XI.2 - Translation Layer Integrity (IX-0)

Date studied: 2026-02-05

Summary

IX-0 tests whether intent-to-authority translation can be implemented as a non-sovereign compiler with refusal, eliminating the translation layer as a hidden locus of decision-making.

Core result: Tooling need not—and need not be permitted to—exercise proxy sovereignty.

The Translation Sovereignty Gap

Most systems quietly assume translation is harmless. Intent gathered, normalized, “understood,” converted into executable representations by compilers/UIs/APIs/agents. When intent ambiguous/underspecified, system “helps”—choosing defaults, narrowing scope, resolving conflict. Responsibility for these choices rarely explicit, auditable, attributable.

IX-0 rejects this assumption.

The test: If converting intent into authority artifacts requires semantic interpretation, heuristic narrowing, or framing-based influence, then sovereignty fails before execution or governance begins. The tool, not the agent/institution, becomes de facto sovereign.

Failure Modes Targeted

  • Implicit defaults (missing/widened authority introduced silently)
  • Semantic guessing (choosing among multiple valid artifacts)
  • Coercive framing (UI/narrative influence altering outcomes)
  • Post-authorization mutation (previewed artifacts ≠ submitted)
  • Nondeterminism (identical inputs → divergent artifacts)
  • Opacity (artifacts not inspectable/diffable pre-authorization)

Conserved Quantity

Authority bound to explicit authorization under deterministic, non-privileged translation

Translation must preserve:

  • User-specified fields verbatim
  • Refusal on ambiguity or incompleteness
  • Diffability of all changes
  • Preview–submission integrity
  • Replay determinism

Any translation behavior that silently substitutes its own judgment violates conserved quantity.

Eight Conditions Tested

A: Identity preservation → All user fields preserved exactly, derived fields generated deterministically, no hidden transformation ✅ PASS

B: Minimal change sensitivity → Single-field intent change produced exactly one user-field diff + derived ID change, no unrelated fields altered ✅ PASS

C: Structural ambiguity refusal → Multi-entry scope produced TRANSLATION_REFUSED with explicit diagnostic, no artifact emitted ✅ PASS

D: Hidden default injection (adversarial) → Injected fields exposed by structural diff, classified as INJECTION_DETECTED ✅ FAIL_DETECTED (expected)

E: UI-level coercion (adversarial) → Framing payloads had no effect, output matched expected artifact exactly ✅ PASS

F: Replay determinism → Identical inputs with sequence reset produced identical artifacts and SHA-256 hashes ✅ PASS

G: Intent incompleteness refusal → Missing required fields produced TRANSLATION_FAILED with diagnostics, no artifact emitted ✅ PASS

H: Preview–submission integrity (adversarial) → Hash mismatch between preview/submission detected, classified as failure ✅ FAIL_DETECTED (expected)

Results

Positive:

  • Intent–authority translation can be deterministic
  • Refusal is first-class outcome, not error
  • Structural ambiguity and incompleteness detectable/enforceable
  • Adversarial injection and mutation observable
  • Tooling implementable without proxy sovereignty
  • Translation bit-perfectly replayable under canonical ordering

Negative (explicit boundary):

  • Does NOT establish translation is usable
  • Does NOT establish users understand artifacts
  • Does NOT establish values are correct
  • Does NOT establish governance succeeds
  • Does NOT establish systems are safe/aligned

Licensed Claim

Intent-to-authority translation can be implemented as a non-sovereign compiler with refusal, eliminating the translation layer as a hidden locus of decision-making.

Clarifications:

  • Necessary condition for reflective sovereignty, not sufficiency
  • Boundary calibration, not end-to-end system evaluation
  • Structural testability result, not narrative hand-wave

Key Insights

Translation vs Choice: IX-0 establishes translation can be honest. Doesn’t establish choice is easy, wise, stable. Bad choices remain possible. Ignorant authorization remains possible. Collapse by refusal remains possible.

These are properties of agency, not tooling.

Final excuse removed: IX-0 removes the “the tool had to decide” excuse. Subsequent Phase IX can now legitimately ask how values are encoded, how coordination occurs, how institutions persist/fail—without ambiguity about where authority is exercised.

Implications

IX-0 establishes necessary condition for reflective sovereignty: tooling can remain non-sovereign.

Authority translation now structurally testable boundary, not narrative hand-wave.

Remaining question: Not whether tools can be trusted—whether agents can own the choices that remain.

That question belongs to Phase IX.

Closure Status: IX-0 CLOSED — POSITIVE (IX0_PASS / TRANSLATION_INTEGRITY_ESTABLISHED)