Independent Validation and Model-Routing Summary

Status: Completed research pass Date: 2026-09-20

Objective

Obtain evidence for the selective Communication Engineering preflight outside the GPT-5.6 Sol execution path used for Experiments 01-05.

Experiment 06 — explicit Claude Sonnet 4.6

Target:

• GitHub Copilot CLI
• claude-sonnet-4.6

Result:

• Copilot CLI and GitHub Actions integration worked.
• The requested model was not available to this repository/account.
• No condition output was produced.
• No fallback was substituted.

Disposition:

Incomplete. Not scored.

Model availability probes

Explicit selection was also unavailable for the tested:

• Claude Haiku 4.5
• Gemini 3.1 Pro Preview
• Gemini 3.5 Flash
• GPT-5.4
• GPT-5.3 Codex
• MAI-Code-1.1-Flash

The account can nevertheless use Copilot's auto route.

OpenTelemetry tracing showed that auto can resolve to different models depending on the request.

Experiment 07 — full-context auto-route test

Pre-registered requirement:

Both A and B must resolve to mai-code-1.1-flash.

Observed:

• A ordinary request -> mai-code-1.1-flash
• B full Communication Engineering context -> gpt-5.6-luna

The model-equality assertion failed.

Disposition:

Invalidated. Not scored.

Experiment 08 — compact-preflight auto-route test

To reduce treatment size, B received only a compact generic preflight kernel rather than the five operational context files.

Pre-registered requirement remained:

Both A and B must resolve to mai-code-1.1-flash.

Observed:

• A ordinary request -> mai-code-1.1-flash
• B compact preflight -> gpt-5.6-luna

The model-equality assertion failed again.

Disposition:

Invalidated. Not scored.

New supported finding: model-routing is part of treatment cost

The routing effect reproduced across:

• the full operational context;
• the compact preflight context.

Therefore, under an auto-routing model surface, adding Communication Engineering context can change:

• the model used;
• model capability;
• latency;
• inference/token economics.

This makes naïve A/B comparisons invalid unless model identity is controlled.

It also means prompt-context size/structure is itself an operational cost of the method.

Engineering consequence

Do not implement Communication Engineering primarily as ambient prompt text.

Move stable preflight mechanics into deterministic typed code where possible.

The model/human should handle only the unresolved judgment:

• classify communication fields;
• decide materiality;
• judge safe inference;
• resolve novel ambiguity.

The executable preflight should compute:

• validation;
• missing required state;
• optional unknown state;
• clarification grouping;
• activation decision;
• bounded-progress permission;
• verification escalation.

F# prototype

Research implementation:

research/prototypes/communication-preflight-fsharp-v0.1/

It is deliberately:

• dependency-free beyond the .NET SDK;
• typed;
• deterministic;
• provider-neutral;
• free of natural-language parsing;
• free of AI SDK dependencies;
• not an Ordo runtime dependency.

It follows the project's preferred split:

natural-language interpretation
        ↓
Decide / Deliberate
        ↓
typed communication state
        ↓
Compute
        ↓
preflight result

Independent-model evidence status

No valid controlled independent-model comparison was completed.

Do not claim one.

The attempt did, however, reveal a reproducible routing confound and strengthened the case for deterministic preflight semantics.

Next evidence gate

Preferred next evidence:

1. human/agent workflow validation using the executable preflight;
2. a provider/model surface that supports explicit model pinning;
3. downstream participant evidence tied to purpose-specific outcomes.

Production lifecycle packaging remains gated.