[Evreu1pro]

What

Adds two related adaptive long-memory strategies for the iterated Prisoner's Dilemma:

1. ZeroResp v5.4.4 — noise ladder (v5.3 basis) plus responsiveness-gated apology and a cycler contour (periods 2–6, noise-tolerant, mirror-guarded). Epoch debt, delayed retaliation, noise-aware forgiveness, opening disambiguation, finite-horizon harvest.
2. ZeroResp v2 (implementation revision 2.2, tournament candidate) — earlier backbone with early-sharp retaliation, echo-shield / contrition, noise-aware forgive, deadlock break, and smart end-game harvest.

This PR supersedes #1495 (ZeroResp alone). Prefer reviewing this PR.

Use of AI tools

Strategy logic and tests for ZeroResp v5.4.4 were developed with AI assistance (primarily GLM / Z.ai) on top of the pre-existing ZeroResp v5.3 basis. I set the research questions, accepted or rejected each change, and verified behaviour against paired benchmarks. Grok Bot helped with repository/site publishing and CI monitoring, not with core strategy design. I understand and can explain the submitted code. The change is strategy class(es) plus tests (isolated).

Why

ZeroResp v5.4.4 is the current production build from the project package. At 5% channel noise on the full Axelrod library it is the best of our family (seed 42: full-library #2). Classifiers are honest: manipulates_state: True because of optional class-level opponent profiles (switchable off via use_profiles / Features).

External benchmarks (reproducible package)

Full report and site: https://github.com/Evreu1pro/ZeroResp · https://evreu1pro.github.io/ZeroResp/

Setting	Result
Full library · 5% noise · seed 42	ZeroResp v5.4.4 #2
Rank stability @ 5% (seeds 42/43/44)	#2 / #5 / #11
Cycler contour vs v5.4.2	+0.36 SPT on CyclerCCD (fresh seeds)
ZeroResp v2 Higher League Top-40 · 400×2 · 7 seeds	avg rank #2 (documented on package)

How to test

pytest axelrod/tests/strategies/test_zeroresp.py axelrod/tests/strategies/test_zeroresp_v2.py -q

Smoke:

import axelrod as axl
m = axl.Match((axl.ZeroResp(), axl.TitForTat()), turns=100, seed=0)
m.play()
print(m.final_score(), axl.ZeroResp().name)  # "ZeroResp v5.4.4"

Coverage

• Initial move always C; reset / clone for multi-rep tournaments
• Noise ladder, responsiveness-gated apology, cycler contour
• Feature flags / ablation paths
• ZeroResp v2 early-sharp, echo shield, noise forgive, deadlock, harvest

Changes

Path	Role
axelrod/strategies/zeroresp.py	ZeroResp v5.4.4
axelrod/strategies/zeroresp_v2.py	ZeroResp v2 (rev 2.2)
axelrod/tests/strategies/test_zeroresp*.py	TestPlayer suites
_strategies.py / classifiers / docs	registration

Breaking

No — additive only. Display name is ZeroResp v5.4.4 (was v5.3 on this branch earlier).

Review notes for maintainers

• Happy to drop ZeroResp v2 and keep only ZeroResp v5.4.4 if you prefer a single strategy.
• manipulates_state is True (class-level profiles); set use_profiles=False for a match-local player.
• Project site + research notebook: https://evreu1pro.github.io/ZeroResp/

[Evreu1pro]

Update for maintainers (2026-08-04)

Pushed revision 2.2 of ZeroResp v2 (short adaptive buffer, dual noise-aware forgiveness, smart harvest / grim probe).

CI was already green on the previous commit; this push should re-run the matrix.

Evidence package (algorithm + unit tests + Higher League / deep-validation tables):
https://github.com/Evreu1pro/ZeroResp

Why this is ready for the strategy list

• Non-cheating classifiers (no source/state manipulation)
• Full TestPlayer suite + registration + docs automodule
• Stable elite placement (Added docstring to Grudger's strategy method #1–Get results graph to plot but sort strategies by median score #3) vs classic + Axelrod elites; H2H with Omega TFT stays near-ideal mutual cooperation
• Honest limitation documented: sustained 1–3% channel noise

Request: please review when you have bandwidth. If two strategies are too many, I can collapse to a single ZeroResp entry with the 2.2 logic.

Thank you for maintaining Axelrod — happy to address any requested API/doc changes promptly.

[Evreu1pro]

CI fix for revision 2.2 (just pushed)

Addressed the two failures from the previous matrix run:

1. test_repr — Player.__repr__ appends base_epoch, so the TestPlayer name is now ZeroResp v2: 25 (was ZeroResp v2).
2. Coverage 99.95% — added targeted tests for smart harvest / grim-probe paths (last-safe D, probe_fired, stochastic probe, non-grim probe window).

What is in this PR (for clarity):

• ZeroResp — v1 backbone
• ZeroResp v2 — class API, implementation revision 2.2 (short adaptive buffer, dual noise forgive, smart harvest / grim probe)

Happy to rename the display string to ZeroResp v2.2 if maintainers prefer that in tournament plots.

[drvinceknight]

Can you fix the merge conflicts (#1497 fixed a problem with numpy).

[Evreu1pro]

Thanks @drvinceknight — merge conflicts with dev are resolved.

Took numpy>=2.1 from #1497 and kept the float() cast in fingerprint point generation. CI should re-run on this push.