DE | EN
☰

Insight: The Adaptive Automation Process

Adapting automation requires a multi-stage agentic workflow driven by the following components and logic.

1. Main Objectives

There are two opposing main objectives governing the workflow:

  • Preserve data fidelity over schema flexibility: Attempt to reconcile unexpected structure or structure with unexpected syntax with the original schema.
  • Conservatively introduce probabilistic inference: Attempt to incorporate data with unexpected semantics into the original schema.

2. The Constitutional Prompt

This component establishes the immutable truth of the system.

  • Business Intent: Describes the original, invariant business intent.
  • Schema Definitions: Contains the original input/output schema definitions, including:
    • Syntactic rules.
    • Mandatory/optional structure.
    • Types, ranges, and memberships of data.
  • Benchmarking: Uses several manually selected pairs of optimal input/output pairs as a benchmark.
    • Described invariants are immutable to avoid a semantic drift.
    • The size of the benchmark is monotonically increasing (existing benchmarks are immutable, but new ones can be added).

3. The Input Delta Phase

The Input Delta Heuristic

  1. Validates against the original input schema definition.
  2. Accumulates validation errors and provides them to the Input Delta Prompt.

The Input Delta Prompt

This prompt compares the current input with the benchmark while taking validation errors into account. It classifies the resulting delta according to the following categories:

A. Continuity (No Delta)

  • Consequence: No adaptation necessary.

B. Evolution (Benign Delta)

  • Types:
    • Resolvable syntactic deviation.
    • Resolvable semantic deviation.
  • Allowed Categories:
    1. Terminological equivalence: Synonyms.
    2. Taxonomic/hierarchical alignment: Generalization or Specialization.
      • Constraint: Only allowed if the specific input implies the general category with near-100% probability (conservative inference).
    3. Compositional transformation: Split or Merge.
      • Merge: Concatenate forename and surname into full name.
      • Split: Split 1.99€ into amount and currency.
      • Justification: No data is invented; it is merely restructured based on semantic understanding of the components.
    4. Contextual derivation: Logical implication.
      • Example: Schema expects PaymentStatus; input is missing this field but contains a CancellationDate. Inference: PaymentStatus = Cancelled.
      • Conservative Limit: Must be strictly distinguished from missing data (which falls under corruption) by insisting on guaranteed implication.
    5. Semantic normalization:
      • Example: Schema expects ISO country code but input provides "Germany".
      • Conservative Limit: Only apply if there are no ambiguities.
  • Explicitly Excluded Categories:
    1. Probabilistic filling (Hallucination): Guessing a missing value based on likely patterns is forbidden (e.g., assuming country is Germany because the phone number looks German).
    2. Ambiguous resolution: e.g., "10/02/2025" may be either 10. February or 2. October.
    3. Intent modification: If the input is an "offer" and the schema expects an "invoice," this is a Domain Shift because the business intent differs.
  • Consequence: Adapt specification accordingly.

C. Corruption (Invalid Delta)

  • Types:
    • Unresolvable syntactic deviation.
    • Missing data.
    • Ambiguous data.
    • Contradictory data.
  • Consequence: Halt and request human review.

D. Domain Shift (Invalid Delta)

  • Types:
    • Deviating business entity.
  • Consequence: Halt and request human review.

4. The Legislative Prompt

Determines the implementation path based on the input analysis:

  • If Delta = 0: Reuse the existing implementation.
  • Else if Delta > 0: Implement adapted specification.
    • Apply implementation on current input (OutputA).
    • Apply implementation on benchmark input (OutputB).

5. The Output Delta Phase

The Output Delta Heuristic

  1. Validates OutputA against the original output schema definition.
  2. Validates OutputB against the original output schema definition.
  3. Accumulates validation errors and provides them to the Output Delta Prompt.

The Output Delta Prompt

  1. Compares OutputA with the benchmark output while taking validation errors into account.
  2. Compares OutputB with the benchmark output while taking validation errors into account.
  3. Verifies whether OutputA or OutputB violates the constitution.
    • On Violation: Halt and request human review.

6. The Confidence Scoring Prompt

Quantifies the system's confidence that the output is semantically valid by:

  • Taking all deltas and validation errors into account.
  • Making a weighted average of the self-assessment and the deterministic results of heuristics.