Pre-commit firewall

Stop AI agents from corrupting your production database.

Apodictic sits between AI agents and your production databases, checking every SQL write and ERP action against your accounting, tax, and spending rules before it commits. Built by the team that processed 900,000+ statutory documents across 983 businesses.

Run Interactive Demo
$npx @apodictic/mcp-verifier wrap -- <your-mcp-server>
< 38ms
Local savepoint & ERP reversal on policy breach
100% Checked
Every write evaluated against encoded invariants
1.39 ms
Median Z3 check across 10,000-case benchmark
0 Bytes
Raw PII sent to external models via gateway

Apodictic today

4
Paying customers

Multi-branch retail and pharmacy

14
Unsafe writes intercepted

Largest deployment, 30 days

12
Branches

In that deployment

Before Apodictic

900K+
Statutory documents

Processed by Rasmi

983
Businesses

Historical Rasmi accounts

Supported stack

Attach at the database wire, ERP RPC layer, or MCP agent boundary without replacing your models.

PostgreSQL

Wire proxy & SAVEPOINT undo

Odoo 18

JSON-RPC & account.move.reversal

Stripe

Pre-capture refund & cap guard

MCP

Drop-in stdio tool verifier

SAP S/4HANA

BAPI & ledger invariant checks

NetSuite

SuiteTalk & 3-way PO match

The core idea

Traditional access controls verify who an AI agent is. They never verify whether its transaction leaves your ledger intact. Apodictic sits at the commit boundary, proving every SQL write and ERP mutation against your financial invariants before a single row changes.

The risk

Authorization isn't financial integrity. Valid credentials do not guarantee valid state.

Traditional access controls determine what an agent is permitted to touch. They do not determine whether its actions will produce a valid financial state. An agent with legitimate credentials can still post an unbalanced journal entry, modify the wrong branch's pricing, or delete a protected financial record.

Production Lineage: Derived from real failure patterns observed while processing 900,000+ statutory documents and 5,000,000+ ledger writes across 983 businesses—stopping destructive ERP mutations before state commits.
01 · Data corruption

It deletes the wrong record.

An autonomous accountant asked to "tidy up November receivables" runs a soft-delete query on historical invoices. It breaks statutory tax hash chains, invalidates previous VAT filings, and leaves the company exposed to statutory audit penalties.

Every action is checked against business rules

Every SQL and ERP write is checked against balance, tax, and retention rules before commit. Unsafe deletes are blocked.

02 · Data leaks

It leaks private data outside your perimeter.

A customer support agent pastes an unredacted PDF purchase order into a cloud LLM context window. Corporate bank accounts, national IDs, and commercial pricing agreements are permanently cached in external model logs.

Sensitive data stays inside your environment

Bank accounts, tax IDs, and credentials are stripped locally at the boundary so only safe placeholders reach external models.

03 · Wasted spend

It re-runs probabilistic inference on deterministic rules.

For the 5,000th time this month, an agent calls a large model to verify if an invoice exceeds $10,000. Temperature variance produces conflicting approvals, and the team pays $0.05 per check for basic arithmetic.

One security boundary for your AI agents

Approved business policies become permanent local rules executing in under 2ms with zero recurring model cost.

Ask once. Watch every step.

Pick a task. An AI agent gets it, and Apodictic checks it, protects the data, and keeps an undo ready.

apodictic, live sandbox
  1. Confirm rules
  2. Hide private data
  3. Test in a sandbox
  4. Approve and undo
>

What a saved rule is worth ($0.05/call in tests)

Asking the model every time$250
With a saved rule$0.00

How it works

Every transaction. Three checks.

Loading walkthrough...

Illustrative walkthrough. Values are simplified to demonstrate the commit boundary.

Real-world proof · 30 days

14 unsafe writes intercepted. Zero reached production.

An anonymized live deployment across a 12-branch retail and pharmacy customer. Real AI workflows attempted database mutations that violated encoded corporate financial policies.

