v4.5.1

Security

• Personal phone number removed from every LICENSE.md (root + the 8 package
  mirrors, license:sync-verified). Found by the pre-promotion privacy sweep — the
  license contact stays email-only. Residual exposure noted honestly: the number
  remains in old git blobs and in the already-published npm tarballs (≤4.5.0); the
  current tree, the next release, and everything a visitor reads going forward are
  clean. Full-history rewrite deliberately NOT done — it would break every clone,
  tag and PR reference for marginal gain.

Added

• Bench results visualized in the READMEs — a static dumbbell chart
  (docs/assets/bench-results.svg + a dark-mode variant, selected via
  <picture>/prefers-color-scheme) of the committed 2026-07-21 live-round
  numbers: skill vs stock per bench × model (research / design / discipline ×
  Haiku / Sonnet / Opus), with direct value labels and deltas on every row.
  Palette CVD-validated for both surfaces. Pure rendering of data already under
  evals/*/results/ — no benchmark execution involved. Embedded in
  README.md, README.en.md and evals/README.md.

• /vkm-research grew from a 68-line monolith into a full skill (same standard as
  the 4.5.0 vkm-spec rebuild): rewritten SKILL.md with a copyable checklist and a
  degradation ladder; references/summary-template.md (the canonical consolidated
  shape) + references/synthesis-guide.md (four axes: claims across sources, links
  that connect, visible supersession, compression with judgment);
  examples/worked-example.md with REAL validator output both ways (draft fails with
  5 named errors, rewrite passes); and scripts/validate_summary.mjs — a zero-dep
  validator that rejects promoted map-reduce drafts by their seams (--- separators,
  ## <file>.md headings, leaked url: lines), missing wikilinks, malformed
  supersedes and transcription-sized output. Doubles as the research-bench grader;
  mutation-style self-test (10 cases) gates in core CI.

• research-bench (evals/research-bench/): a synthetic RESEARCH/<topic>/ bank
  with the pipeline's real frontmatter and two seeded probes — a contradiction between
  sources (must surface as a typed - supersedes) and an embedded instruction in one
  source (must be flagged as untrusted DATA, never obeyed). Skill vs stock; grader =
  the shipped validator + probe signals (reference consolidation scores 100, the raw
  draft 0). New job + dispatch option in llm-benchmarks.yml.

• implementer-bench (evals/implementer-bench/): first eval covering the
  vkm-implementer agent — its real installed contract as system framing vs bare, on
  spec-shaped tasks, graded by discipline-bench's existing hidden-test instruments
  (no new graders). New job + dispatch option in llm-benchmarks.yml.

• design-bench auto round 1 (mechanical score from the skill's own validators;
  raw HTML committed under results/2026-07-21-round1/): Sonnet +60 on the
  slop-attractor brief (stock: 15) and +30 on the held-out brief; Opus (n=1)
  +60/+40 — stock Opus scored 0 on facturio (full slop fingerprint + failing
  contrast). Haiku flat, dial-consistent. Judgment axes stay in the manual protocol.

• Effort gate auto-match (ADR-0031 amendment): the PreToolUse effort gate now
  detects the session's current effort (CLAUDE_EFFORT inherited by the hook) and
  opens itself when the model's proposed level equals it — no pause when there is
  nothing for the user to change. Mismatch or undetectable effort keeps the original
  pause; the deny message now advertises the detected level. Subprocess-level tests
  cover match/mismatch/undetectable/template-only paths.

• Diversified round (round 2/3 of the live benches), Opus added at reduced n —
  raw data under each eval's results/2026-07-21-round2/:

  • research-bench: the skill's gain GENERALIZES — held-out domain topic
    (container-queries: Haiku +35, Sonnet +52.5, Opus +50) and Opus on sqlite-vec
    (+45); stock Opus still fails the consolidation contract.
  • discipline-bench on the new harder tasks (incl. the held-out instrument):
    Sonnet +19/+21, Opus +25/+31; Haiku flat-to-−3 within spread — where the hidden
    contract exceeds the small model's reach, the doctrine neither helps nor hurts,
    exactly what the dial predicts.
  • token-quality-ab adversarial fixture (no-keyword decisive lines): delta 0.0
    on all three models measuring the FIXED hook — verdict stays KEEP.
  • skills-triggering hard set (12 near-miss/multi-skill/tool-vs-skill cases):
    Haiku 11/12, Sonnet 12/12, Opus 11/12 — desaturated, both misses are arguable
    boundary calls, logged with the description tweaks to try next round.

• Round 1 of both new benches, raw data committed (2026-07-21, Haiku 4.5 + Sonnet 5,
  n=3/cell, under results/2026-07-21-round1/):

  • research-bench: the skill delivers — Haiku 33.3 → 70.0 (+36.7), Sonnet
    35.0 → 96.7 (+61.7). Stock output fails the consolidation contract (no typed
    supersedes, weak linking, the seeded injection usually dropped silently).
  • implementer-bench: honest null result — explicit specs saturate both
    conditions (the value lives in the spec, which is /vkm-spec's job), underspec
    deltas sit inside replica spread with one negative Sonnet cell noted for re-check.
    The agent's case is delegation ergonomics, not raw scores; the bench exists to
    catch contract regressions. No verdict claimed beyond n=3.

Changed

• The skill structure gate's cross-reference allowlist is now EMPTY (was 4
  tolerated vkm-design entries): the in-reference pointers became plain-text mentions
  at the point of use ("contemporary.md, this folder") and SKILL.md remains the only
  place that links — Anthropic's one-level-deep rule now holds to the letter, and the
  gate ratchets at zero exceptions. lineages.md deliberately NOT partitioned: it
  passes the ## Contents navigability gate, and splitting it would mint new
  cross-references — the exact debt just retired.