v0.8: The Grammar of Simulation

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

v0.7: The Grammar of Simulation

The Grammar of Simulation, version 0.7 (2026-11-03). This version presents the same results as v0.6, rewritten to lead with each finding and to state each limit once.

• Dataset: sha256 dc951a62c44ca9786ba5286ba421d7b1ac8447981c43bf3c82b03c381172bb7a, unchanged from v0.6.
• Read it: at https://simgrammar.com/paper/, or as the attached PDF (119 pages, down from 124).
• Cite it: CITATION.cff, with the dataset hash. The concept DOI 10.5281/zenodo.23846209 resolves to the newest version.

What changed since v0.6 (#86):

• Abstract: cut to about 500 words. Six findings, each with its key number, and one Reliability paragraph on cross-model agreement.
• Body:
  • Every section leads with its finding.
  • Robustness checks are reported as what was tested and what held.
  • A check that failed is stated once, where the finding is made.
• Limits: §2.6 is now five short limits: sample, raters, agreement, relations, and coverage and recency. The full detail is in a new Appendix F (Reliability): ratings and calibration, the second rater, the rated mechanics, the coding sample, the cross-model check, and the rating sample-size simulation.
• Legibility (§5.6): the correlation comes first (ρ = −0.42), then the checks it survives, then the second model's weaker reading (ρ = −0.21). The section now says plainly that legibility tracks variety, not the number of couplings.
• Corrections:
  • Sentences that the cross-model run had made out of date now report what it found.
  • The site's home strip now matches §6.4: the ten breaking techniques lose lockstep play and re-simulation, and logging their outputs restores replay.
• Book: the same rule applies across all chapters, and chapter 13 is the single home for the book's limits.

No record, score, statistic, or stored rank changed.

🤖 Generated with Claude Code

v0.6: The Grammar of Simulation

The Grammar of Simulation, version 0.6 (2026-11-03). This version answers Explore Science's review of v0.5.1. That review scored the paper 93 (Platinum) and raised 47 issues, all minor.

• Dataset: sha256 dc951a62c44ca9786ba5286ba421d7b1ac8447981c43bf3c82b03c381172bb7a. It changed only by two new source records.
• Read it: at https://simgrammar.com/paper/, or as the attached PDF.
• Cite it: CITATION.cff, with the dataset hash. The concept DOI 10.5281/zenodo.23846209 resolves to the newest version.

What changed since v0.5.1 (#84, #85):

• Subtitle: now "…the mathematical, AI, ML, and LLM results that could change them". An affects edge is an asserted potential impact, not a demonstrated change.
• Opportunity ranking:
  • The ranking table shows ties as places, for example "2 to 5" and "8 to 14", instead of tiers.
  • A new table rebuilds the top 8 under equal weights, a mean of ranks, random weights, two feasibility caps, and a whole-scale rater shift. The why-query stays first under every weighting but falls to second under a cost cap: any opportunity that calls a model during play with no stated budget is capped at feasibility 3.
  • The cross-model table flags the 18 opportunities whose two readings are more than 10 places apart.
  • §9.4 now starts from the five opportunities that both readings put in their top 15.
• New checks, some of them negative results:
  • The coupling recorded between mechanics does not track legibility (ρ = −0.07).
  • The second model rejected only 2 of 12 proposed negatives at admission.
  • With nothing named in advance, 11 of 19 clusters keep their five most central games together.
  • 38 of 42 predictions would show a feature spreading rather than arriving, including all 9 due in 2027.
• Predictions (§11): a rule for scoring the whole set, fixed before any prediction resolves. Only technique tests confirmed with the thesis evidence count. The thesis is supported if half or more confirm, and weakened below a quarter.
• Reporting:
  • Monte Carlo p-values print as "≤" at their floor.
  • The curveball null is cited and defined, and Appendix E defines every null.
  • The determinism classes, the ceiling frames, and the post-2010 count (now a range, 33 to 44) are worded to what the data show.
• Figures:
  • Figure 10 has a size key.
  • Figure 14 highlights all opportunities tied at the eighth-place priority.
  • Figures 4, 6, and 9 have clearer labels.
  • The PDF bookmarks no longer show layout markup.

No record, score, or stored rank changed. Still open, and now listed in §11.1:

• agreement with a second person;
• recoding the determinism classes against written tests;
• an audience bar for the console clause;
• error-rate checks of game tags and of the prior-art searches.

🤖 Generated with Claude Code

v0.5.1: The Grammar of Simulation

The Grammar of Simulation, version 0.5.1 (2026-11-02). This patch answers what Explore Science's fourth review left open on finding E31, the ranking of opportunities. That review addressed 67 of 68 findings and scored the paper 90.

