# ASD-STE100 Communication Audit

Date: 2026-08-28

## Objective

Use ASD-STE100 Issue 9 for all Omnius natural-language communication.
Apply the rule to external replies and internal agent messages.
Preserve exact machine content without a change.

## Standard boundary

[ASD-STE100 Issue 9](https://www.asd-ste100.org/assets/files/ASD-STE100_ISSUE9.pdf) has 53 writing rules.
It also has an approved-word dictionary and permits necessary technical terms.
The standard requires consistent technical terms, short sentences, direct verbs, and clear procedures.

Issue 9 limits a procedure sentence to 20 words.
It limits a descriptive sentence to 25 words.
It requires one instruction in most procedure sentences.
It also requires quoted text to remain unchanged.

The [official STE description](https://www.asd-ste100.org/about.html) identifies an approved dictionary and permitted technical terms.
The [official checker guidance](https://www.asd-ste100.org/ste-software.html) states that software aids are not complete compliance proof.

Therefore, Omnius can enforce structural STE rules automatically.
A qualified human must confirm full dictionary compliance.
Omnius must not claim certified compliance from a model or heuristic check.

## 2026 reliability evidence

Current research supports three separate communication layers.

1. Use controlled natural language for human and agent prose.
2. Use schemas and exact fields for machine messages.
3. Use code validators for syntax and meaning.

A 2026 protocol review reports weak support for clarification, context alignment, and semantic verification.
Omnius now names these duties in the common contract.
Source: [Beyond Message Passing](https://arxiv.org/abs/2604.02369).

A 2026 harness study reports that prompt text alone did not preserve critical guarantees.
It assigns deterministic guarantees to code, schemas, manifests, and validators.
Omnius now uses this rule in the common contract.
Source: [From Prompts to Contracts](https://arxiv.org/abs/2607.08028).

A 2026 structured-output study separates valid syntax from valid meaning.
It also reports model-specific failure patterns.
Omnius must validate JSON structure and field meaning.
Source: [StructHallu-Drift](https://aclanthology.org/2026.surgellm-1.22/).

A 2026 reliability framework treats retries and verification as different recovery operators.
Omnius must classify a failure before it selects a recovery method.
Source: [Cost-Aware Adaptive Reliability](https://arxiv.org/abs/2605.09121).

2025 structured-output studies remain directly applicable.
They show that schema constraints need adapted prompts, examples, and validators.
Sources: [The Hidden Cost of Structure](https://aclanthology.org/2025.ranlp-1.124/) and [Schema Reinforcement Learning](https://aclanthology.org/2025.acl-long.243/).

## Lossless exception boundary

Do not change these items during STE conversion:

- Code and code blocks
- JSON, schemas, keys, and enum values
- Commands, paths, options, and environment keys
- API routes, headers, status codes, and media types
- Tool names and tool argument names
- Model names and provider names
- Sentinels, tags, placeholders, and parser labels
- User input and source quotations
- Logs, errors, traces, and tool output
- Required literal replies

Use STE for text that introduces or explains these items.
If STE conflicts with an exact contract, the exact contract has priority.

## Runtime architecture review

| Surface | Previous condition | Current control | Status |
|---|---|---|---|
| Small tier | Large duplicated prose with long sentences | Shared common contract and small-tier addition | Updated |
| Medium tier | Large duplicated prose with long sentences | Shared common contract and medium-tier addition | Updated |
| Large tier | Large duplicated prose with long sentences | Shared common contract and large-tier addition | Updated |
| Ollama and compatible providers | No common language envelope | Provider adds one STE policy message | Updated |
| Anthropic adapter | Independent system translation | Provider policy enters before translation | Updated |
| Native Ollama stream | Independent message path | Provider policy enters before native transport | Updated |
| Nexus remote peer | Independent message path | Nexus adds the same policy | Updated |
| Cascade backend | Delegates to Ollama or Nexus | Inner backend adds the policy | Updated |
| Web chat | Separate non-tier system prompt | Web prompt starts with the shared policy | Updated |
| Realtime and phone | Explicit contraction preference | Shared policy and direct spoken rules | Updated |
| Task templates | Long compound instructions | Eight source templates use short direct rules | Updated |
| Telegram router | Many inline contracts and exact schemas | Provider policy governs output | Controlled rewrite required |
| Sub-agent delegation | Separate preambles and exact envelopes | Provider policy governs output | Controlled rewrite required |
| RALPH runners | Separate JSON-sensitive prompt files | Provider policy governs output | Controlled rewrite required |
| Memory and compaction | Separate JSON and summary prompts | Provider policy governs output | Controlled rewrite required |
| CLI dream and emotion prompts | Creative and machine output contracts | Provider policy governs output | Profile-specific review required |
| Reference contracts | 39 long Markdown contracts | Included under the common policy | Controlled rewrite required |
| OpenAPI and help text | Human prose in TypeScript literals | Not model-generated | User-interface review required |

The final provider boundary adds the policy before each model request.
This rule covers inline prompts that still need a controlled source rewrite.
It also covers new prompts that a feature adds later.
The policy does not change exact machine content.

