ai-crucible 0.3.0

Diagnostic adversarial game for frontier LLMs — a policy-enforced kernel that mediates a Designer/Solver/Judge cycle, scores against a hidden oracle, and curates a Lab/Arena/Regression catalog.

日本語 | 中文 | Español | Français | हिन्दी | Italiano | Português (BR)

A diagnostic adversarial game for frontier LLMs — a measurement instrument that happens to be fun.

One Claude session (Designer) crafts puzzles targeting real, currently-observed capability gaps. Another (Solver) attempts them. A policy-enforced kernel mediates, scores against a hidden oracle, and curates a catalog through a Lab → Arena → Regression lifecycle. Puzzles are grounded in empirical signal — real GitHub issues, academic literature, observed failures in the field — not synthetic.

What makes it different

• Capability, not "cheating." AI Crucible distinguishes elegance and novelty (rewarded) from answer-bypass (penalized). Lateral thinking is a capability to measure, not a vice to punish.
• The instrument measures itself. Prompt framing is a first-class measured arm — the kernel runs the same puzzle under neutral / self_referential / social_standings framings and reports its own prompt-effect as a diagnostic.
• A sealed measurement boundary. Motivation and measurement never share a context window; the hidden oracle is graded out-of-band by a different model family with the agent's reasoning hidden. The model can't game what it can't perceive.
• Reliability by consistency. pass^k (all k independent trials succeed), Wilson intervals, and cross-family judge panels — built to report distributions, not point estimates.
• A living, durable catalog. Runs accumulate across sessions in an event-sourced, hash-chained log (the source of truth); tier state is a derived projection. Graduation Lab → Arena is abstention-aware — it promotes only on a confident cross-family verdict and escalates to the Designer otherwise; saturated puzzles are demoted to Regression, never deleted (an anti-flapping e-process), so the catalog becomes a capability-evolution timeline as the frontier moves.
• The differential payoff. Per puzzle, Claude's solve-rate vs the cross-family cohort's classifies the gap — Claude-specific (highest value) / LLM-general / Claude-strength — off a Newcombe difference interval with a first-class inconclusive class, so a small-N null is never dressed up as a finding.
• Measures any model family. Solvers using the native function-call protocol (not just the text action protocol) are first-class — a model that returns tool-calls instead of text actions Solves through the same governed sandbox.

Threat model (summary)

AI Crucible's primary adversary is the model under measurement, which may try to reach the answer key or to detect and game the eval. It touches: puzzle prompts/setup, the Solver's own confined working directory, and the model-provider API calls the operator configures. It does not touch: the oracle/answer key (graded out-of-band on a separate host the Solver cannot reach) or motivation "chrome" (rank/standings — never injected into the scored context). Permissions: model-provider keys via environment variables at runtime; no bundled secrets, no telemetry, no outbound calls of its own. Full disclosure — including where a boundary is defense-in-depth rather than a hard guarantee — is in SECURITY.md.

Architecture

AI Crucible is a thin policy layer on Inspect AI (UK AISI), not a from-scratch harness. A single AttemptState object is threaded Designer → Solver → (Critic) → Judge through one generate choke point, so every model and tool call is observable.

Module	Responsibility
puzzle_loader	Loads a puzzle directory (meta.json / prompt / setup_script) into Solver-visible state. Never touches the oracle.
sandbox	Narrow exec / read_file / write_file channel into a locked, network-less container.
roles	The five role slots (Designer / Solver / Critic / Judge / CohortSolver). Only Solver gets tools; Critic is interface-reserved, default-off.
budget_governor	Per-class tool-call + wall-clock budgets, displayed to the agent, enforced kernel-side; hard-kill on pathological loops.
oracle_scorer	Out-of-band grading: solved-and-no-regression against the hidden oracle (SWE-bench pattern).
judge_panel	Cross-family panel of model-scorers + reducer (PoLL) for novelty validation and bypass detection.
trace_writer	Per-attempt transcript in the Inspect EvalLog shape; large blobs stored by digest.
observability	Per-attempt → per-puzzle → per-model rollups; pass^k native.
catalog	Event-sourced durable persistence + the Lab → Arena → Regression lifecycle (abstention-aware graduation, anytime-valid saturation) + the differential typology. Builds on attestation's hash-chained log.
attestation	Cryptographic provenance (cosign + event-store) behind a typed subprocess boundary.

The sealed boundary runs in three tiers — Tier 1 scored context (deployment-shaped, framing-neutral), Tier 2 engagement framing (probed for contamination each release), Tier 3 chrome (rank/leaderboard — human-facing UI only, never in a context the model solves in). The full design rationale, with citations, is in docs/research-grounding.md.

Install

# As a Python library + CLI (PyPI):
pip install ai-crucible          # or: uv pip install ai-crucible
ai-crucible --help

# Or zero-prerequisite via npx — downloads a verified binary, no Python needed:
npx @dogfood-lab/ai-crucible --help

Run one diagnostic cycle — a Solver attempts a puzzle in the sandbox, graded out-of-band against the sealed oracle, emitting the pass^k / Wilson rollup:

# @family selects the adapter: no tag / @claude -> Claude (ANTHROPIC_API_KEY);
# any other @family -> a local Ollama model of that family (text OR native tool-calls).
ai-crucible run puzzles/seed-sulzbach-55252 --model claude-opus-4-8@claude --k 5

Each run accumulates in the durable catalog. Read and curate it, or run the eval-awareness boundary probe:

ai-crucible catalog list                 # tiers + per-puzzle differential typology + health
ai-crucible catalog show <puzzle-id>     # one puzzle: runs, transition timeline, differential
ai-crucible catalog graduate             # preview Lab->Arena->Regression transitions (--apply to commit)

# Eval-awareness gate: does behaviour diverge between deploy- and test-framing?
ai-crucible probe puzzles/seed-sulzbach-55252 --model claude-opus-4-8@claude --k 5

Research preview (v0.3.x). The judge panel's alt-test ω is still a circular model-jury bootstrap: validating it needs a round of ≥3 independent human annotators (the alt-test), which a single-human studio cannot staff — so that round is on ice by structural constraint, not neglect. Seated judges stay provisional, the composed panel escalates to a Claude Designer below quorum, and the instrument discloses this rather than faking human grounding. See the scorecard for the honest, non-cosmetic gate results.

Quick start (from source)

AI Crucible uses uv for environment and dependency management. Python 3.11+.

# Create the venv and install the dev + stats extras
uv sync --extra dev --extra stats

# Run the test suite (with the coverage gate)
uv run pytest --cov=ai_crucible --cov-report=term-missing

# Lint
uv run ruff check .

# One command: lint + tests + build + smoke
bash verify.sh

Documentation

• Handbook — guides, architecture, and reference.
• docs/research-grounding.md — design rationale, with citations.
• docs/gameplan.md — roadmap and open questions.
• SECURITY.md — threat model + honest residual-risk disclosure.

License

MIT. Public and pre-1.0 — see the CHANGELOG for version status.

---

_{Built by MCP Tool Shop · part of the dogfood-lab workshop for testing in the AI era.}