[Naoray]

What & why

Append independent CC0 generator-v15 cells for numeric JSON person-linked IDs, bare JSON null/true/false, and prose placeholders following quoted labels. A quote around a key or label does not quote its value. This closes a benchmark blind spot before measuring the recognizers in #773.

Each partition adds 30 A positives, 115 D counterweights and 30 R repeat documents. Numeric values are JSON integers; each keyword document has a distinct public tool-turn field. Dev/test values and template IDs are disjoint. Cell definitions derive from JSON semantics and agent record traffic, with no recognizer-pattern imports. All v3–v14 corpora and frozen contracts remain byte-identical. The full partition is A/D/R 3,728 / 2,144 / 280 documents.

Benchmark evidence

Re-measured every displayed release on the current v15 harness (56b20a2575d543da898a5aa158fb1e7afa908176) against its own freshly built tagged binary. Each v0.15.x and v0.16.0 arm uses its own fresh setup policy. v0.14.0 retains the documented common v0.15.1 model settings, tokenize-only actions and composite-source splitting.

Release / arm	Leaked A / D / R bytes	FP A / D / R bytes	Refused A / D / R
v0.16.0 / policy-file	17135 / 0 / 9894	1966 / 6898 / 479	0 / 0 / 0
v0.15.1 / policy-file	33228 / 0 / 10034	2823 / 8058 / 463	0 / 0 / 0
v0.15.0 / policy-file	33401 / 0 / 10034	2823 / 8058 / 463	0 / 0 / 0
v0.14.0 / full-stack-kiji-resolve	43733 / 0 / 10244	10161 / 7846 / 323	0 / 0 / 0
v0.14.0 / pass2-ner	44476 / 0 / 10270	2282 / 5977 / 200	0 / 0 / 0

All v0.16.0 and v0.14.0 NER restores are exact with valid manifests. The existing legacy limits remain: v0.15.x have four non-exact restores and four invalid manifests in A; v0.14.0 Kiji has 54 / 15 / 1 non-exact restores and no invalid manifests. No refusals occurred. These are historical measurements, not a gain claim for #773.

Evidence: docs/reference/benchmarks/agentic-adjacency-v15-history.json, four agentic-layers-v15-*.json.gz scorecards and four bound observations-agentic-v15-*.jsonl.gz records. The ledger pins binary revisions, binary/scorecard digests, corpus identity and legacy semantics. v14 evidence stays intact. Test corpus SHA-256: 0a5f6576f2d5f5c8143628e88ec9d2985197fa3bd74c3823de86f611979c09ba.

Commands: each tagged CLI's setup --non-interactive --force, release clean_for_bench builds (Nym for v0.15.x/v0.16.0; Kiji for v0.14.0), current agentic_layers.py measure, and render_agentic_adjacency_doc.py --generator-version 15 --record .... Primary corpus inputs/scoring and frozen competitor-comparison inputs did not change, so their committed measurements remain applicable; this remeasurement updates A/D/R only.

Machine: MacBook Pro, Apple M5 Max, 18 cores, 64 GB, macOS 26.5. Byte evidence only; concurrent slots supply no latency evidence.

Validation

• CI-equivalent unittest discovery: 1,025 passed. Initial subprocess protocol/deadline timing failure preserved; isolated retry and final full run passed without a timeout change.
• CI-equivalent pytest groups: 274 passed, 3 skipped, 10 subtests passed.
• Focused syntax tests: 4 passed; agentic-layer tests: 248 passed.
• All eight docs checks passed; 183 renderer tests passed; changed-Markdown scrub passed for three files.
• Additional inherited Keep the MCP tier mutation probe from discarding uncommitted edits #780 tests: ten passed, one interrupt-matrix case timed out on macOS Bash 3.2/Python 3.13. Reported separately; no source or timeout changes.
• Signed commits with DCO; synthetic fixtures only.

Axes and limits

No runtime detection or restore change is introduced. The new counterweights strengthen precision evidence; bound observations retain deterministic, auditable provenance. Every numeric ID in this slice has six digits. Shorter unlabelled repeats remain outside this slice. Generator v15 has no curly or typographic value-quote cases and no rejected-placeholder-then-valid-cue cases; these results do not establish coverage for either behavior. Independent positive, repeat and benign counterweight cells for both boundaries are deferred. Both limitations are disclosed in the scalar-syntax document and adjacent README coverage paragraph.