Abbey Research Artifact Creation Design

July 23, 2026

Tags: abbey-research • voice-analysis • architecture • provenance • validation

Abbey Research Artifact Creation Design

Summary

The three completed Voice Analysis artifact chains were never intended to establish a manual conversion queue.

They were prototypes that demonstrated the desired canonical lifecycle:

OBS → EVID → HYP → VAL

The useful next step is to turn the lessons from those prototypes into a controlled creation process.

Work Completed

Designed an Abbey Research workflow that separates:

• Candidate generation.
• Structural normalization.
• Deterministic sanitization.
• Metadata and citation validation.
• Human research review.
• Canonical promotion.

The design gives each research attempt a persistent run workspace containing its source fingerprints, model and prompt provenance, raw output, normalized output, candidate, validation results, and review record.

Why the Separation Matters

AI output is useful for proposing research artifacts, but it should not become canonical merely because its Markdown is well formed.

The workflow therefore treats a candidate and an accepted artifact as different things.

Only explicit promotion may write to canonical research directories.

Human Authority

Abbey can verify whether:

• A source identifier exists.
• A date and quotation match the corpus.
• Metadata is complete.
• Parent artifacts resolve.
• Scores and totals are internally consistent.

Abbey cannot deterministically decide whether:

• Evidence is representative.
• A score is persuasive.
• A hypothesis is the best explanation.
• Confidence is justified.
• A validation outcome should be accepted.

Those remain human review decisions.

Initial Implementation

The first implementation should orchestrate the existing observation workflow:

abbey research create --type observation

This vertical slice will prove run manifests, provenance, immutable raw output, safe working paths, and review-ready candidates before Abbey adds canonical promotion or downstream artifact stages.

Validation also found that the current abbey research run command fails on the Mac when no optional input files are supplied because its empty array is expanded under set -u. The first implementation phase must repair this portability issue before using the command as an orchestration primitive.

Lesson

The completed chains are most valuable as executable expectations.

Their purpose is not to encourage more hand-built artifacts. Their purpose is to teach Abbey how to create future artifacts safely, consistently, and transparently.