[B·01]
See what the change really touches
Every PR shows the parts of the codebase a change actually affects — not just the files in the diff.
§00 / INDEX
FOR TEAMS MERGING AI-WRITTEN CODE
FOR TEAMS MERGING AI-WRITTEN CODE
Your AI writes code faster than you can vet it.
Reter gives every change an evidence-backed impact review — derived from a live model of your codebase, not an LLM's opinion. Posted on the PR in shadow mode, so it never blocks your merge.
Fig. 01 / impact review · illustrative example
§01/ Before / After, same query, two agents
- Accent color
- structurally resolved fact
- Warn color
- trust-boundary rule
- Danger color
- violation / taint flow
Before and after comparison of the same query: “where do we authenticate webhook signatures?”. A vanilla agent pattern-matches text and returns a single file that looks related. The Reter-assisted agent uses symbolic reasoning to return the three call sites that actually verify webhook signatures, names the trust boundary rule they enforce, and flags one bypass that violates layer policy.
Fig. 02 / same query, two agents
Left finds code that looks related. Right understands what it means.
§02/ Signal · what's already shipped
[01]
We've contributed accepted security fixes to major open-source projects, including NVIDIA's NemoClaw.
Fixes accepted and merged by the maintaining team.
↳ real-world audit
Source: public open-source contributions
[02]
Broad multi-language coverage.
Built on PhD research in symbolic AI.
↳ depth, not breadth
Source: product capabilities
[03]
Decades-proven symbolic reasoning.
A purpose-built analysis engine over your codebase's structural facts — not a wrapper on an LLM.
↳ engineering, not hype
Source: product architecture
§03/ Problem
Generation got cheap. The merge decision did not.
Agents open pull requests at machine speed. But every diff still reaches code nobody on the PR was looking at — a caller three hops away, a test that never runs, an architecture boundary quietly crossed.
Reviewers can't trace that by eye at agent speed, so “LGTM” does the work — and the regression ships. The constraint moved from writing the change to deciding it's safe to merge.
AuthController
UserService
TokenStore
Session
WebhookVerifier
RulesEngine
AuditLog
DB.Policy
TrustBoundary
ExtAPI
THE PR SHOWS ▼
a fraction of the graph
FIG. 03 · WHAT THE DIFF HIDES
§05/ Benefits · what changes on day one06 EFFECTS
[B·02]
Catch the regression the diff doesn't show
The risks a reviewer can't trace by eye at agent speed — surfaced automatically.
[B·03]
Run the tests that matter, skip the ones that don't
Reter points you at the existing tests that exercise the change — and the changed paths nothing covers.
[B·04]
Evidence, not an LLM's opinion
Every line traces back to a fact about your code you can check yourself. When it's wrong, you can see why.
[B·05]
Honest about what it can't prove
When it can't determine the full reach, it says so instead of showing a false green. "Can't prove this is safe" beats a confident miss.
[B·06]
Shadow mode first — it never blocks your merge
Posts on the PR as a neutral check. We track which predictions held and which didn't, so it gets more accurate over time.
§06/ Differentiator · the reframe
Not just RAG. Not just a vector search. A complete semantic model of your entire codebase.
Unlike approaches that work from keyword matching or vector similarity over text, we work with structural facts about how your code actually composes — call relationships, data flows, architectural layers — derived automatically from your repository.
§06.1
thesis
thesis
"Generation got cheap. The merge decision didn't. The constraint on shipping AI-written code isn't writing it — it's deciding, fast and with evidence, that it's safe to merge."
↳ the bet Reter is built on
We track which predictions held and which didn't, so the system gets more accurate over time.
§07/ Built on research, not hype
[01]The Engine
Reter applies decades-proven symbolic reasoning over your codebase's structural facts — a purpose-built analysis engine, not a wrapper on an LLM.
[02]The Founder
Pawel Kaplanski, PhD AI; published research in symbolic reasoning and knowledge representation — the foundations behind Reter's engine.
[03]The Depth
Broad multi-language coverage across security, architecture, code quality, and more. A purpose-built analysis engine, not a prompt.
02
proven on real codebases
Reter found security vulnerabilities in Nvidia's NemoClaw project. 2 already fixed by their team.
§04/ How it works03 STEPS · 06 ACTIONS
From raw repo to an evidence-backed merge decision, with no extra prompting.
§04.1 / onboarding scan
01step 1
Connect your repo
GitHub, GitLab, Bitbucket, or self-hosted. One-click OAuth, or install the GitHub App.
02step 2
Reter scans your codebase
Analysis builds a structural model of your code: classes, calls, layers, security boundaries, data flows.
