Zetetic AI · Verification-first · Open source

Agents that say
“I don’t know.”

97 reasoning patterns drawn from primary sources — Dijkstra, Curie, Pearl, Peirce, Feynman, Cochrane. Each agent cites its primary paper and documents the conditions under which it must refuse. A pre-commit hook blocks any constant lacking a source. Not a prompt library. A methodology with commit-time enforcement.

> /plugin marketplace add cdeust/zetetic-team-subagents
> /plugin install zetetic-team-subagents

Full distribution for Claude Code: the plugin installer copies agents, skills, hooks, and tools into ~/.claude/. A portable evidence-synthesis package runs in Codex and Gemini CLI.

Zetetic Agents in 90 seconds. Silent. Generated from cdeust/zetetic-team-subagents at v2.46.0 (); every figure it prints is read from site-facts.json.
What the film says, scene by scene
  1. A role prompt is not a method. Each agent is a sourced reasoning pattern whose body documents the conditions under which it refuses. Those conditions are intent statements, not enforced contracts.
  2. No source, no commit. The commit hook blocks unsourced certainties at any profile. $ git commit -m "tune retry backoff". UNSOURCED (error) retry.py:1: # It always works. MAGIC_NUMBER (error) retry.py:2: DELAY = 2.741592. Profile: strict (staged mode) FAILED: 2 blocking violation(s). BLOCKED: Zetetic violations in staged files. Output as printed in the README. Magic numbers and bare TODOs warn by default; they block under the strict profile.
  3. A new test must fail first. At push, the tests a diff adds are run against the old code. They fail: as they must: the tests detect the change. They pass: VACUOUS: a warning by default, blocking under strict. No verdict: INCONCLUSIVE, never a pass.
  4. Mechanical thresholds, judgement left to you. A file over 500 lines blocks; function, class, parameter and nesting limits block for recognized languages. Judgment rules (SRP, OCP, LSP, ISP) are deliberately not mechanized. A prose scan names candidates for a rewrite; it does not judge quality. Thresholds are tunable per repository in .craftsmanship.conf.
  5. Choose a method by the shape of the problem. You name the shape; a skill loads the reasoning procedures that fit. problem-shaped skills: 15. sourced reasoning patterns: 97. problem-shape triggers: 650+.
  6. A build loop that does not certify itself. The loop drafts and converges a candidate; a deterministic gate decides. A criterion passes if and only if its shell command exits 0. It fails closed, never touches main, never pushes.
  7. Zetetic Agents. 97 agents, 86 skills, 28 lifecycle hooks. Quality gates on Claude Code and Codex, portable skills on Codex and Gemini CLI. $ claude plugin marketplace add cdeust/zetetic-team-subagents $ claude plugin install zetetic-team-subagents

Sources: Sources: README and tools of cdeust/zetetic-team-subagents at the v2.46.0 tag (3caec26). The blocked-commit output is the README’s composite of the checker and the hook wrapper. Figures: ai-architect.tools/site-facts.json.

The system enforcing its own standard

When you commit code, the hook runs tools/zetetic-checker.sh against the staged changes. Unsourced absolute claims and magic floats fail the gate.

zsh — pre-commit
$ git commit -m "tune retry backoff"

UNSOURCED    retry.py:1: # It always works
MAGIC_NUMBER retry.py:2: DELAY = 2.741592

Profile: strict  (staged mode)
Errors:   2  (blocking)
Warnings: 0  (informational — promoted to errors when profile=strict)
FAILED: 2 blocking violation(s).

BLOCKED: Zetetic violations in staged files.

The commit re-runs once each flagged line carries a # source: comment, a benchmark reference, or a measured-on note.

What you actually get

A collection of reasoning procedures with citations — not personas pretending to be smart.

97 Reasoning agents, each citing its primary paper
86 Multi-step skills
28 Lifecycle hooks
3 Hosts · Claude Code, Codex, Gemini
CapabilityWhat it gives you
97 documented refusalsEach genius agent's body documents conditions under which it refuses (when to stop, what to cite, when to hand off).
86 multi-step skills15 problem-shaped skills route you to the right procedure; 71 category skills run full pipelines. Type one slash command, get a sourced research brief, debugging trace, or ADR.
Commit-time gatesUNSOURCED keywords (always, never, obviously) block at any profile. MAGIC_NUMBER floats and TODO_NO_REF warn by default, block under ZETETIC_PROFILE=strict.
Problem-shape routingNatural-language problem descriptions map to a reasoning method via shape tags documented per agent — not by field or job title.

Routed by problem shape, not field

Most AI agent libraries ship “pretend to be Einstein.” This ships Einstein's method — gedankenexperiment, operational definitions, equivalence-principle reasoning — with citations, canonical moves, documented blind spots, and the conditions under which the agent must refuse.

