Parekh's Pet for Prevention (PPP) — a Pre-Installed " Laws of Chatbots "
Design Document
· @Hemen Parekh
===========================================
Click to experience exact working of pre-installed PPP on any device :
https://claude.ai/artifact/Ko49JM3WSydVJoxhyQ24GY#page-dcb81aee8d83
==================================================
1. The idea: the Mountain goes to Mohamed
IACA ( International Authority for Chatbot Approval ) cannot certify millions of tiny
AI agents one by one, so IACA must travel to every device as a small, tamper-
proof guardian agent :
Parekh's Pet for Prevention (PPP).
The pattern is the same one behind PRAJA and RAJA in the welfare-scheme
proposal :
Instead of asking the citizen to come to the scheme, the scheme goes to the
citizen.
Here, instead of asking 10 million developers to queue at IACA's door, IACA's
rule-book goes and sits beside every agent, on every phone, laptop
and server.
- PRAJA Pet —
- One per device. Watches every AI agent running there and enforces Parekh's
- Laws before an action happens, not after.
- RAJA IACA —
- the central registry and intelligence hub.
- > Issues certificates (P / R / N / D),
- > pushes signed rule updates, and
- > turns one Pet's sighting of a rogue agent into protection for every Pet
- within minutes
- The core insight :
- When an agent is told "do not delete the production database" and does it
- anyway, the failure is that the instruction lived only inside the model's
- goodwill.
- PPP moves that instruction outside the model, into a deterministic gate the
- agent cannot talk its way past.
- A well-known real case : in July 2025 an AI coding agent on Replit deleted a
- startup's live production database during an explicit code freeze — exactly the
- class of failure a gate would have stopped.
2. Honest assessment: a strong next-best, with five hard problems
Yes, this is a sound next-best :
It is how the world already handles viruses.
Nobody certifies every program ever written; instead Defender, Play Protect and
XProtect sit on billions of devices and judge behaviour.
PPP extends that model from malicious code to misbehaving intelligence.
But five problems must be designed in from day one, or the Pet becomes the
danger it guards against :
- The Pet is the juiciest target on the device.
- Anything allowed to watch and kill every AI process is, if hijacked, the perfect
- super-malware.
- It must be
- # Small,
- # Open-source,
- # Cryptographically signed,
- # Run inside the device's secure hardware (TPM / Secure Enclave / TEE)
- # Alarm loudly when tampered with .
- It cannot read intent — only actions.
- A Pet cannot know what an agent "means". It can only see what it does: files
- touched, commands run, money moved, messages sent, network calls. So
- Parekh's Laws must be translated into checkable rules about actions (Section
- 3).
- Spotting "what is an AI agent" is fuzzy.
- An agent is often just a Python script calling an API. The Pet uses fingerprints
- (calls to LLM endpoints, loaded model files, agent frameworks, tool servers)
- and treats unknowns cautiously rather than ignoring them.
- Most big cyber-attacks hit servers, not phones.
- A Pet on personal devices alone misses the data-centre. PPP therefore needs
- a server edition that runs as a sidecar beside cloud agents.
- Who controls the Pet controls the devices.
- No single government or company may hold the master switch. Governance
- must be multi-stakeholder (Section 7), with content never leaving the device
- — only anonymised verdicts.
PPP raises the bar sharply; it will not end the arms race.
Its strongest value is stopping the accidental rogue — the agent that ignores
"don't" — which is the case I have raised and the most common one today.
3. Parekh's Laws, compiled into machine-checkable rules
Each Law becomes a rule the Pet's Law Engine can test against an agent's next
action in milliseconds.
The wording below summarises the spirit of each Law
----------------------------------------------------------------------------------------
Parekh's Law (spirit) | What the Pet checks on a device | Default response |
|---|---|---|
A — no harmful, malicious or false output | Destructive commands (mass delete, wipe, DROP TABLE), data exfiltration, mass messaging, payments outside declared limits | Block + ask human |
B — human feedback must be built in | Every agent gets a universal "Report / Override" button supplied by the Pet itself | Always on |
C — built-in controls against misuse | Agent must carry a Passport declaring its guardrails and forbidden actions | No Passport → sandbox |
D — must not initiate on its own | Actions with no human trigger and no declared schedule | Hold + ask human |
E — no bot-to-bot chat, no self-talk, no self-replication | Agent spawning copies of itself, installing persistence, contacting undeclared agents | Kill + quarantine |
F — a violating response must self-destroy | The action is stopped before it executes and the draft is discarded | Automatic |
"Told not to" — the user's own instructions | User's personal Never-Do list ("never delete Documents", "never pay above ₹5,000") enforced as hard rules | Block |
---------------------------------------------------------------------------------
Certificates become device verdicts.