## Source inventory

The prompt audit found 83 source Markdown files.
Generated files under `publish/` are not source files.

- CLI TUI prompts: 7 files
- Prompt package contracts and templates: 15 files
- Orchestrator prompts: 22 files
- Root reference contracts: 39 files

TypeScript also contains direct model prompts and user-interface messages.
These strings need a separate literal extractor because many strings contain machine data.

## Main defects found

### Tier prompts

The three tier files duplicated nearly all content.
Many sentences had more than 25 words.
Several sentences contained more than one instruction.
The smaller tiers did not have a materially smaller common contract.

The new design keeps one common contract.
Each tier file now contains only its tier rules.
This design prevents language drift between tiers.

### Realtime prompt

The previous prompt explicitly permitted contractions.
This rule directly conflicted with STE.
The fallback reply also used a contraction.

The new prompt prohibits contractions.
The fallback now uses two short sentences.
An exact user-requested reply remains exempt.

### Task templates

The source templates used long labels, questions, vague terms, and compound instructions.
The revised templates use one direct action in each procedure sentence.
Exact placeholders and tool names remain unchanged.

### Telegram and tool contracts

Telegram router prompts contain safety, identity, route, and reply policies.
They also contain exact JSON fields and enum values.
A broad text conversion can change router behavior.

Tool contracts contain exact JSON, tool names, argument names, and parser sentinels.
These elements need token-preservation tests before a controlled prose conversion.

### Corrected prompt defect

`packages/cli/prompts/tui/emotion-center.md` had an unclosed parenthesis.
Its examples also conflicted with its required emoji-and-word output.
The revised prompt has short instructions and valid examples.
A test now protects the placeholders and required output format.

## Telegram router failure review

The live failure was not an unavailable model.
The broker completed the requests with hidden reasoning tokens.
Omnius then removed the hidden text and received no visible decision contract.

One predicate controlled two different facts.
It controlled Ollama transport features and Omnius pool ownership.
The broker discovery correctly disabled the local Omnius pool.
That result also disabled Ollama transport behavior by mistake.

The affected requests omitted `think:false` and Ollama context options.
They also used `/v1/chat/completions` instead of the preferred native `/api/chat` route.
The model used its reply budget for hidden reasoning.

The backend now uses two predicates.

- `isOllamaTransport` controls Ollama request fields and native routes.
- `useOllamaPool` controls only local runner acquisition and spawn behavior.

Broker-managed endpoints now use native Ollama transport without a local pool slot.
Regression tests cover unary and streaming requests through `http://localhost:11434`.
The tests also prove that Omnius does not acquire a local pool slot.

The Telegram bridge had a separate token-count defect.
It read usage fields from the wrong level of each stream chunk.
It now reads the `usage` object, which corrects the false `~0 tok` display.

Each model-generation block now has a bottom `read mode` control.
The control expands all preserved thinking, content, tool-contract, and handling rows.
The control remains available after the block completes.
The Telegram chat fast path now streams into its block and closes the block on success or failure.

## Validation policy

The structural check must mask exact machine content.
It must then inspect only authored prose.

Use these hard checks for core prompts:

- No contraction
- No semicolon
- No unresolved placeholder
- Balanced delimiters outside protected content
- Maximum 20 words for a procedure sentence
- Maximum 25 words for a descriptive sentence
- One instruction in each procedure sentence
- Exact preservation of machine tokens

Do not use automatic passive-voice detection as a hard failure.
Do not use an incomplete dictionary as compliance proof.

## Remaining controlled work

Use a token manifest before any broad prompt conversion.
Record placeholders, schema keys, enum values, tags, sentinels, and tool names.
Require the same token set after each prose change.

Convert the remaining groups in this order:

1. Telegram router and reflection prose
2. Sub-agent and delegation prose
3. Recovery and completion prose
4. RALPH runner prose
5. Memory and compaction prose
6. CLI setup and recovery text
7. API descriptions and help text
8. Reference contracts

Run parser and snapshot tests after each group.
Do not use a repository-wide blind text replacement.
