PSDG

Appearance

Research pitch ​

Use the block below in email or DM when you want a one-screen intro. Swap in your name and a single sentence on why this might fit their work (evals, robustness, alignment, game theory, etc.).

Canonical links (site is static; paste absolute URLs — use paths without *.html to match the live host and avoid 308 redirects):

Suggested subject lines (pick one or shorten) ​

  • PSDG: exact solver can lose to a worse opponent — rules decouple placement from possession
  • PSDG — small exact game, oracle benchmarks (static vs re-solving)
  • Compact benchmark: when “optimal” deployment still loses to a late blunder
  • A checkable toy for proxy misspecification + deployment fragility

Email body (copy from here) ​

Copy from “I’m sharing…” through “[Your name]” (include the quick links). You may omit the Optional tighter thesis block if you want a shorter note. Stop at the horizontal rule after your sign-off—the material below that rule is not part of the email.

Optional one-line dek (first sentence or subject-adjacent hook; buys patience before the longer “result” line):

PSDG — Philosopher's Stone Dice Game. An exact solver can lose to a worse opponent — not from noise, randomness or hidden information, but because the rules decouple placement from possession. The act of choosing is long over by the time its consequences are realized. PSDG is the smallest game where this gap is exactly measurable.

I’m sharing a compact research environment you might find useful: PSDG (Philosopher’s Stone Dice Game)—a perfect-information, deterministic-after-setup tabletop game with a published exact solver and reproducible seeded benchmarks.

Result in one line: In a standard protocol, A can follow a line derived from the oracle and still lose to B roughly 6–9% of the time when B makes a single legal but suboptimal move on the last draft twist—not because of noise or hidden information, but because of irrevocable commitments and what gets re-optimized at the Exchange (static principal-line Gift vs re-solving on the realized crucibles; sequential vs simultaneous timing moves the headline rate). The home page table pins the counts.

Why it might be useful: PSDG is a small, checkable place where proxy misspecification, latent rule structure, and deployment brittleness can all be stressed against oracle truth—without hand-waving about scale.

Optional tighter thesis (one sentence; paste after the line above—or use instead of it—if you want a single dense hook; drop it if the email already feels long):

PSDG simultaneously illustrates proxy misspecification, latent commitment structure, deployment brittleness, and the structural limits of human oversight — all within a single tiny, exactly solved, empirically benchmarked game where no claim rests on interpretation because every measurement reduces to oracle truth.

Quick links

[One sentence: how this connects to their papers or agenda.]

Best,
[Your name]

Reminder for the sender (not for the email) ​

The psdg repository on GitHub is public—direct links to blob/main/... and git clone should resolve for recipients. If an institutional firewall blocks github.com, fall back to Home — Solver, benchmarks, and GitHub for context and paths.

The ~6–9% band is the static principal-line rows in the standard six-dice blunder suite; ~5.7% is re-solving at the Exchange—a different protocol. All figures are embedding-specific. If someone pushes back, sharper scoped language lives on AI safety and FAQ.