01. Architecture

Product definition (single source of truth)

Falsify is an evidence-driven decision gate.

It tries to disprove high-risk claims made by AI or humans, then emits a scoped verdict (PASS, PASS_WITH_DEBT, or BLOCK) from:

an explicit authority path (what system/state is final),
raw artifacts (not summaries alone),
and an explicit policy version.

Multi-model / multi-agent review is an optional attacker, never the trust root.

Models may propose charges. They may not manufacture facts. Hard BLOCK should prefer deterministic policy, missing evidence, or verifiable state conflict.

Current public MVP surface: GitHub workflows that gate claims surrounding a change (PR narrative, deploy plan, decision docs)—not “full automatic verification of every cloud deployment.”

The problem

AI systems report completion with confident prose:

CI green, logs complete, “another AI reviewed it”
yet the target state never changed
or the evidence surface was already biased before the metric gates ran

Code review and lint ask: does the diff look right?
Falsify asks: is this claim defensible against the authority path?

The gate pattern (public core)

One inspectable loop:

Frame — name the claim, owner, authority path, and claim ceiling
L0 Brooks-Lint (Framework) — structural decay / auditability surface before adversarial attack (marketing alias:「框架审计」; protocol name is Brooks-Lint)
Attack — seek the cheapest counterproof (deterministic checks first)
Recompute / re-read — hit the real state, calculation, command, or raw source
Cutline — Must Fix / Known Debt / Delete (includes L0 structural Must Fix)
Receipt — preserve verdict, evidence path, policy/tool versions, freshness limits, and a brooks_lint block proving L0 ran (or was explicitly scope-refused / skipped with hard cap)

PASS is not permanent. Receipts expire when environment, artifacts, policy, freshness, or authority path change. Claim-bearing review / run without L0 proof cannot yield PASS / PASS_WITH_DEBT.

Not the same tool: falsify lint is a markdown tag/blocker static check (L2 gate path). It is not Brooks-Lint. See Brooks-Lint.

What multi-agent review is (and is not)

Historically this repo also documented a two-agent drafting pattern (Agent A writes, Agent B audits into a shared Obsidian vault over git). That pattern remains a useful collaboration adapter for research write-ups.

It is not the product’s root of trust:

Role	Trust?
Deterministic probes + policy	Yes — primary
Raw artifact hashes / authority reads	Yes — primary
Second model / second agent	Optional attacker only
“Two models rarely share the same error”	Not claimed — unmeasured slogans were removed

If both agents agree on a false assumption, the gate still fails unless the authority path is checked.

Optional topology: shared vault collaboration

            ┌──────────────────────────────────┐
            │   GitHub repo (private)           │
            │   = single source of truth        │
            └──────────────────────────────────┘
              ▲          ▲           ▲
       git push│   git push│   Obsidian Git
              │          │           │
       ┌──────┴──┐  ┌────┴────┐  ┌───┴────────┐
       │ Agent A │  │ Agent B │  │ Your laptop │
       │ (draft) │  │ (attack)│  │ human read  │
       └─────────┘  └─────────┘  └─────────────┘

Three writers can share one truth for research. The gate still bottoms out in authority path + artifacts + policy—not in agent consensus.

Why git + plain markdown still matter

Concern	Why it helps the gate
Diffable history	Receipts and claims are reviewable
Author identity	Provenance for who asserted what
Local-first vault	Human can open the same files agents attacked
GitHub Action surface	Where the MVP ships first

Failure modes the gate targets

Logs ≠ state — “deploy succeeded” with unchanged target
Derived freshness — today’s signal timestamp over stale inputs
Mirror drift — docs/runtime disagree
Metric theater — gates run after an already-shaped evidence surface
Opinion stacking — second AI agreement treated as evidence