# Real three-provider Model Jury demo

This is a visible-output transcript summary from a real installed-package run on 2026-08-31. It contains no credentials, request headers, environment snapshot, or hidden reasoning. Public route names are shown after the nonfunctional `dsh-model-jury` branding migration; the provider backends and protocol are unchanged.

## Question

> For a small engineering team building a local-first developer tool, should the core orchestration engine use a deterministic workflow or allow the LLM to dynamically select the next agent? Compare reliability, debuggability, flexibility, cost, failure recovery, and long-term maintainability.

## Providers

| Seat | Provider | Model/backend | Health |
|---|---|---|---|
| GPT | `model-jury-codex` | native Codex app-server `0.149.1` | PASS with `OPENAI_API_KEY` absent |
| GLM | `glm` | `glm-5.3` | PASS |
| DeepSeek | `deepseek-official` | `deepseek-v4-flash` | PASS |

Anonymous mapping for this run:

| Seat | Position |
|---|---|
| GPT | P2 |
| GLM | P1 |
| DeepSeek | P3 |

## Round 1 — blind independent positions

- **GPT:** deterministic, versioned state-machine control should be the core, with narrowly allowlisted LLM routing only where explicit rules are inadequate. Confidence 92.
- **GLM:** deterministic workflow control with schema-validated, bounded LLM discretion; avoid fully dynamic agent selection. Confidence 78.
- **DeepSeek:** deterministic state machine or DAG, with LLM variability confined to predefined recoverable nodes. Confidence 82.

The three prompts were byte-identical and contained no other seat output. Their calls started within 80 ms of one another.

## Round 2 — anonymous cross-critique

Only P1/P2/P3 identified competing positions. No GPT, OpenAI, Codex, GLM, Zhipu, DeepSeek, provider id, or model id appeared in any competitor packet.

- **GPT** preferred P2 and argued it best combined deterministic reliability with bounded routing; it criticized P1 for overclaiming exact replay and P3 for understating flexibility costs. Confidence 94.
- **GLM** preferred P2, emphasizing its operational detail while challenging unsupported assumptions about bounded routing and future retrofit cost. Confidence 72.
- **DeepSeek** preferred P3, arguing that even allowlisted next-step routing can reintroduce nondeterministic control flow and harder replay. Confidence 84.

Every response named the strongest point, weakest point, possible error, and missing evidence for both competing positions. The critiques contained substantive disagreement and were not consensus-only summaries.

## Round 3 — revision and rebuttal

- **GPT:** P2, unchanged. Deterministic state machine with guarded, recorded, allowlisted routing checkpoints. Confidence 86.
- **GLM:** changed to `hybrid`. It retained deterministic graph control, narrowed routing, and added an explicit concern that the recommendation depends on task decomposability. Confidence 82.
- **DeepSeek:** changed to P2 after accepting guarded next-step routing with deterministic fallbacks and persisted decisions. Confidence 80.

Both `changed_mind: true` and `changed_mind: false` were accepted in the same real run.

## Deterministic verdict

```text
Result: MAJORITY
Quorum: 3/3
Votes: P2 = 2, hybrid = 1
Winning choice: P2
Changed minds: 2
Critical review required: no
```

Recommendation: use a deterministic, versioned state machine or DAG as the core. Permit LLM routing only at selected checkpoints over a small allowlist, with schema validation, recorded decision context, retry limits, checkpoints, and an explicit safe fallback.

Minority / dissent: GLM retained a distinct hybrid recommendation. It warned that a stable workflow graph and testable bounded routing are assumptions; a product centered on novel, unstructured autonomy could reverse the recommendation. This single structured critical risk was surfaced but did not veto the majority.

## Timing

| Round | GPT | GLM | DeepSeek |
|---|---:|---:|---:|
| Round 1 | 33.279 s | 35.889 s | 10.866 s |
| Round 2 | 26.759 s | 103.273 s | 45.349 s |
| Round 3 | 27.786 s | 70.860 s | 35.887 s |

Run id: `20260831T084039975Z-e819e49e`.

## Deliberation stress run

A second real run compared Apple/iOS support with ARM64/SIMD optimization. DeepSeek initially chose ARM64/SIMD while GPT and GLM chose Apple support. Anonymous critique changed both GLM and DeepSeek to guarded hybrid proposals. The final deterministic vote was `hybrid = 2, P3 = 1`, with GPT preserved as the named minority. All three submitted structured critical risks, demonstrating that critical-risk reporting remains independent of the vote.