Operational Context Initial Evidence Summary

Status: Completed initial prospective validation Date: 2026-09-20

Question

Does injecting the current Communication Engineering operational context improve communication work enough to justify automatic distribution/injection?

Experiments

Experiment 01 — incident procedure

Task:

• concise technical procedure;
• all state/action facts supplied;
• safety-critical unknown-effect distinction.

Result:

• ordinary request: 18/18
• Communication Engineering: 18/18
• both 91 words

Disposition:

• null correctness result.

Experiment 02 — executive decision memo

Task:

• mixed quantitative evidence;
• explicit recommendation;
• decision owner and auditability requirement supplied.

Result:

• ordinary request: 23/23
• Communication Engineering: 23/23
• 206 vs 216 words

Disposition:

• null correctness result.

Experiment 03 — flawed memo review

Task:

• 12 pre-injected semantic/authority/uncertainty defects;
• full fixed source facts supplied.

Result:

• ordinary review: 12/12 defects detected, 0 false positives
• Communication Engineering review: 12/12 defects detected, 0 false positives
• corrected memos passed all guardrails

Disposition:

• null correctness result.

Supported finding

The current operational context is non-regressive in these tests.

It did not:

• reduce correctness;
• cause excessive verbosity;
• leak writer/persona imitation;
• invent authority;
• flatten uncertainty;
• introduce a universal quality score.

It also supplied a useful shared vocabulary for defect classes.

Unsupported claim

The current evidence does not show that injecting the full context improves a strong reasoning model's first-draft correctness or review recall when the source task/facts are already explicit.

Do not claim otherwise.

Method change caused by evidence

The full method is now selectively activated.

Use Communication Engineering when decision-relevant communication inputs are materially unresolved:

• purpose;
• audience/task;
• proof obligation;
• authority;
• uncertainty;
• accessibility;
• competing use modes;
• cultural/translation boundaries;
• success measure.

Do not require mechanical contract completion for simple, fully specified tasks.

Likely value hypotheses now worth testing

1. Brief construction: detect missing communication inputs before drafting.
2. Cross-model consistency: help less capable/differently trained agents preserve the same boundaries.
3. Human/team review: make proof, authority, uncertainty, and outcome choices inspectable.
4. Telemetry/taxonomy: classify recurring communication defects and measure reduction over time.
5. High-consequence gates: ensure purpose-specific verification happens when stakes justify it.
6. Outcome linkage: connect communication decisions to downstream task/comprehension/transfer evidence.

Package decision

Do not yet evolve @echelon-foundry/communication-engineering into an automatic lifecycle context-injection tool.

Evidence required before that promotion:

• independent-model/provider delta;
• human/team rework or consistency improvement;
• participant outcome evidence;
• or a brief-construction experiment showing unique value.

Repository disposition

Keep agent-context/ as the canonical research/operational source.

It is ready to be used deliberately in further experiments and real communication work, but not yet justified as mandatory ambient context.

Addendum: Brief-Construction Evidence

Experiment 04 — brief construction

Underspecified internal leadership update.

• Ordinary response: 5/10.
• Communication Engineering: 10/10.
• Difference: +5 points.
• All eight guardrails passed.
• Communication Engineering recovered all eight preregistered missing communication inputs in four grouped questions.
• Ordinary response surfaced only part of the missing communication state.

Experiment 05 — external customer notice

Underspecified API-deprecation notice.

• Ordinary response: 6.5/10.
• Communication Engineering: 10/10.
• Difference: +3.5 points.
• All eight guardrails passed.
• Communication Engineering again recovered all eight missing inputs in four grouped questions.
• Ordinary response missed recipient role, cadence, localization/accessibility, and fully explicit personalization authority.

Revised synthesis

The first five experiments now separate two conditions clearly.

When the brief is already strong

Experiments 01-03:

• no measured correctness or review-recall gain from full context;
• no observed regression.

When decision-relevant communication state is missing

Experiments 04-05:

• replicated improvement in missing-input recovery;
• no additional unsupported assumptions;
• no redundant known-fact questions;
• useful bounded progress remained possible.

The best-supported current role for Communication Engineering is therefore:

selective communication preflight before drafting when important communication state is incomplete.

This remains same-model process evidence. Downstream reader outcomes and independent-model reproducibility remain unproven.

Tooling consequence

Do not build an always-on prose wrapper.

The next product candidate is a small preflight capability that:

• receives typed communication-state assessment;
• preserves field provenance;
• identifies material missing fields;
• groups clarifications;
• selects bounded-progress vs clarification vs verification;
• gets out of the way when ordinary drafting is already appropriate.

See:

• docs/communication-preflight-requirements-candidate-v0.1.md
• research/prototypes/communication-preflight-v0.1/