Measurement

Measurement & Signal

“The measurement exceeds what known parts predict.”

  • Curie — residual carrier
  • Ekman — anatomical anchoring
  • Wu — testing the obvious
  • Mandelbrot — fat-tail detection
Causal

Causal & Abductive

“Does X cause Y, or is it confounded?”

  • Pearl — do-calculus, ladder of causation
  • Peirce — abductive inference cycle
  • Snow / Hill — epidemiology, criteria of causation
  • Mill — methods of agreement / difference
Correctness

Formal & Correctness

“Can we prove this correct?”

  • Dijkstra — precondition / postcondition
  • Lamport — happens-before, no global now
  • Pāṇini — generative specifications
  • Gödel — incompleteness limits
  • Turing — computability before optimization
Resilience

Failure & Resilience

“What happens when everything goes wrong?”

  • Hamilton — priority-displaced scheduling
  • Taleb — fragile / robust / antifragile
  • Carnot — theoretical efficiency limits
  • Boyd — OODA, fast transients
Decision

Decision & Bias

“Is this decision driven by bias?”

  • Kahneman — System 1/2 debiasing
  • Schön — reflection-in-action
  • Roger Fisher — principled negotiation, BATNA
  • Simon — bounded rationality, satisficing
Ethics

Ethics & Justice

“Who benefits and who bears the cost?”

  • Rawls — veil of ignorance, difference principle
  • Arendt — thoughtlessness audit
  • Le Guin — what will be lost, narrative frame
  • Ostrom — commons governance, eight design principles

Full routing table — problem shapes, pairings, composition chains — in agents/genius/INDEX.md.

What you type → what happens

Each is a multi-step skill that names the procedure used, surfaces blind spots in its output, and refuses to ship if a step fails.

Audit

/paper-vs-code-audit

Extracts every claim from a paper, finds corresponding code, flags mismatches, produces a traceability matrix.

Iterate

/autoresearch-loop

Hypothesis, implement, commit, benchmark, keep or revert, iterate until diminishing returns.

Research

/deep-research

Plans search, runs parallel researchers, synthesizes, verifies citations, writes a cited brief with a provenance sidecar.

Forensic

/incident-investigation

Forensic timeline, three-timescale decomposition, common vs special cause, structural root cause, remediation.

Routing

/genius route

Routes a problem description (“p99 latency exceeds the sum of profiled components”) to the reasoning procedure that fits its shape.

Compose

/genius compose

Chains multiple reasoning procedures in sequence — a Curie residual, a Pearl confounding check, a Dijkstra correctness proof, one pipeline.

Pairs with Cortex memory

Each zetetic agent writes to its own topic in Cortex. Decisions auto-propagate. The next session inherits not just what was decided but who decided it.

Specialization

Engineer's debugging notes don't clutter tester's recall. Each agent has its own memory namespace.

Coordination

Decisions auto-propagate (Wegner 1987). When engineer decides “use Redis over Memcached,” every agent sees it next session.

Briefing

SubagentStart hook extracts task keywords, queries prior work, fetches team decisions, and injects them as a context prefix.

Sourcing chain

When an agent cites a paper, that citation gets indexed in Cortex. Future sessions retrieve the citation along with the claim.

Install

One command. The plugin installer copies agents, skills, hooks, and tools into ~/.claude/.

Option A — Claude Code Marketplace (recommended)

> /plugin marketplace add cdeust/zetetic-team-subagents
> /plugin install zetetic-team-subagents

zetetic-team-subagents ships as its own Claude Code marketplace (zetetic-marketplace); the gates alone install as zetetic-gates. Codex and Gemini CLI get the portable package (see the README). It pairs with Cortex (persistent memory), ai-architect-mcp-codebase (codebase intelligence), and ai-architect-mcp-spec (verified PRDs). After install, restart your Claude Code session. Agents become available via the Agent tool with subagent_type; skills become available via /<skill-name> slash commands.

Option B — Manual clone

$ git clone https://github.com/cdeust/zetetic-team-subagents.git
$ cd zetetic-team-subagents
$ bash scripts/setup.sh install

For advanced configuration — selective agent install, custom hooks, ZETETIC_PROFILE tuning — see docs/INSTALL.md.

Configure the strict profile (optional)

$ export ZETETIC_PROFILE=strict

Strict mode promotes MAGIC_NUMBER and TODO_NO_REF from warnings to blocking errors. Default mode keeps them informational. UNSOURCED always blocks.

Free & open source

MIT licensed. 97 agents. 86 skills. 28 hooks. Every commit is checked.

Part of the engagement journey · Exhibit 02

Where Zetetic Agents fit in an engagement

The 97 agents are wired in during Claude Code activation and keep enforcing sourced claims and coding standards through the paid pilot — the same commit-time gate shown above, running against your own repositories.

Stage Exhibit 02 · Claude Code activation