03step 3
Every PR gets an impact review
An evidence-backed review posted on the PR in shadow mode — and queryable by any MCP client (Cursor, Claude Code, Copilot).
§04.2 / actions
once scanned, your agent can
scans · 02
Full Codebase Scan
Builds the full structural model of your codebase.
↳ violations report
File Change Scan
Delta scan on every push, incremental fact update.
↳ incremental violations report
requests · 04
Agent Context Query
AI agents query MCP for architectural context.
↳ facts, rules, violations for the slice
Modify Graph
Assert new facts at runtime: layers, patterns, quality annotations.
↳ modification receipt
Human-readable Diagram
Ask for a UML view of any section of the codebase.
↳ UML view
Custom Query
Run predefined or custom queries against your codebase.
↳ query-specific output
› One-click setup
GitHub App or local CLI. Pick repos, set scope.
› Automatic analysis
Scans incrementally on every push. Fresh model, always.
◆ privacy Source stays yours
Reter analyzes your codebase and stores only the structural model, never your raw code.
§04A/ Customize · the rule layerONE ENGINE
The built-in rules cover the common cases. For everything else, you write your own.
Reter ships with detectors for security, architecture, and code smells. When your team's conventions aren't in that list — and they never fully are — you encode them yourself: a rule layer for codifying architectural patterns, a query interface for asking questions of your codebase, and a detector library for surfacing specific concerns.
[01]team policy & architectural rules
Codify your team's conventions as rules. Violations surface automatically — and reviewers read the rule, not the code that enforces it.
[02]ad-hoc questions about your codebase
Ask questions of your code — like which functions are never called — and get structured answers, not a text search.
[03]custom audits over code structure
Build custom checks that match on how code is actually structured, and emit tasks your agent can pick up.
The detectors that ship with Reter are written in the same layer you extend. Nothing is hidden behind the engine. If you can read it, you can extend it — and the agent can read it too.
§09/ About / The founder01 FOUNDER
Pawel Kaplanski. Forty years of symbolic AI, distilled into one engine.
Pawel is the inventor behind Reter's core technology. He holds a PhD in Artificial Intelligence from Gdańsk University of Technology, where his research focused on automated reasoning and knowledge representation. His published work includes papers on automated reasoning-based user interfaces (Expert Systems with Applications) and high-level models for smart city systems (Springer).
Before Reter, Pawel built and shipped production systems across AI, data science, and enterprise software, including roles as Lead Data Scientist, CTO, and founding engineer at multiple ventures. At Reter, he leads all product development, from the symbolic reasoning engine to the MCP integration layer.
PhD · Artificial IntelligenceGdańsk Univ. of TechnologySymbolic reasoning · 40+ years
§10/ FAQ07 QUESTIONS
Common questions, answered up front.
Q.01
Where does my source code go?
Reter analyzes your repository and stores only the resulting structural model — classes, calls, layers, security boundaries, dependency edges. Your raw source code stays in your repo. We read during scan, write nothing back, and persist only the derived facts needed to answer your agent's queries.
Q.02
What languages do you support?
Broad multi-language coverage across most major programming languages, including TypeScript, JavaScript, Python, Go, Rust, Java, C#, C++, Ruby, and PHP.
Q.03
Can I write custom rules for my own conventions?
Yes. Reter gives you a rule layer for codifying architectural patterns, a query interface for asking questions of your codebase, and a detector library for surfacing specific concerns. Every built-in detector is written in the same layer you extend, so anything Reter does, you can extend or override. See the Customize section.
Q.04
Does it work with my AI agent?
Reter exposes its model through MCP (Model Context Protocol). Any MCP-aware client — Cursor, Claude Code, Copilot, Continue, and custom agents — can query Reter for architectural context. One-time setup: install the MCP server, point it at your scanned repo, and every query the agent makes gains architectural awareness.
Q.05
Is this on-prem or SaaS?
SaaS by default, with on-prem available for enterprise customers who need code never to leave their network.
Q.06
How long does the first scan take?
Initial scan time scales with repo size and language mix. Small repos finish in minutes; large monorepos take longer. Incremental scans on every push complete in seconds — only changed files re-parse, and the model updates in place.
Q.07
How is this different from RAG, AST search, or static analysis?
RAG retrieves text that looks similar. AST tools see syntax structure. Static analysis runs predefined rules. Reter builds a structural model of your codebase — how code calls, depends, and composes across files — and reasons over it. It can explain why a query returned what it did, trace dependency chains across files, and surface architectural issues that pattern-matching tools cannot see.