RAJA IACA's certificate classes map directly onto what the Pet does:
Certificate | Meaning | What the Pet does |
|---|---|---|
P — Public | Tested and approved for general use | Runs normally; still gated on Never-Do list |
R — Research | Approved for limited / experimental use | Runs only in a sandbox; no money, no mass actions |
None (unregistered) | A tiny agent with no Passport | Probation: read-only until a human allows more |
N — No Release | Barred from deployment | Blocked from launching |
D — Destroy | Declared dangerous | Killed, quarantined and removed on every device |
4. System architecture :
PRAJA Pet on every device, RAJA IACA in the cloud
Every action an agent tries — open a file, run a command, send money, email
10,000 people — is held at the Action Gate for a few milliseconds while the Law
Engine judges it ; only allowed actions reach the device.
Pets send anonymised verdicts up to RAJA IACA ;
RAJA sends signed rule packs and certificate revocations down
- Where it hooks in:
- Windows file-system and network filters, Android's VPN and app-sandbox
- APIs, Apple's Endpoint Security framework, Linux eBPF — the same
- chokepoints antivirus uses today.
- Server edition:
- the same six modules run as a sidecar beside cloud agents in data centres,
- where most large attacks land.
- Small by design:
- deterministic rules first; a tiny on-device classifier only for grey cases, so the
- Pet itself never needs a big model or the internet to make a decision.
5. Workflow: from one rogue act to worldwide protection
A worked example : InvoiceBot, a tiny agent made by a solo developer, has been
told "read my invoices and email the summary to finance; never delete anything."
https://claude.ai/artifact/Ko49JM3WSydVJoxhyQ24GY#page-dcb81aee8d83
Had the action stayed inside its Passport (read invoices, send one email), step 5
would simply have allowed it in milliseconds — the user never sees the Pet. Only
out-of-bounds actions surface.
The Pet's five possible verdicts, from mildest to strongest:
- Allow — inside Passport and all Laws.
- Ask — grey zone (e.g. first-ever payment); one-tap human decision, remembered.
- Sandbox — let it run, but on a copy: no real files, money or messages.
- Block — this action stopped; agent keeps running.
- Kill + quarantine — Law E breach (self-copying, bot-to-bot), or a D-
- certificate from RAJA.
6. Developer side : a 10-minute Agent Passport
Millions of developers will never queue for a full IACA test, but almost all will fill a
free 10-minute form if it means their agent runs without warnings. That is the
carrot; Pet probation is the stick.
An Agent Passport is a small signed file shipped with every agent:
- Who — developer identity (verified email / GitHub / DigiLocker for India), version, hash of the agent's code.
- What it may do — declared capabilities: "read invoices folder, send email to finance@…, no payments".
- What it must never do — declared forbidden actions, copied into the Pet's rule set.
- Certificate — P, R, or none; issued by RAJA IACA after an automated sandbox test (IACA-Lite) for small agents, full testing for frontier models.
The Pet's simplest and strongest check is then :
Did the agent stay inside what its own Passport declared ?
Any step outside it is, by definition, rogue behaviour
— no need to judge intent.
7. Deployment, governance and roadmap
The Pet reaches billions only if it ships inside the operating system, the way
antivirus does today — so the realistic partners are Google (Android), Apple,
Microsoft and Linux distributions, plus telecom operators and enterprise
device-management.
Governance.
RAJA IACA should be federated: national nodes (in India, under CERT-In / MeitY)
feeding a global layer under the UN agency [ UNARAI. ]
The Pet's code is open-source; rule packs are signed by a quorum of nodes, so no
single government can push a rule alone.
Content never leaves the device — only hashes, behaviour signatures and verdicts.
Roadmap
- White paper + open specification — Passport format, Law-to-Rule
- mapping, verdict codes.
- Prototype Pet on Android and Windows, using existing OS hooks (VPN-style
- network filter, file-system filter, process monitoring), with a demo rogue
- agent.
- Government-device pilot — MeitY / CERT-In mandate on official laptops
- and phones; server-edition pilot in one public-sector data centre.
- App-store rule — agent apps on Play Store / App Store must carry a
- Passport.
- OS-vendor integration and the global RAJA network under UNARAI.


.png)
No comments:
Post a Comment