Versioned System Prompts Without Losing History

Edit a system prompt and you lose history. Edit a kernel and you create a new version. That one difference decides whether the rules your AI follows can be trusted, reused, and audited, or whether they are just text that quietly drifts.

What we set out to change

Most AI "knowledge" lives as a prompt string in a file. Change it and the old version is gone: you have no record of which version produced yesterday's output, rolling back means git archaeology, and a good rule is not reusable in another bot without copy-pasting. A prompt is a contract written in pencil.

The second is discovery. When a company accumulates rules, nobody can answer "which bot already knows how to do this", so teams re-write the same instruction over and over, and the copies drift apart. (This is one level of the Raisy company operating system; the same immutability is why bot training never breaks the audit trail. Browse the blog.)

The idea / how Raisy fixes it

A kernel is a unit of knowledge with an identity: a stable key, a version, a dimension classification, and a scope. When you edit it, Raisy creates a new version and keeps the old one, so every bot call that used the old rule can still be reconstructed against the rule that actually ran. When you teach one bot, every bot that shares that kernel improves with it, and the change is a dated event.

How it works

• Immutable by default. Each kernel version is a row with its own content hash; updating creates a new version row, never an overwrite.
• Stable key across versions. The kernel key never changes, so bots and logs keep referencing the same contract as it improves.
• Classified, not free-text. A kernel carries BD/KD codes from the dimension indexes, so the system can match a request to the right kernel instead of guessing.
• Scoped to global or private. A global kernel is shared and reusable; a private kernel belongs to one entity and is not exposed to others.
• Reuse before create. When a bot needs a shared slot, Raisy reuses an existing active kernel at the right scope before minting a new one: leaner bots, fewer copies.

What the user actually sees

You edit the rule that governs how your sales bot qualifies leads. Instead of a silent overwrite, you get a version: the new kernel, the old kernel, and the date the bot switched. Every past run shows which version was in effect. You can promote the new version, roll back to the old one, or share the improved rule with the recruiting bot that uses the same slot.

The chart above shows why this is not optional. Between March and June 2023, GPT-4, the same model name, with no version bump a user could see, dropped from 84.0% to 51.1% accuracy on prime-vs-composite classification, from 83.6% to 35.2% on happy-number counting, and from 52.0% to 10.0% on directly-executable code generation, while its response rate on sensitive questions fell from 21.0% to 5.0% (Chen, Zaharia & Zou, "How Is ChatGPT's Behavior Changing over Time?"). If the model underneath your bot silently changes behavior within three months, the rules you control are the only stable ground you have. That is why a behavioral contract must be pinned and versioned, not edited in place.

Why this matters beyond convenience

Rules you can version are rules you can audit, and rules you can audit are rules you can trust. The moment AI knowledge stops being throwaway text and becomes versioned contracts, three things change: outputs become reconstructible, good rules stop being re-invented, and a company's accumulated know-how becomes an asset instead of a pile of prompts.

What this post does not answer

This post describes the kernel model and why versioning matters. It does not publish the dimension-index contents, the classifier's selection rules, or any kernel prompt text, those are internal. It does not measure reuse rates or version churn empirically (no public kernel API yet). It does not compare kernel versioning against prompt-engineering workflows from other tools. Those are next steps once the kernel API is public.

Version your first kernel at raisy.team