• Dataset: sha256 3c96f240392d1eb3d49c3d5ca85d751283aba26a082850481fc2b460ed745d69, unchanged from v0.5.
• Read it: at https://simgrammar.com/paper/, or as the attached PDF.
• Cite it: CITATION.cff, with the dataset hash. The concept DOI 10.5281/zenodo.23846209 resolves to the newest version.

What changed since v0.5 (#83):

• The second model's ranking is in §9.1. The priority table has a cross-model rank column. A new table gives all 42 opportunities' places under three readings: the original rater's, the second model's, and their lower median.
• Ties are counted, not broken by id. On the second model's scores, the why-query shares one priority with eight other opportunities. v0.5 reported it as 27th, but that place came from alphabetical order. It is tied for places 26 to 34, and the abstract, §9.1 and book chapter 10 now say so.
  • On the lower median, the why-query is tied for places 15 to 23, and language modding is first on its own.
  • Counting ties, 5 of the stored top 15 reach the second model's top 15, not 4.
• PDF: short table values such as "11 to 17" no longer break across two lines.

No record, score or stored rank changed. The stored ranking stays the original rater's. Still open: agreement with a second person, which tasks/second-rating.md sets up.

🤖 Generated with Claude Code

v0.5: The Grammar of Simulation

The Grammar of Simulation, version 0.5 (2026-11-02). This revision answers the eight findings that Explore Science's third review (score 90) left open. Each of them asks for a second rater. Where no other person was available, it does the analyses that need no rater, and adds a blind check by a second model.

• Dataset: sha256 3c96f240392d1eb3d49c3d5ca85d751283aba26a082850481fc2b460ed745d69.
• Read it: at https://simgrammar.com/paper/, or as the attached PDF.
• Cite it: CITATION.cff, with the dataset hash. The concept DOI 10.5281/zenodo.23846209 resolves to the newest version.

What changed since v0.4.1 (#72–#81):

• A cross-model check. A second model, gpt-6-sol, coded and rated blind samples of every coding and rating decision under the same operator.
  • It agrees closely on families, simulation-state classes and technique maturity (κ 0.82 to 0.90), less on effect types (κ 0.73), and little on opportunity scores (α 0.28 to 0.48).
  • The paper reports it as agreement across models, not independent agreement.
• The bets ranking does not replicate. The two readings' priority orders correlate at ρ = 0.12. The stored ranking stays the original rater's, and the abstract now says so.
• Effect types. An audit applied the written §8.1 order to every flagged edge, relabelling 41 of them. The second model had already given 8 of the 9 sampled relabelled edges their new label. Replacing a mechanic's substrate is now the most common effect, and agents & behavior the most exposed family.
• Edge recall. A keyword sweep of 231 cited sources found 1 missing edge. A blind recall check found 72% of the recorded edges and 53 extra ones, so the paper now treats an absent edge as weak evidence.
• Mechanic state classes. Every mechanic is classed as simulation state, control, meta-structure or presentation. Clustering and co-occurrence are re-run on the simulation-state subset.
• Family sensitivity. White space and exposure are re-run with borderline mechanics moved to their next-closest family. Small families now carry bootstrap intervals.
• Rating reliability. A simulation finds that a reliable alpha would need 120 to 150 doubly rated mechanics. Rating noise at the observed agreement weakens the score correlations but keeps their sign.
• Predictions. Each now has a pre-specified reference class, generic-diffusion baseline and catalog base rate, and three criteria gained size bars.
• Prior art. A recorded search documents every novelty score. It lowered 16 scores and raised none, and three TRLs were corrected.
• Videos. Eight narrated scenes are re-voiced to match. They are published separately from yoheinakajima/simgrammar-media.

Still open: agreement with a second person. tasks/second-rating.md lets one independent contributor run every sheet in a single pass.

🤖 Generated with Claude Code

v0.4.1: The Grammar of Simulation

The Grammar of Simulation, version 0.4.1 (2026-10-31). This revision answers the Explore Science re-review of v0.4, which scored it 89, up from 83 for v0.3.

• Dataset: sha256 a8db9b2e7a2da59882dc42db02da6a887c10c49a96e3905839930b0a991940a4.
• Read it: at https://simgrammar.com/paper/, or as the attached PDF.
• Cite it: CITATION.cff, with the dataset hash. The concept DOI 10.5281/zenodo.23846209 resolves to the newest version.

