Reproducible context evidence · MIT + CC BY 4.0

Context assembly manifest fixture

Inspect how one fixed task filters nine synthetic candidates by tenant, scope, data class, and freshness before it ranks or budgets evidence. The bundle includes strict TypeScript, sixteen tests, the reviewed manifest, and its SVG diagram.

The run tests context assembly policy, not model quality. Its token counts are supplied fixture values and do not represent any provider tokenizer.

Context contract passing through policy gates into typed evidence sections and a context manifest

Files

context_assembler.ts Strict TypeScript policy and manifest builder fixture.ts Synthetic contract and nine source candidates context_assembler_test.ts Policy, failure, injection, and determinism tests run_fixture.ts Deterministic manifest runner context-manifest-2026-08-27.json Reviewed happy-path manifest render_manifest.ts Dependency-free SVG renderer context-assembly-manifest-2026-08-27.svg Diagram generated from the saved manifest evidence_artifact_test.ts Stored JSON, SVG, and credential-pattern checks evidence-run-2026-08-27.md Commands, results, hashes, and evidence limits README.md File map and independent rerun commands

Policy before ranking

Cross-tenant, forbidden, unauthorized, and stale candidates are rejected before relevance or authority can affect selection.

Fail-closed evidence

Missing, stale, or oversized required evidence stops assembly before a model payload exists.

Inspectable decisions

Included and excluded sources retain versions, trust labels, token counts, and machine-readable decision reasons.

Verified scope

Sixteen tests cover required-source failures, freshness, tenant and scope isolation, forbidden data, distractor overload, budgeting, prompt injection, deterministic output, stored evidence, and the SVG. The happy path selected 424 of 430 supplied source tokens.

MIT license CC BY 4.0 license Read the context engineering guide