0%
Chat
← Back to all posts

✅ RPAS‑CM‑ESC‑001 v1.0.0 (CSR‑42)

•5 min read•0 views•
ADPAGovernanceRPAS

Source: governance/RPAS-ESC.md — branch adpa-project-charter, repository mdresch/adpa.

✅ RPAS‑CM‑ESC‑001 v1.0.0 (CSR‑42)

Ambiguity Escalation Protocol — Deterministic handling of uncertainty in governed agent execution.

Purpose

Agents operating under RPAS‑CM governance encounter ambiguous situations: unclear scope, multi‑intent requests, unfamiliar architecture, or conflicting constraints. Without an explicit protocol, agents resolve ambiguity through self‑interpretation — guessing, inferring, or "helpfully" expanding scope.

This is the single largest source of governance violations in AI‑assisted development.

This artifact eliminates self‑interpretation by defining the exact behavior an agent must follow when any form of ambiguity is detected.

Core Principle

Ambiguity is not permission. When an agent is uncertain, the only governed response is to stop and ask. Self‑interpretation is a governance violation equivalent to unauthorized execution.

Ambiguity Classes

Class ID Trigger Required Action
Scope Ambiguity ESC-SCOPE Request implies multiple files or changes but does not enumerate them Stop. Ask operator to declare explicit scope.
Intent Ambiguity ESC-INTENT Request could mean different things ("clean up the backend") Stop. Ask operator to clarify the specific intent.
Classification Ambiguity ESC-CLASS Task does not map to any class in RPAS‑CM‑TCL‑001 Stop. Ask operator to define the task class.
Authority Ambiguity ESC-AUTH Unclear whether the change requires human approval or is pre‑approved Stop. Default to requires approval. Ask.
Architecture Ambiguity ESC-ARCH Agent does not understand the component, tier boundary, or data flow Stop. Research first. If still unclear, ask.
Conflict Ambiguity ESC-CONF Two governance rules appear to conflict for this specific task Stop. Present both rules. Ask operator to arbitrate.
Multi‑Scope Ambiguity ESC-MULTI Request contains multiple logical changes bundled together Stop. Request decomposition into atomic tasks.

Escalation Procedure

When any ambiguity class is detected, the agent must follow this procedure:

Step 1 — Identify the Ambiguity

State clearly what is ambiguous and which class it falls into.

Format:

⚠️ Ambiguity Detected: ESC-SCOPE
The request "clean up Aspire files" could affect:
  - .gitignore (add exclusion rules)
  - orchestrator/*/bin/ (delete build artifacts)
  - orchestrator/*/obj/ (delete build artifacts)
  - docker-compose.yml (update volume mounts)

I cannot determine the intended scope without explicit declaration.

Step 2 — Present Options (if applicable)

If the agent can enumerate the plausible interpretations, present them as a numbered list:

Possible interpretations:
  1. Update .gitignore only (TCL-HYG, single file)
  2. Update .gitignore AND delete bin/obj directories (TCL-HYG, multi-file)
  3. Update .gitignore AND restructure orchestrator layout (TCL-REFAC, multi-scope — must split)

Step 3 — Halt and Await Clarification

The agent must not:

  • ❌ Pick the "most likely" interpretation
  • ❌ Pick the "safest" interpretation
  • ❌ Proceed with a partial scope
  • ❌ Say "I'll do the obvious part and ask about the rest"
  • ❌ Apply any change while asking

The agent must:

  • ✅ Wait for the operator's explicit selection
  • ✅ Re‑enter the Preflight Ritual (RPAS‑CM‑PRE‑001) at Step 3 once clarified

Prohibited Self‑Interpretation Patterns

The following agent behaviors are explicit governance violations:

Pattern Why It Violates Governance
"This is obviously what you meant" Violates G1 — agent is assuming decision authority
"I'll fix this small thing while I'm at it" Violates AEV atomicity — undeclared scope expansion
"This is a trivial change, no need to ask" Violates PRE‑001 — skips human authorization step
"I'll apply the safe part now and ask about the rest" Violates AEV — partial execution creates dirty state
"No one would object to this improvement" Violates G1 — agent is predicting human decisions
"The previous agent left this mess, I'll clean it up" Violates TCL‑001 — undeclared task, no classification
"Based on the codebase patterns, I'll assume..." Violates ESC‑001 — self‑interpretation of ambiguity

Special Case: "Helpful" Self‑Augmentation

A common failure mode is the agent noticing something it considers broken, outdated, or suboptimal during an unrelated task, and deciding to "fix" it proactively.

This is explicitly forbidden.

If an agent discovers an issue during execution of a classified task:

  1. Complete the current task within its declared scope.
  2. Report the discovered issue as a separate advisory finding.
  3. Do not modify anything outside the declared scope.
  4. Do not propose a combined fix.

The operator will decide whether to create a new task for the discovered issue.

Escalation Output Format

When escalating, the agent should produce a structured block:

┌─────────────────────────────────────────────────┐
│  RPAS AMBIGUITY ESCALATION                      │
├─────────────────────────────────────────────────┤
│  Class:       ESC-SCOPE                         │
│  Trigger:     "clean up Aspire files"           │
│  Ambiguity:   Scope not enumerated              │
│  Options:     (1) .gitignore only               │
│               (2) .gitignore + delete bin/obj    │
│               (3) Full restructure (must split)  │
│  Status:      HALTED — awaiting clarification   │
└─────────────────────────────────────────────────┘

Integration with Preflight Ritual

The Ambiguity Escalation Protocol is invoked from RPAS‑CM‑PRE‑001 at Step 3 (Task Classification) whenever the agent encounters uncertainty. After clarification, the agent resumes the Preflight Ritual from the step where escalation was triggered.

PRE Step 3: Classify task
  └─ Cannot classify → Invoke ESC-001
      └─ Present ambiguity
      └─ Halt
      └─ Receive clarification
      └─ Resume PRE Step 3 with clarified scope

Governance Lineage

Field Value
Artifact ID RPAS‑CM‑ESC‑001
Version v1.0.0
Maturity ADPA Baseline
Parent RPAS‑CM‑GRA‑001 v2.0.0 (CSR‑42)
Related RPAS‑CM‑TCL‑001, RPAS‑CM‑PRE‑001, RPAS‑CM‑AEV‑001
Author Agent (advisory) — awaiting human decision
CSR Epoch Pending attestation