Scale Assumption Register¶

Documentation register — instantiates Scale-Bridging Translation

A living ledger of the assumptions a scale translation rests on — each tagged with what must stay true, how far the supporting evidence can travel, where the rule is valid, and who owns it.

Scale Assumption Register is the standing ledger that makes a translation's load-bearing beliefs explicit and owned. Every row is one assumption the bridge depends on — a source condition expected to persist at the target scale — tagged with its current evidence status, an explicit statement of how far that evidence is licensed to travel, the boundary within which the assumption is claimed to hold, and a named owner accountable for revisiting it. Its defining move is turning "it should still work at scale" into a diagnosable list: when the translation later fails, you read down the register to find which assumption broke, instead of relitigating the whole bridge. Where the Construct Mapping Table records what terms mean, this register records what must stay true.

Example¶

A bank has a fraud-detection rule that works well in its home market and now wants to run it across all regions. Rather than shipping it, the risk team opens a Scale Assumption Register. Row one: "transaction-timing patterns resemble the home market" — evidence status: unverified abroad; transport rule: home-market backtest may size the effect but may not confirm it elsewhere; validity boundary: markets with comparable card-present mix; owner: regional risk lead. Row two: "customers dispute false positives through the same channel" — evidence: contradicted in two markets with no call center; boundary: excludes those markets; owner: ops. Row three flags an assumption about holiday-season volume that must be retested, not merely carried over.

The register does not decide anything, and it does not test anything. Its whole payoff comes months later: when the rule misfires in one region, the postmortem opens the register, finds the timing-pattern assumption was never verified there, and fixes that row — instead of concluding the model is "just broken" and starting over.

How it works¶

What distinguishes the register from a generic risk log is that every row is scoped to scale:

Persist vs. redesign. Each assumption is classified as a source condition expected to hold at the target scale, or one that must be redesigned, retested, or bounded before transfer. This split is the register's core work.
An evidence-transport tag per row. Beside each assumption sits an explicit rule for how the source-scale evidence may be used at the target scale — sizing versus confirming, illustrative versus decisive — so pilot data and lab results are not quietly over-cited.
A validity boundary per row. Each assumption carries the conditions under which it is claimed, and named owners who must revisit it as evidence lands.

Tuning parameters¶

Assumption granularity — a few headline beliefs or an exhaustive enumeration. Exhaustive registers catch subtle dependencies but decay into unread paperwork.
Evidence-status scheme — how strictly a row is graded (verified / sized-only / assumed / contradicted) and what status is required before transfer proceeds.
Boundary tightness — how narrowly each assumption's validity is scoped; tight boundaries are safer but multiply rows and exceptions.
Ownership and cadence — whether every row has a named owner and a review trigger, or the register is written once and abandoned.

When it helps, and when it misleads¶

Its strength is diagnosability: a failure at the target scale becomes a specific broken row rather than a mystery, and the "must-retest" rows stop a team from silently importing a source condition that no longer holds.

Its failure mode is the register that is written to satisfy a checklist and never opened again — assumptions go stale, owners move on, and the artifact gives false comfort precisely because it looks thorough. It is also easily gamed: the most convenient assumptions get graded "verified" on thin evidence, letting a decision already made pass a governance gate. The register lists what must hold and how far evidence travels, but it does not itself run the test that would confirm any row — mistaking the ledger for evidence is its classic misuse. The discipline is live ownership: every row has a person and a trigger, and the "sized-only" and "contradicted" tags are honored rather than upgraded under pressure.

How it implements the components¶

assumption_log — the register is the log: one row per assumption, split into persist / redesign / retest.
evidence_transport_rule — each row's tag stating how source evidence may and may not be used at the target scale.
validity_boundary — each row's stated conditions of applicability, marking where the assumption is claimed and where it is not.

It logs assumptions but does not remap the constructs those assumptions are about — that is the Construct Mapping Table — and it never runs a test to close a row; correspondence checking belongs to Multi-Level Model Check and Stratified Target-Scale Rollout.

Instantiates: Scale-Bridging Translation — the register keeps the bridge's load-bearing beliefs explicit, owned, and diagnosable.
Sibling mechanisms: Construct Mapping Table · Pilot-to-Scale Translation · Lab-to-Field Translation · Micro-to-Macro Model Translation · Macro-to-Micro Operational Translation · Individual-to-Population Policy Translation · Ecological Scale Translation · Multi-Level Model Check · Stratified Target-Scale Rollout · Team-to-Organization Process Translation