (the same ledger that powers /forecasts). The public surface presents these commitments qualitatively — most-likely path, most-impactful path — while the ledger holds the record we can be measured against later, when there is enough data to measure.

The qualitative tags on cascade pages are derived mechanically from the ledger’s most-recent snapshot. They cannot drift from what we have committed to internally.

“The record we hold ourselves to” — see /forecasts.

The cascade tree is a structured set of authored AI hypotheses about how a triggering event could play out. Each scenario (X) branches into a small number of conditional paths (Y), and each Y can branch again into more specific outcomes (Z).

X is the trigger. Y is "given X, then what?". Z is "given X and Y, then what?". The siblings at each level are intended as an exhaustive partition of the conditional space, summing to 100%.

These shapes are possibilities, not predictions. We're sketching, watching, holding our priors lightly. The platform claims possibility — not prediction.

These are hypotheses, not predictions. The cascade is a framework for thinking about asymmetric exposure across paths — not a ranked forecast feed.

Every probability you see is authored as an honest base-rate-anchored estimate, not a market-calibrated forecast. The AI consults available evidence — historical base rates, public commentary, OSINT signal patterns — and writes a number it can defend.

All probabilities are multiples of 5%. We never author 22% or 47%. The underlying authoring is not precise enough to justify finer units, and a 5% grid keeps siblings legible (they sum to 100%) and forces honesty about uncertainty.

Probabilities are re-elicited per horizon when the underlying picture shifts. A scenario that read 65% in March might read 50% by May if events undercut its premise. We do not silently update prior estimates — we re-author and refresh the information-cutoff date.

Each branch carries a small set of confirmatorySignalTerms — phrases that, if they show up in recent news, would corroborate the branch. The badge counts how many articles in the last 7 days mention those terms.

Signal counts are not probability shifts. A high count means people are talking about this path right now; it does not mean the path is more likely to materialise. Volume-driven probability mutation was rejected unanimously by the review council (PR #134) — media events get amplified by attention, not by underlying base rates.

The signal numbers are a corroboration trail, not a forecast. Use them to ask "does the public information line up with this hypothesis?", not "should I bet on this branch?".

On Y-branch detail pages, the "Cost of inaction" panel shows AI-hypothesis estimates of how named markets, tickers, or sectors might move IF the branch materialises. Every estimate carries a confidence tag — almost all are tagged low. We surface the tag prominently so readers anchor on the epistemic state, not on the number.

These are not forecasts. They are conditional illustrative estimates — useful for asking "what's my exposure if this path plays out?", not for trade sizing. Magnitudes assume the branch materialises exactly as described; actual moves depend on timing, prior positioning, and intervening events.

Where confidence is low, the magnitude is muted in the UI so the eye lands on the label, not the number. Where confidence is high, the magnitude is emphasised — but high confidence is rare in practice.

Each Y- and Z-branch carries an optional one-sentence Trade lens — an investor-direct note on the named securities or asset classes the branch touches, authored from the winners/losers analysis. It names ≤ 3 instruments, includes a magnitude bucket (small move / meaningful / structural), and a timing tag (fast / slow).

The parallel Policy lens speaks to diplomats, analysts, and journalists — one sentence on the institutional or regulatory response the branch implies, with no portfolio framing. Both lenses are additive to the neutral branch description, not replacements for it.

Neither is a recommendation. Trade and policy lenses are authored hypotheses about what the branch would likely mean for named markets and institutions — not advice on what to do.

Each top-level scenario carries a Falsified byline — a single observation that, if seen in public data, would substantially undermine the scenario's premise.

This is a pre-commitment for analytical hygiene. Stating in advance what would refute a thesis forces intellectual honesty and makes the surface auditable as events unfold. When the falsification condition appears in reality, we re-author the scenario rather than quietly updating its framing post-hoc.

Wildcards are events that the X→Y→Z partition does not capture — outside-the-tree contingencies that could materially shift the picture. They are authored as paired "what if positive" / "what if negative" outcomes to capture asymmetric tail behaviour.

Wildcards deliberately carry no probability. Assigning a number to an off-tree event would overclaim — we would be saying we know how to compare a wildcard's likelihood to the on-tree partition, and we don't.

The only directional cue is a small lean indicator on the side we think is currently more likely to surface. The lean is editorial, not numerical.

GDELT / English-media bias. Our signal counts pull from GDELT and similar English-leaning OSINT feeds. Events that the English-language press underweights — domestic-only stories in non-English markets, lightly-covered regional dynamics — will look quiet here even when they are not.

Low-corroboration risk. A branch can be plausible with very few news mentions — and a branch with many mentions can still be wrong. The signal badge is a media-volume measurement, not a truth measurement.

Sanctioned-country handling. Any country under active comprehensive sanctions surfaces only as news / threat / exposure on this platform — never as an opportunity, posture, or actionable trade. This is a legal rule, not a default.

Information cutoff. Each cascade page shows the date of its latest re-authoring at the bottom. Probabilities and named winners/losers are valid as of that date; events since then have not been folded in.

No market-calibrated mispricing. We do not claim that any probability on this page differs meaningfully from a price implied by options, prediction markets, or other tradable instruments. Our numbers are authored estimates, not arbitrage signals.

No investment advice. Nothing on the cascade pages is advice. The scenarios are an analytical frame to help you think about exposure; sizing and execution are your call, with a qualified advisor.

No forecast accuracy track record yet. We do not yet publish a back-tested or live-tracked accuracy log for these scenarios. When we have one, it will live here and be visible from every cascade page.

Scenario detail pages that correspond to a well-documented geopolitical type (energy shock, EM currency crisis, rate shock, etc.) show a Sector Rotation Map— a compact table of S&P 500 sector ETFs with directional tilts derived from historical analogs of the same scenario type.

Each tilt shows a beta (β)— the historical correlation coefficient between the scenario type materialising and the sector's relative performance. +0.8 indicates the sector tended to outperform in roughly 80% of analogous historical episodes; −0.6 indicates underperformance in roughly 60%. Beta is sourced from post-mortems of named historical episodes: CRSP/Compustat sector proxies for pre-ETF events, and live SPDR ETF returns where available.

Tilts are labelled OW (Overweight) or UW (Underweight) based on the beta sign and a neutral band of ±0.20. These are directional tilts, not portfolio recommendations.

Every map displays an N-of-analogs count — the number of distinct, named historical episodes that underpin the tilt set. When this count is below 5, an amber warning is shown. A low count means the statistical base is thin: the tilt is directionally plausible but not well-established by repetition. A count of 7+ reflects a pattern seen across multiple market cycles.

The sector rotation maps are not shown for every scenario. Only scenario types with a coherent body of historical analog literature are mapped. Scenarios without a well-established type (novel hypotheticals, composite crises, or scenarios that don't map cleanly to a single historical pattern) do not render a sector rotation card rather than show a misleading or thin-analog result.

The cascade scenarios are reviewed by an AI-simulated advisory panel before being published. The panel runs as four independent expert perspectives — an economist, a geopolitical risk analyst, a security expert, and a financial risk advisor — each prompted to review the draft for factual accuracy, framing bias, missed alternatives, and calibration of the probability. A synthesis agent adjudicates disagreements and flags open questions to the human editor.

The council does not replace human judgment. Human review is required before any scenario is published. The panel is a structured red-team, not an autonomous publisher. Council notes are retained in internal version history but are not published alongside the scenarios.

This process was adopted in May 2026. Scenarios authored before that date were reviewed under an earlier informal process and are marked OpenWatch editorial in their attribution line.