generated: '2026-09-19'
method: searched
source: https://raw.githubusercontent.com/ryuxik/snhp/main/gametheory/CHANGELOG.md (200, 19,453 bytes, fetched 2026-09-19);
  rendered at https://github.com/ryuxik/snhp/blob/main/gametheory/CHANGELOG.md
checked: '2026-09-19'
summary: 'The only dated changelog is a repository file, gametheory/CHANGELOG.md, with three entries all from 2026-05-01/02.
  It is an EXPERIMENT and TUNING log rather than an API release log: it records paired LLM-tournament results (lifts,
  confidence intervals, p-values, dollar cost per run), the magic-number tuning framework, and a variant that was
  "shipped, reverted, and what we learned". It contains no endpoint additions, no breaking-change notices and no
  version numbers, and it has not been updated since — although the repository was pushed through 2026-07-26 and
  the deployment rebuilt 2026-09-16. The site does not serve it (/CHANGELOG.md 404) even though llms.txt says "See
  CHANGELOG.md" twice. Release tags exist on GitHub (v0.1.0 2026-06-30, v0.2.0 2026-07-10) without release notes.'
scheme: dated headings (YYYY-MM-DD — title); no semver in the log
current_version: 0.1.0 (product) / 0.2.0 (PyPI) — neither appears in the changelog
window: complete file (3 entries)
entries:
- date: '2026-05-02'
  title: Hassabis experiment portfolio (5 experiments, $5.88 total)
  kind: experiment-log
  breaking: false
  sections:
  - Experiments + verdicts
  - Strategic implications
  - What we deliberately did NOT do
  highlights_verbatim:
  - E1 — H3 fresh-seed replication ($0.69)
  - E2 — H1 power-up at N=50 ($3.01)
  - E3 — Sim-only commit-reveal sweep ($0.00)
  - E4 — Oracle-reveal LLM-loop validation ($1.48)
  - E5 — Stuck-deal stratification ($0.00, re-mining existing data)
  - E6 — Cross-vendor benchmark ($0.73, post-YC v5 review)
- date: '2026-05-01'
  title: Magic-number framework + Phase 1-2 tuning (in progress)
  kind: tuning-framework
  breaking: false
  sections:
  - Context
  - What shipped
  - Methodology lessons
  - Methodology lessons
  highlights_verbatim:
  - '`gametheory/negotiation/_config.py`'
  - 20 were tagged `magic-tunable`
  - Env-var override mechanism
  - '`gametheory/negotiation/_sim.py`'
  - Sim validation gate
  - Phase 1.5 sensitivity analysis
- date: '2026-05-01'
  title: 'peer_cs variant: shipped, reverted, and what we learned'
  kind: experiment-log
  breaking: false
  sections:
  - Summary
  - Timeline
  - What was reverted
  - What was preserved
  - Lessons for future tuning work
  - Live N=20 numbers (preserved for honest framing)
  highlights_verbatim:
  - Hypothesis
  - v1 (Optuna-tuned)
  - +0.297, 95% CI [+0.279, +0.315], Wilcoxon p<1e-11.
  - N=20 LLM validation at n_steps=6 with v1
  - p=0.13 — not stat-sig.
  - maximized opponent utility subject to self-target
releases_without_notes:
- tag: v0.2.0
  date: '2026-07-10'
  notes: none
- tag: v0.1.0
  date: '2026-06-30'
  notes: none
see:
  lifecycle: lifecycle/snhp-dev-lifecycle.yml