Skip to content

v0.8: The Grammar of Simulation

Latest
Latest

Choose a tag to compare

View all tags
@yoheinakajima yoheinakajima released this 04 Nov 22:17
· 2 commits to main since this release
v0.8
6e79a46
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

The Grammar of Simulation, version 0.8 (2026-11-04). This version answers the sixth outside review with new analysis rather than caveats, and classifies the runtime techniques' determinism by written tests.

  • Dataset: sha256 f9ac0d88ac86202e2f48efd28caea7807f09bbde00cfb9ae99b2bea5ab38106e. Twenty technique records changed determinism class.
  • Read it: at https://simgrammar.com/paper/, or as the attached PDF (127 pages, up from 119).
  • Cite it: use CITATION.cff with the dataset hash. The concept DOI 10.5281/zenodo.23846209 resolves to the newest version.

Determinism by written tests (#89):

  • Each of the 35 techniques that can run during play is checked against four tests, asked in order:
    1. Does its output enter simulation state?
    2. Does it give the same output on one machine from the same seed and inputs?
    3. If not, can the developer make it repeat by fixing a seed, weights it runs, or batch and kernel settings?
    4. Is the output identical across machines?
  • The answers are in report/audits/determinism-tests.yaml, and the build fails when any record disagrees with them.
  • Counts go from 9 / 16 / 10 (preserves / mitigable / breaks) to 19 / 11 / 5.
    • Fixed-weight networks and iteration-counted search now preserve.
    • Open-weight world models and language characters on weights the developer controls are mitigable.
    • Agents on closed hosted models break.
  • Lockstep is the tighter limit. Only 12 of the 35 could run in lockstep without one machine streaming its outputs to the others, because GPU inference need not match across machines. The abstract's Finding 5 and the site's home strip now lead with this.
  • Agreement: two independent codings agree on 31 of 35 (κ = 0.81). An earlier blind recode under v0.7's wording had agreed with the old classes on only 14.
  • Figures 11 and 12 relabel the middle class "reproducible once seed and weights are fixed". §6.4 adds a table that names the test deciding each class, and the appendix catalog gains a decided-by column.

The sixth review's other issues (#88):

  • Dating check: dating every early-access game by its 1.0 release, and every update by its year, leaves the post-2010 count at 33.
  • Mechanic pairs: coverage over pairs from different families, where no pair is exclusive by design, is 27.4% against a null of 36.0%. This widens the gap.
  • Rater agreement: with a one-anchor offset removed, tuning-difficulty agreement goes from α 0.09 to 0.84. The raters agree on order, not level.
  • Ranking: under a maturity cap on feasibility, the why-query stays first, and the three ideas §9.4 recommends starting with stay in the top 15.
  • Prior-art recall: a known-item check found that the logged searches retrieved 9 of 17 shipped titles that do the feature. Five new finds are listed, and none lowers a novelty score.
  • Predictions: each prediction says whether a hit would show spread or first arrival, and a new table lists the prior titles whose updates and mods don't count.
  • Smaller fixes: each review bar has a stated reason, the sample-size simulation includes a rater shift, and the contents fit on page 1.

No opportunity score or stored rank changed.

🤖 Generated with Claude Code

Assets 3
Loading