Skip to content

Releases: th3vib3coder/vibe-science

Vibe Science v7.0 TRACE + TRACE+ADAPT V0

26 Mar 17:01
@th3vib3coder th3vib3coder
v7.0
3e4f672

Choose a tag to compare

View all tags
Vibe Science v7.0 TRACE + TRACE+ADAPT V0 Latest
Latest

Vibe Science v7.0 TRACE

This release closes the runtime loop of Vibe Science.

TRACE is the point where the plugin stops being just surrounding infrastructure and becomes an active research runtime: state is persisted, key scientific checks are enforced in code, and session continuity becomes real.

What is operational in this release

  • 7 Claude Code lifecycle hooks
  • SQLite persistence for sessions, claim lifecycle, R2 reviews, serendipity seeds, gate checks, citations, observer alerts, patterns, and benchmarks
  • runtime enforcement for DQ4, L-1+, L0, D1, and stop-time review blocking
  • citation extraction plus bounded DOI / PMID / arXiv verification
  • FTS5 / BM25 retrieval with fallback tiers
  • SessionStart context recovery, observer alerts, and cross-session pattern carry-over

TRACE+ADAPT V0

This release also includes TRACE+ADAPT V0.

This is not a new release codename: it is a deterministic adaptation layer inside v7.0 TRACE.

What it adds:

  • short carry-over hints injected at SessionStart
  • hints derived only from already-persisted runtime signals such as recurring gate failures and observer alerts
  • deterministic activation rules and cooldowns
  • no schema changes
  • no LLM calls in hooks
  • no protocol rewrites

Core boundary:

the harness may adapt, the truth model may not

Vibe Science can improve how it reminds and steers the workflow, but it does not redefine evidence standards, gate semantics, or scientific truth.

Why this matters

Vibe Science is built against a specific scientific failure mode:

an AI agent can analyze data, read papers, and produce a convincing conclusion faster than it can verify whether that conclusion is actually robust.

This release makes that problem harder to ignore by enforcing more of the research loop at runtime.

Validation

  • 169/169 end-to-end tests passing
  • smoke passing
  • readiness passing

Scope note

Not every methodological rule is hard-enforced in the plugin yet. Some methodology still lives at skill / review level rather than code-level enforcement.

But the core value proposition is now real:

  • harder to rush
  • harder to fake
  • easier to audit
Assets 2
Loading

Vibe Science v5.0 — IUDEX (Claude Code Edition)

16 Feb 22:26
@th3vib3coder th3vib3coder
v5.0.1
093327b

Choose a tag to compare

View all tags
Vibe Science v5.0 — IUDEX (Claude Code Edition)

Vibe Science v5.0 — IUDEX (Claude Code Edition)

Agentic tree search for scientific research. Structurally unbypassable adversarial review.

This is the Claude Code edition of Vibe Science v5.0 IUDEX — the full-featured version designed for Anthropic's Claude Code environment.

Platform

  • Target: Claude Code (Anthropic)
    • Installation: claude plugins add ./vibe-science/vibe-science-v5.0
    • Directory: vibe-science-v5.0/

What's Included

  • SKILL.md (~1,150 lines) — Complete skill definition
    • CLAUDE.md — Constitutional enforcement
    • 21 protocols — Full protocol suite
    • 9 schema files — JSON Schema gate validation
    • 8 assets — Supporting resources
    • 4 Claude Code hooks — System-level enforcement

Key Features (v5.0 IUDEX)

  • Seeded Fault Injection (SFI) — Known-false claims planted before R2 review
    • Blind Falsification Protocol (BFP) — Anonymized claim review
    • Structured Verdict Grammar (SVG) — Machine-parseable YAML verdicts
    • Judge Agent (R3) — Meta-reviewer scoring R2's review quality
    • Circuit Breaker — Auto-termination after 3 consecutive R2 failures
    • 10 Constitutional Laws | 27 Quality Gates | SOLO + TEAM modes

Runtime Modes

  • SOLO — Single agent, all roles in one context. Lower cost.
    • TEAM — Agent Teams with R2 in its own context. Maximum rigor.

Authors

License

Apache 2.0

Loading