Portfolio regimes

A second axis, parallel to the per-subject gap regimes and read the same way: a regime, not a score. It names what kind of multi-repo situation an estate is in, and the one structural move that fits it.

Regime	Shape	Steer
Federated	Shared habitat, bound across subjects, narrow gap spread	Hold. Maintain the binding; watch the pin ages.
Distributed	Shared habitat present, unbound or stale in several subjects	Bind what already exists — usually the cheapest uplift in the estate.
Fragmented	Per-subject habitats, high near-duplication, diverged	Extract the common layer.
Islanded	Per-subject habitats, little duplication, wide spread	Transfer knowledge. One team is materially ahead; move people, not files.

See Extract or bind for why the steers differ so sharply, and how to tell Fragmented from Islanded.

The four inputs

Selection is deliberately coarse. A regime a reader cannot check for themselves is worse than no regime at all.

Input	What it contributes
Declared habitats	Picks the axis. Declared → Federated / Distributed. None → Fragmented / Islanded.
Provenance proportions	inherited against inherited-unbound across assessed subjects. Separates a harness that binds from one that only exists.
Near-duplication	Whether per-subject control surfaces share an ancestor. Separates Fragmented from Islanded.
Gap spread	The range already computed for the spread section.

The decision

• Shared habitat, mostly inherited, narrow spread → Federated
• Shared habitat, inherited-unbound or stale pins in several subjects → Distributed
• No shared habitat, several near-identical but diverged control surfaces → Fragmented
• No shared habitat, little duplication, wide spread → Islanded

Where the inputs disagree — a declared habitat that binds nowhere and heavy duplication — the report names both, rather than forcing one label onto an estate that is genuinely in two states at once.

Regime claims carry their evidence

A regime is a claim about somebody's estate, so the report always states what produced it: the provenance counts, the spread, how many habitats were declared, and which duplicated artefacts were found.

An unevidenced classification is indistinguishable from a guess, and it gets argued with rather than acted on.

Near-duplicate detection

A duplicated control surface reads as all three of:

1. the same artefact name across subjects — HARNESS.md, .github/workflows/ci.yml, CLAUDE.md;
2. substantially overlapping content — the same sections, rules and structure;
3. divergent specifics — the copies have drifted apart.

All three together is reported as fragmented with lineage: the artefacts share an ancestor and have been maintained separately since. The report names where they diverged, because that is what an extraction has to reconcile — and it is the part nobody remembers.

Two subjects with the same filename and unrelated content are not duplication, and are not reported as such.

Declared variance

Divergence is not always drift. A manifest can declare it:

subjects:
  - id: legacy-batch
    path: ./legacy-batch
    justified_variance:
      - dimension: testing
        reason: COBOL batch; harness test tooling does not apply

Named dimensions are suppressed from drift findings and extraction candidates, and listed in the report as declared variance with their reason.

Listed, not hidden — a reader must be able to see what was excluded and disagree with it. Suppressed, because an estate where testing or observability genuinely cannot be uniform should not be nagged toward a convergence that would make it worse.

What a regime is not

It is not a score, a grade, or a ranking of estates. Like the gap regimes it tells you which way to invest, not how well you are doing. An Islanded estate is not worse than a Federated one — it needs different work.