What changed since v0.4 (#66, #67):

• Predictions are classified. All 42 are now technique tests (13) or feature tests (29), and only technique tests score where techniques land. Each criterion names the technique class that counts, including the classical methods that satisfy a feature test.
• One shared family list. §5.1 and §8.2 use the same set of non-state families.
• A rater-comparison figure. New Figure 8 compares the original and second ratings on each scale.
• Second-coder study. A seeded sample, coding sheets, and kappa/alpha scoring for family assignment, admission tests and effect types are in place. Agreement fills in once an independent coder submits codes; none are collected yet, and §2.6 says so.

v0.4: The Grammar of Simulation

The Grammar of Simulation, version 0.4 (2026-10-31). This is the revision that answers the Explore Science review of v0.3, which raised 66 minor findings and no major ones.

• Dataset: sha256 08d8ed7099162c1fb1a6213db456013cd136f59a957909eb365642626119a45e.
• Read it: at https://simgrammar.com/paper/, or as the attached PDF (90 pages, built by make pdf from this commit).
• Cite it: CITATION.cff, with the dataset hash. The concept DOI 10.5281/zenodo.23846209 resolves to the newest version.

What changed since v0.3 (#58–#64):

• Vocabulary growth and reuse. The early-coverage baseline is corrected: v0.3's null favoured the early games. Reuse and pair coverage are tested against size-preserving nulls. The result is that later games are built from old parts as the catalog's makeup predicts. A first catalog year is now correctly called an upper bound on a mechanic's origin year.
• Clusters. The 22-way partition is one admissible cut among its neighbours. Families are named only where resampling and a second method agree (21 of 22). Agreement with subgenres is reported against shuffled-label nulls.
• Scores. Chance-corrected rater agreement and family-resampled intervals are added. The score correlations are presented as exploratory rubric-consistency findings.
• Techniques and bets. Exposure weights and the priority formula are printed, with sensitivity and perturbation checks. Effect-type, maturity and inclusion rules are written down. Prediction criteria name the technique class they test.
• Framing and reporting.
  • "Grammar" is defined.
  • Related work on system dynamics, agent-based modelling and gaming simulation is added.
  • Family and admission rules are written down.
  • Appendix E gives definitions.
  • The build's checks are described precisely.
  • Figures are numbered and the PDF is cleaned up.

The episodes and shorts are re-voiced and re-rendered separately.

Videos 2026-10-31c

All 12 episodes and 12 shorts, re-rendered from the dataset after the v0.4 revision (#58–#64), which answers the Explore Science review of v0.3.

• 17 scenes re-voiced in episodes 2, 4, 5, 7, 8, 9, 11 and 12:
  • A first catalog date is now "at least that old", not "a floor".
  • Episode 2 uses the corrected baseline for early coverage.
  • The score relationships are described as ratings, not measurements.
  • Exposure measures where the recorded results point.
  • Vacant lots are vacant in the catalog.
  • A refuted prediction may count against timing rather than the thesis.
  • Episode 9's horizon counts are updated.
• Every frame's numbers come from the current data.
• Japanese subtitle tracks are updated for the re-voiced lines and timed from these renders.

media/published.yaml lists each file with its sha256, and the Pages workflow checks them on deploy.

v0.3: The Grammar of Simulation

The Grammar of Simulation, version 0.3 (2026-10-31). This is the first tagged release of the paper and of the dataset it is built from.

• Dataset: sha256 6e87b0ed28e958d2f6a7476f7ceb09782b896f524fc8d0527b7477b8ce179a77. It holds 179 games, 199 mechanics, 46 techniques, 43 opportunities and 709 sources, recorded across 69 runs under contract 1.2.
• Read it: at https://simgrammar.com/paper/, or as the attached PDF. The PDF is built by make pdf with tectonic, deterministically, from this commit.
• Cite it: CITATION.cff at the repository root. Please add the dataset hash above, because the data keeps changing between versions.

Since the outside reviews of 2026-10-31, this version adds:

• a related-work section, a data availability appendix, and a second-rater robustness check (§5.6);
• a tag audit of all 179 games, with every tag it changed now sourced, and the number of clusters re-selected by its stability rule (22);
• a citation file, a PDF build, and site fixes found in review, including at phone width.

The episodes and shorts that go with this version are in release videos-2026-10-31b.

Videos 2026-10-31b

All 12 episodes and 12 shorts, re-rendered from the dataset after run 2026-10-31-tag-audit-follow-up.

• 19 narration lines in episodes 1, 2, 3, 8 and 12 and shorts 1 and 8 are re-voiced where the wording changed. The changes: production chains now in 42 games, Factorio and Anno 1800 sharing six mechanics in neighbouring families, 22 families picked by the stability rule, three games that combine building placement with event chains, and the vacant lots that now hold two or three games.
• Every frame's numbers come from the current data.
• Japanese subtitle tracks on the site are timed from these renders.

media/published.yaml lists each file with its sha256, and the Pages workflow checks them on deploy.