SISO People Graph

SISO KnowledgeSISO Knowledge owns canonical-service governance for durable graph and index semantics, while the People Graph keeps independent repository identity, release lineage, and data-plane boundaries.

SISO FoundryFoundry may discover and evaluate source universes and emit source observations, but it is not the canonical truth store for person identity, claims, or merges.

SISO Evidence EnginesEvidence Engines transform and adjudicate source-grounded observations into claims or knowledge; they do not silently mutate canonical identity.

SISO Book LibraryThe Book Library is one independently released source and contract surface that can emit people, work, role, subject, and locator observations for the graph.

The exact source tree contains one v2 SQL schema, identity and enrichment loaders, source-to-graph loaders, and a read-only query router; no generated database is selected by this Work record.Open receipt ↗

The public program spine documents ownership, control-plane and data-plane boundaries, observation interchange, parallel lanes, and honest release claims.docs/people-graph-program/README.md

The published graph-v2 asset was retrieved to a clean directory and queried on 2026-08-04, reporting 339,217 people, 901,524 person-to-work edges, 1,272,495 person-to-person edges, 1,131 organisations and 442,499 external identities across five domains; graph-v1 reported 280,708 people with no relational layer. CORRECTION 2026-08-07: this is a single unreplicated observation, recorded as a repository claim rather than a verified measurement. The figures survive only as this prose summary -- the value 339217 appears in no file on any branch of siso-people-graph, so no query log or receipt backs it. Neither of the two independent audits could re-count them (red team: 'No production SQLite assets were opened. Dataset-wide counts and collision rates are therefore not claimed'; first-principles: 'Published row counts are therefore treated as repository claims, not independently re-counted observations'). Separately, red-team findings PGRT-002/003/009/016 establish that person rows are not defensible as distinct-human counts under name fusion, 0.98 attribute auto-accept and CJK label collapse. Re-counting against the release asset, with the query recorded as a receipt, is required before any of these magnitudes is cited as verified.local:people_graph_v2.sqlite:counts-2026-08-04

The relational layer is derived from 32,243,556 declared package dependency rows, of which 27,847,334 runtime edges were kept and 4,396,222 development and build edges deliberately excluded as a different claim. An edge therefore means one person's software does not build without another's, which is machine-checked rather than inferred. It answers who is load-bearing rather than who is popular: the most depended-upon person is relied on by 45,037 distinct others. Two earlier attempts at this layer were measured and rejected, one yielding 720,107,620 meaningless co-membership pairs and one yielding five usable edges.local:people_graph_v2.sqlite:person-person-2026-08-04

Membership is evidence-led rather than popularity-led. 58,509 people were admitted because a third party measured use of their published work, not because their repositories cleared a 100-star threshold; the preceding population matched that threshold to within nine people. Among those admitted are maintainers of infrastructure with billions of recorded installs whose repositories fall below the threshold, which is the case against it. People whose output is not yet linked are recorded as tracked rather than omitted.local:people_graph_v2.sqlite:admission-rule-2026-08-04

The schema encodes the graph's design laws: satellites are referenced rather than duplicated, so content_ref holds a repository name or catalog identifier and the bytes stay in the domain that owns them; every cross-source match is an explicit, evidenced, reversible identity_claim rather than a silent merge; role sits on the edge rather than the person; and person.state distinguishes a tracked person with no linked output from a broken row.Open receipt ↗

Fifteen enrichment rounds are recorded with the command and raw output behind every number, including three negative results that stopped expensive work: a GitHub-to-book identity stitch capped at 419 candidate people and yielding zero plausible matches, naive person-to-person co-membership measured at 720,107,620 pairs and rejected, and a repository source-diversity signal shown to be a restatement of stars once controlled for.Open receipt ↗