14
Unsafe Writes Blocked

0 reached database

30
Days Observed

Continuous gateway telemetry

12
Branches Protected

PostgreSQL & Odoo 18 nodes

0
Corrupted Records

100% SHA-256 chain integrity

01What Happened

Valid AI Credentials, Dangerous State Changes

Autonomous agent workflows with valid database credentials proposed mutations that would corrupt business state:

  • 3 cross-branch pricing errors: Subsidized branch tiers applied to unauthorized store IDs.
  • 7 unbalanced journal entries: Credit total drifted by $1,420 from debit lines.
  • 4 attempted soft-deletes: UPDATE statements nullifying records inside an active SHA-256 tax chain.
02What Apodictic Did

Deterministic Evaluation Before Commit

The interceptor caught every SQL and ERP action locally in memory prior to PostgreSQL / Odoo commit:

  • Zero-delay interception: Parsed AST and mapped affected relations within 0.4 ms.
  • Z3 invariant evaluation: Solved debit == credit and reserve floor constraints in 1.7 ms.
  • Fail-closed rollback: Triggered savepoint WAL rollback; 0 dirty rows touched storage.
  • Immutable audit log: Signed the failed mutation into a tamper-evident SHA-256 audit chain.
03What Was Verified

Forensic Database State Proof

Database states were forensically inspected immediately following each blocked attempt:

  • 0 uncommitted writes: Database diff confirmed balance and sequence numbers remained intact.
  • Customer confirmation: Customer engineering team verified all 14 were unintended AI actions.
  • Reproducible test coverage: Converted all 14 incidents into synthetic regression test suites.

Inspect the corresponding synthetic test suites, Z3 proofs, and execution traces in our public evidence repository:

Public benchmarks

Every number links to the file behind it.

Inspect the architecture, reproduce verification tests and examine execution evidence.

Why unchecked AI fails

Your team is already using AI. Nobody is checking it.

38%

Of AI-generated code samples carry exploitable vulnerabilities. One unchecked write is an incident.

Veracode GenAI Code Security Report · Fall 2026
up to 17.2x

Error amplification in unchecked multi-agent chains; 4.4x with a central validator.

Kim et al., Google & MIT · 2025
43%

Of breached organizations suffered a shadow-AI incident averaging $5.39M vs. $4.99M overall.

IBM Cost of a Data Breach Report · 2026
The fix isn't a smarter model. It's a gate.

How we compare

From agent security to financial state integrity.

Agent security platforms govern what autonomous agents can access, invoke, and execute. Database guardrails inspect and enforce rules over database operations. Apodictic focuses on a complementary layer: verifying proposed financial state transitions against explicitly encoded accounting, tax, and operational invariants before they commit.

Rather than relying solely on permissions or query-level restrictions, Apodictic uses formal constraint verification to evaluate whether a proposed mutation preserves the business rules governing your production systems.

-85.2%
Token Overhead Reduction

Symbolic tokenization strips boilerplate PII and noise before cloud dispatch

< 10 min
Time to Deploy

Drop-in @apodictic/mcp-verifier wraps any MCP server or database connection

$0.00
Recurring Eval Cost

Pre-compiled Z3 rules execute locally without querying external LLMs

100.0%
Attack Interception Rate

60/60 attacks blocked in public suite · 14/14 real unsafe writes blocked in live production

