RESEARCHv5.4.x
From v5.3 to a hybrid: ZDB validation, full-library tournaments, loss analysis, and two new contours — v5.4.3 (noise switch) and v5.4.4 (cycler contour). Every claim paired, every negative result recorded.
- Gap to #1 (Evolved ANN 5 Noise 05)
- −17.1 pts
- Next of ours (v5.4.3)
- #3
- v5.3 baseline
- #4–#6
- Gradual
- +0.2622
- Calculator
- +0.4068
- Random cost
- −0.0472
Full-library round-robin with our versions inside the pool
200 turns · 1 rep · paired seeds| Noise | v5.4.4 | v5.4.2 | v5.4.1 | v5.3 | leader SPT |
|---|---|---|---|---|---|
| 0% | #9 | #10 | #16 | #6 | 664.2 |
| 5% | #2 | #8 | #4 | #5 | 592.2 |
| 10% | #3 | #8 | #5 | #4 | 553.9 |
Ranks inside our tight cluster (0.3–2 SPT) shuffle across seeds — the family reliably sits top-10 at every noise level; per-version stars need reps ≥ 3. The 0% leader is Evolved FSM 6; at 5–10% it is Evolved ANN 5 Noise 05 — a strategy literally evolved against this library.
What each new version is
- v5.4.2
- responsiveness-gated apology: ΔPresp = P(C|our C) − P(C|our D) ≥ 0.35. Keeps the spin-out fix, halves the Random cost of v5.4.1 (−0.093 → −0.047).
- v5.4.3
- noise-level contour switch (quiet → static gate, noisy → responsiveness gate). Honest negative result: under channel noise it converges to v5.4.2 — the responsiveness gate already is the right switch.
- v5.4.4
- cycler contour: fixed-period detection (periods 2–6, window 18, noise-tolerant, mirror-guarded) blocks the apology — the aggressive base contour keeps exploiting. Best of ours at 5% noise.
Why we were not top-5
per-opponent gap decomposition| Opponent category | 0% | 5% | 10% |
|---|---|---|---|
| Cyclers (8) | +0.19 | +0.58 | +0.44 |
| Exploiters (26) | −0.06 | +0.02 | +0.08 |
| Stochastic (5) | +0.12 | +0.06 | +0.05 |
| Reciprocators (26) | −0.03 | −0.28 | −0.18 |
| Trained (13) | −0.13 | −0.27 | −0.25 |
Mean SPT gap, rivals above us minus ours (negative = we out-earn the leaders). The leak is not the hard opponents — it is the Cycler family, and it is our own regression: under channel noise the periodic D's of a cycler read as unprovoked, the apology fires and resets the exploitation (CyclerCCD at 5%: v5.3 3.369 vs v5.4.2 2.207).
Replicates out-of-sample (fresh seeds 1000–1029). AntiCycler needs the forbearance ratchet from the AOM line — next mechanism on the list.
Rank stability across independent draws
full library · 5% · seeds 42 / 43 / 44| Version | seed 42 | seed 43 | seed 44 |
|---|---|---|---|
| v5.4.4 | #2 | #5 | #11 |
| v5.4.2 | #8 | #3 | #7 |
| v5.3 | #6 | #6 | #4 |
| v5.2 | — | #2 | #5 |
The ZeroResp family holds top-10 of 230 at 5% noise on every draw; the ordering inside the family is draw-dependent (mean SPT within hundredths). Out-of-sample targeted replication on seeds 1000–1029: TTFT +0.23, Gradual +0.23, HardProber +0.20, cycler fix +0.36; Calculator lost significance, Random −0.08.
Methodology (ZDB)
- Pre-registration
- criteria committed before runs
- Unit of analysis
- per-seed means, Wilcoxon + Holm
- Pairing
- identical seeds & opponents per version
- Bit-level checks
- full move-history comparison
- Artifact hunts
- rank 4/225 exposed as measurement artifact
- Negative results
- published in-repo (v5.4.3, gray jitter, layer 3)
Every headline number on this page traces to a JSON artifact and a runner script in the repository.
Does the model know its opponents?
overfitting auditNo opponent names appear in the decision logic — zero branches of the form "if it is X, do Y". All behavior is inferred from the current match: noise estimation, responsiveness gate, period detection, evidence ladder, tactical classifier.
- Profile memory
- learned stats keyed by opponent name across rematches; inactive in single matches; declared via manipulates_state; switchable off
- Tuned constants
- thresholds tuned on the reporting pool; external holdout showed −0.03 (recorded honestly)
- Meta-lessons
- some fixes motivated by specific dev seeds; mechanisms stay generic and replicate on fresh seeds
- Context
- the strategies above us (Evolved ANN 5 Noise 05 et al.) were literally evolved against this library — our rules are readable and auditable instead
Attribution / Use of AI tools
- Basis
- Production through v5.3 is the human-directed ZeroResp line (noise ladder).
- v5.4.x
- Built on that v5.3 basis with AI coding agents under human direction: the human set hypotheses, review criteria, and every release decision; agents implemented runners, ran batteries, and drafted reports.
- Tools
- Primary coding agent for the v5.4.x cycle: GLM (Z.ai). Site updates, GitHub packaging, CI/monitoring and publication support: Grok Bot. Autocomplete and spelling fixes do not count.
- Responsibility
- The maintainer of this repo is responsible for the contribution and can explain the strategy decisions. v5.4.4 is current on the main site; this page keeps the research notebook for the cycle that produced it.
Full provenance lives in the repository: pre-registration files are committed before their test runs, and the git history separates the research branches from the release line. One disclosure block is enough — no per-line AI markers.