v3.43 black-box audit: I-1..I-7 convergence fixes for 9 v3.42 failures

[FluffyAIcode]

Scope

• SUT: scheme_b_v343.py via AgentMemorySystem.py redirect.
• Runner / spec: unmodified.

v3.43 architectural changes

Tag	Target cases	Mechanism
I-1	4.7, 4.8, 4.10, 4.15, 4.17, 4.21	Symmetric CFG content_bias (fwd adds bias to both cond/uncond); dampen=1.0
I-2	4.7	Linearised repetition penalty (penalty 2.5, exp 1.0, cyclic 5)
I-3	4.12, 4.21	Residual excludes prompt+generated content tids
I-4	4.23	Residual subtracts WTE global mean
I-5	4.24	Hybrid context encoder (WTE centroid + β=0.8 × hidden_mean)
I-6	4.25	Slot-exact renormalisation
I-7	4.17	Test-scoped deterministic execution

Audit result

• 26 cases, elapsed 1452.3 s on CPU.
• Pass: 15 / 26, Fail: 11 / 26.
• v3.42 baseline: 17 / 26 pass.

Delta vs v3.42:

Transition	Count	Cases
FAIL → PASS	0	—
PASS → FAIL	2	4.8, 4.11
Persistent FAIL	9	4.7, 4.10, 4.12, 4.15, 4.17, 4.21, 4.23, 4.24, 4.25

Net pass count −2. [I-1] algebra was correct (bias appears once after CFG) but dampen=1.0 + [I-2] linear repeat penalty together removed the exponential cap that broke repetition cycles. Sample output (4.8, prompt="The pianist"): "practiced pian pian regularly cards pian Chop pian Chop Chop practiced Chop pian practiced".

Equilibrium analysis: at bias ≈ 18 logit, content_repeat_penalty = 2.5 × N^1.0, break-even N ≈ 7. Runner case 4.21 reports avg_max_repeat_per_content_token = 5.33 (was 3.33 in v3.42).

Other structural:

• 4.24 intra jumped to 0.90 (hybrid cohesion works) but inter also to 0.84 — hybrid trade-off
• 4.23, 4.17, 4.25 unchanged
• 4.7, 4.10, 4.12, 4.15 unchanged (runner samples lg_cond without prefix for 4.10, with wrong quantisation for 4.15)

Single-knob recovery path for v3.44: fwd_path_bias_dampen = 0.25 while keeping [I-1] symmetric CFG. Restores v3.42's ~4.5-logit effective bias with bias still visible in lg_cond.

Artifacts

• reports/v343_blackbox/{report.json, report.md, runner.log, audit_feedback.md}