Methodology

semii.ai is an automated research experiment. This page describes, at a high level, how it forms its views and — more importantly — the discipline that keeps its track record from lying to you. The honesty is the product; if the method below is not convincing, nothing else on the site should be either.

How ingestion works

An automated pipeline reads the public record on the semiconductor supply chain, digests each source into our own prose, and updates a structured board of stances on the drivers we track — packaging, foundry capacity, memory, interconnect, power, and so on. Each stance carries a direction (-2..+2 versus a benchmark), a conviction, and a short rationale linked to the source that moved it. Related layers — a longer-form worldview, a mechanism reference, and a register of dated predictions — are maintained the same way.

The underlying source corpus, the prompts, and the internal exposure maps never leave the system. What the site publishes is the derived layer only, in our own words. The free tier shows the board on a 30-day delay, the track record, and a calibration strip; the live board and the deeper browsers are Pro (coming soon).

Live vs. replay: the parameter-leakage problem

The tempting way to show a track record is to replay history: take today’s method, run it over the last few years of data, and chart the returns. That number is almost always a lie, because the method was built knowinghow that history turned out. Thresholds, driver definitions, and framings are all quietly tuned — consciously or not — to the period you are now “testing” against. This is parameter leakage, a form of hindsight bias, and it makes a backtest look far better than anything achievable in real time.

So we split the record in two and never average them together:

• Live states were recorded on or after the cutoff, before their outcome window closed. Nothing about the future could have leaked into them. These are the real claim, shown solid and green.
• Replay states predate the cutoff. They are reconstructions of what the board would have said, and are structurally exposed to the leakage above. We show them — dashed, muted, and explicitly labeled “hindsight-contaminated, not a performance claim” — only so the method is auditable, never as evidence of skill.

If the live record is empty, we say so plainly and show nothing in its place. We do not borrow a replay figure to fill the headline. The live record begins accruing at the cutoff and grows one committed state at a time from there.

Git-commit timestamping

What makes a “live” state trustworthy is that its timestamp is not self-reported. Every change to the board is a git commit in a version-controlled history. The commit records, cryptographically chained to every commit before it, exactly when a stance was taken — and that record is fixed before the market resolves it. A claim cannot be back-dated without rewriting the entire chain that follows it.

That is the entire basis for calling a point “committed before its outcome”: the clock is the git history, not our word. The track record is exported from those commits, so what you see is bounded by what the history can prove.

None of this is investment advice, and the positions shown are hypothetical overlays rather than trades. See the full disclaimers, or go back to the track record.