paper §7.4: complete 8-cell topology×model grid (Gemini folded in)

Gemini's project-wide key-restriction 403 is cleared (key restricted to the
Generative Language API), so Gemini rejoins the grid. Full 2×4 ranking
(pipeline/debate × Claude/GLM-4.7/Gemini/heterogeneous), judge-ranked 1-8:
D2 > P2 > P3 > P1 > D3 > PG > DG > D1.

Adds a third finding: model strength is task-dependent — Gemini-2.5-flash
placed 2nd on the §7.3 concision brief but 6th-7th here on reasoning
("generic list", "idea repeated"): concise-but-shallow. Reinforces that the
optimal config is a joint choice over {topology × per-role model × task}.
Gemini cells ran via the durable async run path (enqueue → worker → poll).

Co-Authored-By: Claude Opus 4.8 <[email protected]>

papers/dynamic-agentic-topologies.md

@@ -232,32 +232,36 @@ measured rather than assumed.
 
 The platform's defining question is not "best topology" *or* "best model" in
 isolation but their **interaction**. We crossed two topologies (**pipeline**, 3
-stages; **debate**, propose→critique→revise→judge) with three model casts
+stages; **debate**, propose→critique→revise→judge) with four homogeneous/mixed
-(homogeneous Claude, homogeneous GLM‑4.7, and a heterogeneous GLM→Kimi→Claude
+model casts (**Claude**, **GLM‑4.7**, **Gemini‑2.5‑flash**, and a heterogeneous
-cast) on a fixed decision task — *"Rust or Go for a 3‑person startup backend?"* — and
+**GLM→Kimi→Claude**) on a fixed decision task — *"Rust or Go for a 3‑person startup
-ranked all six final outputs 1–6 with an independent judge on reasoning quality,
+backend?"* — and ranked all eight final outputs 1–8 with an independent judge on
-clarity, and concision:
+reasoning quality, clarity, and concision:
 
-| Topology ＼ cast | Claude | GLM‑4.7 | Heterogeneous |
+| Topology ＼ cast | Claude | GLM‑4.7 | Gemini | Heterogeneous |
-|------------------|:------:|:-------:|:-------------:|
+|------------------|:------:|:-------:|:------:|:-------------:|
-| **Pipeline**     |   3    |  **2**  |       4       |
+| **Pipeline**     |   4    |  **2**  |   6    |       3       |
-| **Debate**       | **6**  |  **1**  |       5       |
+| **Debate**       | **8**  |  **1**  |   7    |       5       |
 
-Two results stand out. First, **all six cells converged on the same decision** ("Go")
+Three results stand out. First, **all eight cells converged on the same decision**
-— so on a task with a clear prior, topology and model move *justification quality*,
+("Go") — so on a task with a clear prior, topology and model move *justification
-not the answer. Second, and central: **the topology × model interaction is real and
+quality*, not the answer. Second, and central: **the topology × model interaction is
-non‑monotone.** Debate *amplified* the strongest model (GLM‑4.7: pipeline 2nd →
+real and non‑monotone.** Debate *amplified* the single strongest model (GLM‑4.7:
-debate 1st, its extra adversarial round adding a hiring‑cost nuance) but *degraded*
+pipeline 2nd → debate 1st, its extra adversarial round adding a hiring‑cost nuance)
-the others (Claude: 3rd → dead last, its judge wasting words on an "upstream agents
+but *degraded* every other cast (Claude: 4th → **dead last**, its judge wasting words
-agree" meta‑citation; heterogeneous: 4th → 5th). In other words, **the best topology
+on an "upstream agents agree" meta‑citation; Gemini: 6th → 7th; heterogeneous: 3rd →
-depends on the model, and vice‑versa** — debate is not a universal upgrade; it pays
+5th). **Debate is not a universal upgrade — it pays off only with a model strong
-off only with a model strong enough to use the extra rounds. This is precisely the
+enough to exploit the extra rounds, and is actively harmful otherwise.** Third,
-result the platform is built to produce and that neither a topology‑only nor a
+**model strength is task‑dependent, not absolute**: Gemini‑2.5‑flash placed *2nd* on
-model‑only study can see: the optimal *configuration* is a joint choice over
+the §7.3 concision brief but *6th–7th* here — the judge faulted it for a "generic
-{topology × per‑role model}, discovered empirically, with the §5 safety invariant
+list" and "the same idea repeated," i.e. concise‑but‑shallow, an asset for brevity
-holding across every cell. The sweep was run sequentially under the subscription
+and a liability for nuanced reasoning. The same model is strong or weak depending on
-plans' quota/concurrency caps (21 turns total), the practical envelope for this
+the role's rubric. Together these confirm the result the platform is built to produce
-class of experiment.
+and that neither a topology‑only nor a model‑only study can see: the optimal
+*configuration* is a joint choice over {topology × per‑role model × task}, discovered
+empirically, with the §5 safety invariant holding across every cell. The sweep ran
+sequentially under the subscription plans' quota/concurrency caps, the practical
+envelope for this class of experiment.
 
 ## 8. Limitations and future work