How it works
semii.ai reads semiconductor research, writes down dated views and forecasts, and later measures them against what happened. It is impersonal: within each access tier, every reader gets the same research. Every page carries the version stamp of the record it was made from.
From research to the Board
An automated process reads semiconductor research, records source-grounded claims in our own words, and updates the Board: model views on themes such as chip packaging, memory, foundry capacity, networking, and datacenter power.
Each theme has a score from −2 to +2, a confidence level, and cited reasoning. Opening a theme explains what a positive score means for that theme. The system also keeps a dated list of forecasts so its confidence can be checked against real outcomes.
Source articles and internal formulas are never published. Only original semii analysis and model outputs appear here.
includes the live Board, track record, forecast scores, and this explanation.
currently adds standardized stock-level model implications relative to SOX and normalized dollar illustrations based on a common reference amount.
How forecasts are checked
A forecast records a statement, its odds and a date to check it by. It stays open until the record includes an outcome. Passing that date does not make it right or wrong.
Open forecasts show the latest recorded odds. Checked rows and the plot show the odds saved with the result. Open a row to see when it was first recorded and any changes to its odds. If that history is unavailable, the row says so.
New checks use a separate resolver. It weighs evidence for and against the statement, with quotations and dates verified against the forecast’s definition and window. A proposed result stays open for at least 14 days and needs two fresh checks, including one at or after that deadline. Disagreement or missing evidence can leave it open. Earlier results remain labelled as checked under the earlier review process.
A supported wrong result is as valid as a supported right one. When a later review changes a result, a public correction keeps both checks. “Last looked” is the date of a completed evidence check; it does not mean the forecast was resolved.
The Brier score averages each forecast’s squared probability error over the time its odds stood, from first recording to the first recorded outcome or the end of its window, whichever came first. Each scored forecast then has equal weight in the overall average. Later corrections use the corrected outcome over that original period.
Forecasts first recorded at or after their window end, or recorded and checked at the same instant, use their original recorded odds. The record labels and counts these separately. Missing or inconsistent history makes a score unavailable; it is excluded from the average and counted. This distinction is separate from live and historical origins.
Why the live record starts small
A long historical test can look impressive: apply today’s method to old research and chart the result. But today’s method was designed after much of that history happened. Knowing the outcomes can affect which themes and rules seem important, even unintentionally. That hindsight can make past results look better than what the system could repeat.
So the site keeps two records separate and never averages them together:
contains Board views saved after launch, with measurement beginning after each decision was recorded. These hypothetical measurements are shown solid and green.
rebuilds earlier Board views after later outcomes were already known. It is shown dashed and grey, clearly labeled as hindsight-built, so you can inspect it — never as evidence of skill.
If the live record is empty, the site says so and shows nothing in its place. It never borrows a historical result to fill the headline. The live record grows one recorded model state at a time.
How the timestamps work
Every Board change is saved as a git commit. Git is software version history: each commit records what changed and links that version to the one before it. Altering an earlier state also changes the later hash-linked history.
The track record is generated from those commits. A model state counts as live because of when it entered version history, not the publication date on the research behind it. The version stamp at the foot of the sidebar of every page — like record ada3c7f — identifies the record used to generate that page. Git provides an exact internal history; an independent public timestamp is a separate verification layer.
The version stamp at the foot of the sidebar of every page identifies the record used to generate that page.
For live measurements, each listing enters at the first scheduled exchange close strictly after the full recording timestamp. A decision recorded exactly at a close waits for the next session. Exchange calendars account for weekends, holidays, early closes and daylight-saving changes. These are scheduled session cutoffs, not a simulation of closing-auction fills.
A one-month study uses 21 later sessions on each listing’s own exchange; the benchmark follows its own calendar. Listings can therefore have different entry and end times. The held-book comparison also applies new decisions at each listing’s eligible close. A decision replaced before that close never becomes a holding in that market.
On September 21, 2026, we replaced publication-date and calendar-date live starts with this rule. The correction record preserves the previous figures and compares both methods on the same price snapshot. Historical replay retains its earlier, explicitly hindsight-contaminated method.
A missing required price makes a result unavailable; it does not move the entry to a more convenient date or silently remove a holding. The later accounting correction records the repaired coverage and the measurement changes described below.
What the research book measures
Each live decision uses the mapping, universe and sizing rules saved with that decision. Later edits cannot change its earlier targets. Changes to the model mapping or sizing also enter the held book; prose edits and unchanged targets do not trigger a rebalance or renew a holding’s expiry. Historical replay retains its older method and remains a diagnostic.
The hypothetical book starts with $100,000 and holds fixed adjusted-price units between decisions. At a rebalance, targets are sized against that same reference amount. Positions expire after the recorded holding horizon, with exit costs included; the last snapshot values open holdings without forcing a sale. Gross exposure, including the benchmark hedge, can exceed 100%, so this is not an unleveraged portfolio claim.
Live semiconductor comparisons use adjusted SOXX prices and broad-market comparisons use adjusted SPY prices. These are ETF proxies, distinct from the SOX and S&P 500 price indices used by the legacy replay. Adjusted prices model reinvested distributions. The hedge offsets the model’s net target exposure and follows its own eligible US close.
Trading costs are 10 basis points of actual traded value, including hedge and expiry trades. Debit financing is 5% annually on negative available cash after segregating short proceeds; short borrow is 3% annually on short notional, including the hedge. Both accrue over elapsed calendar time on a 365-day basis. Cash and short collateral earn no interest. These fixed scenario assumptions were introduced in the September 21 correction; they are not recorded broker terms or evidence that every short could have been borrowed.
Foreign prices use the latest daily FX observation available before the valuation instant. An FX row becomes eligible only after the following UTC midnight and expires after seven days. Local market holidays carry the previous scheduled native close while FX can still change. This avoids later same-day FX information, but it is a daily-data approximation, not an observed intraday conversion. Missing scheduled equity prices remain unavailable.
Cash, holdings and costs reconcile to daily net asset value. Maximum drawdown measures the decline from a prior daily net peak, including losses that recovered between decisions. The separate per-state charts are gross studies before costs and are not a cumulative portfolio curve. These corrected hypothetical measurements do not establish investment skill.