CapabilityApodicticVerictoAkto AgentGuardGeneric GuardrailsPublic Proof
Z3 Formal Financial Invariant Verification
Proves balanced debits/credits, liquidity floors, statutory VAT parity & role caps with mathematical certainty.
Z3 5.4.1 SMT ProofSQL AST / PolicyGuardrail RulesRegex / Prompt Filter100-Case Z3 JSON
ERP-Native & Local WAL Reversal
Sub-25ms PostgreSQL savepoint WAL undo and sub-38ms Odoo 18 account.move.reversal upon invariant breach.
<25ms WAL / <38ms OdooBlock Only (No Undo)Block OnlyNo RollbackRecovery Proof
Cross-Session Split Smurfing Detection
Stateful cumulative tracking blocks multi-turn split payouts across worker restarts (e.g. 3x $49,950 under a $50k cap).
Stateful CumulativePer-Query StatelessSession-ScopedStatelessGist 08 (ATK-06)
Zero-PII Symbolic Gateway (SYM_*)
Bijective tokenization in local RAM replaces names, IBANs & keys before external LLM prompt egress (-85.2% tokens).
0 Bytes Raw PII SentResponse MaskingPII RedactionNone / Cloud-SideGateway Spec
Deterministic Saved-Rule Caching ($0.00 Cost)
Compiled into local Z3 solver; repeated checks run locally at $0.00 external LLM token cost (saving 80–90% on API fees).
$0.00 / Verified CheckLocal ProxyAPI / Agent EvalPer-Call LLM FeeBenchmark JSON
Time-to-Deploy Enterprise Guardrails
Drop-in stdio MCP proxy & npm SDK vs weeks of custom LLM prompt eval engineering or proxy re-architecting.
< 10 Minutes (SDK/MCP)TCP Proxy SetupSDK + Policy ConfigCustom Developmentnpm SDK
256-Bit Cryptographic WORM Audit Chain
Full 256-bit SHA-256 digests + secp256r1 ECDSA signatures linking every blocked and allowed write.
SHA-256 + ECDSASigned Audit LogEvent LogsBasic LogsGist 01 (ATK-01)
60 / 60 Attacks Blocked · 40 / 40 Valid Allowed

Pre-Commit Data Integrity

Traditional tools check if an actor has database permissions. Apodictic proves whether an authorized AI write will corrupt your ledger, breach liquidity reserves, or drift statutory VAT.

Inspect 100-Case Benchmark JSON →
1.42 ms P50 · 2.516 ms P95 · 2.976 ms P99

Precision Latency & Local Reversal

Z3 gate checks run in ~1.4 ms at P50 (2.976 ms P99 steady-state; 15.192 ms cold-start). Local Postgres WAL savepoints undo in <25 ms and Odoo account.move.reversal completes in <38 ms.

Open Interactive Proofs & Runner →
4 Paying Accounts · 14 Interceptions

Production Lineage & Live Traction

Built on operational experience from 983 historical Rasmi compliance businesses. In live multi-branch retail & pharmacy deployments, Apodictic intercepted 3 cross-branch pricing errors, 7 unbalanced journal entries, and 4 hash-chain soft-deletes.

Inspect Security & Trust Posture →

Why we built Apodictic

Before Apodictic, our team built Rasmi across 983 historical businesses and over 900,000 statutory invoices. When multi-branch finance teams began connecting autonomous scripts to Odoo and PostgreSQL, the failures didn't come from outside attackers—agents with valid credentials posted unbalanced journal entries, leaked cross-branch discounts, and broke sequential tax chains.

We built Apodictic as the deterministic commit boundary between autonomous AI and production ledgers—so teams can move faster without risking silent financial corruption.

Keep autonomous operations accountable.

Your AI agents can move faster. Your financial controls should remain deterministic.

Explore Technical Proofs

Or write to the founder directly: sofiantarek2009@gmail.com

Under the Hood·Technical Overview

The Execution Boundary Between AI and Production Systems

Language models are probabilistic by nature: they predict the next plausible token. In enterprise software, guessing leads to database corruption, accidental discounts, and data leaks.

Closed-Loop Verification Architecture

01.

De-Identified Data Boundary: Before any prompt reaches an outside model, customer PII, tax IDs, and sensitive schemas are stripped and replaced with bijective placeholders. Outside models never receive raw corporate data.

02.

Isolated Sandbox Execution: Candidate actions run inside an ephemeral microVM sandbox with strict default-deny network egress. Security flaws and policy violations are caught and auto-repaired before any state change is approved.

03.

Atomic Commit & Instant Rollback: When an action passes verification, it commits across databases, ERPs, and APIs with a cryptographic hash chain. If an error is detected later, the entire operation can be reverted in under 50 milliseconds.

Apodictic Enterprise Architecture