Poke the Bear
Every other page in this cluster states things crisply and links here for the asterisk. This is where the hedging lives — the claims that aren't settled, the numbers that don't mean what people think, and the "facts" the whole field takes on faith. It's the most contrarian page in the set, and the most fun. If you've ever watched a vendor benchmark collapse under a second look, you already have the instinct this page runs on: trust, but verify.
A pressure-release valve for the whole cluster
Here's the architectural honesty. Every narrative page wants to say something clean — "Deep Blue was around 2850," "AlphaZero beat Stockfish," "Rybka copied Fruit." But each of those carries an asterisk, and if I hedged inline on every page the prose would turn to mush. So the other pages state things crisply and link here for the caveat. This page absorbs all the "well, actually," which is exactly what makes it the most citable, most expertise-signaling page in the set. Each section below is a clean split: the popular claim, then the messier truth.
Popular claim vs. messy truth
One engine, one rating body, a 130-point swing
If you want the single cleanest proof that engine Elo is soft, it's this — and it's fully receipted. The same engine, on the same rating body (CCRL), takes four different "current" values depending only on which build, index, and snapshot you cite:
| Engine / build | CCRL list & date | Rating |
|---|---|---|
| Rybka 4, 64-bit 4CPU | 40/15 "best versions only" · Jan 8 2026 | 3186 |
| Rybka 4, 32-bit | 40/15 complete list · Feb 28 2026 | 3057 |
| Rybka 3, 64-bit 4CPU | 40/15 complete list · Feb 28 2026 | 3152 |
| Rybka 3 Human, 64-bit 4CPU | 40/15 complete list · Feb 28 2026 | 3142 |
Roughly 130 points of spread on "the Rybka 4 number" alone — from build, index, and measurement date, with no change to the engine's actual play. That's not a rounding error; it's the whole point.
CCRL is recomputed continuously, so every figure above is a dated snapshot, not a fixed fact — by mid-2026 the same builds had already drifted again (Rybka 4 64-bit 4CPU reading around 3139). That the "current" number won't hold still is the argument.
Same query, different server, build, cache state, or measurement day → a different "the number." Nobody sane trusts a benchmark without its configuration, and nobody should trust an engine Elo without its (build, list/index, snapshot date). "Rybka 4 = 3186" is a benchmark ripped from its config — the kind of claim every DBA has watched fall apart the moment someone asks "measured how, on what, when?" The publishing rule for this whole cluster follows from it: no engine rating ships without its receipt.
The unsettled questions — FAQ
Did AlphaZero really beat Stockfish fairly?
The 2017 result came under conditions widely called unfair to Stockfish (old version, 1 GB hash, fixed move time, no book, no tablebases, TPU vs CPU). The fairer 2018 rematch — proper hash and tablebases, 1,000 games — AlphaZero still won (+155 −6 =839). But AlphaZero was never released, so it's unreproducible, which is why it stays partly contested.
Was Rybka really a clone?
The ICGA ruled in 2011 that it derived substantially from Fruit and banned its author. Many accept the finding but consider the process messy, and "inspiration vs. copying" was never cleanly resolved. Officially ruled against; never fully settled in the community.
Why do the same engine's ratings differ between lists?
A rating is a measurement, not a property. It depends on the build, the list and index, the hardware and time control, and the snapshot date. Rybka 4 shows about 3186 on one CCRL index and about 3057 on another — the same engine, ~130 points apart, from configuration alone.
How strong was Deep Blue, really?
Unknown, and unknowable. It had no official rating, played almost no standard rated games, and was dismantled after 1997, so it can never be retested. Any specific number is a guess.
Is comparing engines across eras meaningful at all?
Only loosely. Within a single rating list and era, relative numbers are informative. Across eras — different hardware, time controls, tablebases, books, and list inflation — the comparison is closer to storytelling than measurement.
Every claim here traces to a real primary source
This is a page about not taking claims on faith, so it can't take shortcuts itself. The material here traces to primary sources — the ICGA report, DeepMind's AlphaZero paper, Stockfish and ChessBase's own legal posts, the CCRL lists, and the Chessprogramming Wiki — not to "as everyone knows." Where a claim is genuinely first-hand or genuinely disputed, it's labeled as such. That's the trust-but-verify loop this whole cluster is built on, running on itself.
- Deep Blue — IBM Research history; contemporaneous 1997 match coverage; Kasparov's public statements.
- AlphaZero vs Stockfish — DeepMind (Science, 2018); Stockfish developer commentary (Tord Romstad); the 2018 rematch results.
- Rybka / ICGA — the ICGA panel report; Chessprogramming Wiki, "Rybka Controversy."
- Rybka rating swing — CCRL 40/15 "best versions only" (Jan 8 2026) and complete list (Feb 28 2026), computerchess.org.uk; figures cross-checked by the author.
- "Fritz" brand lineage — Chessprogramming Wiki (Fritz) and ChessBase release history.
Back to the stories these asterisks belong to
Now that the caveats live here, the narrative pages get to state things cleanly. Read the ones that point here.