Predictive Mapping — Why We Don't Say "AI‑Powered"
The Phrase We Refuse to Use
Walk through any data-integration vendor's booth and you will hear the same phrase: *AI‑powered mapping*. The implication is that a model reads two schemas, decides how they correspond, and writes the transformation. In a regulated environment — healthcare, life sciences, federal — that sentence describes an unsigned decision. And an unsigned decision cannot be deployed.
Magellan does not do autonomous mapping. We do predictive mapping: a deterministic pipeline of proposal, named-SME approval, and signed commit, built on primitives auditors already understand.
The Three-Step Pipeline
Predictive mapping is exactly three steps, in this order, with no shortcuts:
1. Propose. The Semantic Match Engine embeds source and target schema elements, ranks candidate mappings by calibrated cosine similarity, and surfaces the top‑k with a confidence score. Nothing is committed.
2. Approve. A named subject-matter expert — by name, with their organizational role recorded — accepts, rejects, or revises a proposal. The proposal becomes a *decision* only at this step.
3. Sign and append. The decision is electronically signed and appended to a SHA‑256 hash‑chained ledger keyed by ULID. The previous-entry hash makes tampering detectable on the next read.
The engine *proposes*. It does not *decide*. That distinction is not stylistic — it is the entire reason regulators accept the output.
What Lives in a Decision Record
Every accepted mapping becomes a decision record. The record carries enough provenance to be replayed years later, by an auditor who has never spoken to the original SME:
| Field | Example |
|---|---|
decision_id | 01HW3F7Y9KQR8VZP4N2C5MABCD (ULID) |
approved_by | dr.sarah.chen@health.example (CMIO) |
input_hash | sha256:7f3a…b91c (source.field + target.field) |
model_version | semantic-match-v2.4.1 |
confidence | 94 (integer 0–100) |
prev_hash | sha256:c2e1…a4d8 (chains to previous decision) |
signature | Electronic signature, 21 CFR Part 11 §11.50 |
Confidence is an integer 0–100, not a percentage with decimals. It is a calibrated score the SME uses to triage their queue — not a probability the system uses to act unilaterally.
Pinned Versions, Not Floating Mappings
Schemas and mappings version independently. A mapping at v1.7.0 pins source-schema v2.4.1 and target-schema v3.1.0. When the target schema upgrades to v3.2.0, the mapping does not silently migrate. It requires a new, signed decision against the new schema version.
This is the rule that makes predictive mapping reproducible:
- Determinism by construction. Re-running the pipeline against the pinned schema versions yields the same output, byte for byte.
- No silent recovery. Schema drift triggers a pinning violation that blocks deployment until an SME signs against the new versions.
- Auditor-friendly replay. An auditor can reconstruct any historical transformation from the ledger plus the pinned schema artifacts — without access to Magellan's running system.
The only exception is the moment of import: when a mapping spreadsheet is first ingested, the source schema(s), target schema(s), and the new mapping all start at the same version number. After that single act, they version independently and the mapping pins what it was authored against.
Failure Modes Are Explicit
In an autonomous-AI system, failure modes are usually implicit — the model produces an output, the pipeline runs, and you discover the problem in production. In a predictive-mapping system, every failure has a defined response:
| Trigger | System response |
|---|---|
confidence < 70 | Dual‑SME approval required |
| Schema drift detected | Pinning violation · deploy blocked |
| Ledger hash mismatch | Tamper alarm · read fails |
| Missing SME signature | Decision rejected at commit |
There is no silent recovery. There is no "the model figured it out." If the system cannot defend the decision in front of an auditor, it does not commit.
Why Regulators Accept This
The standards an auditor cares about — HIPAA §164.312(b), 21 CFR Part 11 §11.10(e), FedRAMP AU‑2 and AU‑10, SOX §404, GxP ALCOA+ — all require some combination of:
- A named human accountable for the decision
- An electronic signature with non-repudiation
- An audit trail that cannot be silently rewritten
- The ability to reconstruct the decision from preserved evidence
Predictive mapping satisfies all four by design. Autonomous mapping satisfies none of them.
That is why we will keep saying *predictive mapping*, *Semantic Match Engine*, *named‑SME approval*, and *hash‑chained ledger* — and why we will keep refusing to say *AI‑powered*. The words are the contract.
To see predictive mapping in action against a real mapping spreadsheet, request a demo or read the interactive white paper. For a deeper look at the audit primitives underneath, start with Hash-Chain Integrity for the Magellan Audit Ledger and ULID vs UUID for Audit Ledgers.
*Magellan AI Corporation — The system of trust for data transformations.*