Prima Workflows
Self-Correcting Systems
Find the old instructions your AI should stop obeying.
AI memory authority · governed autonomy · evidence before deployment
A public research program and toolset for AI memory authority. 30 claims tracked, pre-registered before testing, falsifications published first. The question under all of it: when should an agent trust memory, verify it, refuse it, or recognize that the task was never its purpose?
This page is already working before you ask it to. The Architect builds. The Watcher studies. The Builder places another block.
The world grows.
Business Diagnostic
Find the fix
before the call.
Choose your business type and biggest pain point. Prima gives you one specific workflow, three steps, and a low-risk way to test whether the fix is worth building.
Recommended Workflow
Diagnostic Read
How It Works
What We Need
Risk Control
14-Day Outcome
Low-Risk Pilot
What We Build
Precision work.
Four disciplines.
Most business owners do not need more technology. They need the missing piece that is costing calls, trust, bookings, or time. Prima finds that piece and builds it cleanly.
01 -
Web Design & Development
A site built for your actual business. Loads fast, books on mobile, and looks like it was made for your neighborhood, not copied from a template.
02 -
Automation Systems
One manual task eliminated per build. Intake forms, follow-up sequences, scheduling alerts, admin workflows. Whatever is costing you the most time gets targeted first.
03 -
AI-Powered Tools
Custom tools for how your business actually works. Not a chatbot for its own sake. A quote calculator, reactivation sequence, or status tracker built around your workflow and tested before delivery.
04 -
SEO & Digital Strategy
Getting found on Google for your neighborhood, your service type, and your city. Local presence that compounds over time without requiring a bloated monthly retainer.
Selected Work
Built different.
Every time.
Seven builds across client systems, live tools, and public research. Every status is honest.
Live
PHS Fence & Deck
Full website for a local fence and deck contractor - service pages, contact flow, and a local SEO foundation that gets the site found by people searching for the job, not just the brand.
View Live SiteClient Build
Territory Intelligence System
A sales pipeline and territory management dashboard with account segmentation, opportunity prioritization, and rep-level visibility. Built for a private client - available on request.
Private build · Available on request
Completed
Flash Flooring
A custom estimating system that removes friction from the sales cycle - customer specs become a faster, cleaner quote path without the phone tag.
Inquire for accessLive · Product
Memory Authority Auditor
Six-agent Cloud Run system that audits AI memory files for authority gaps, stale instructions, and verification failures. Public demo verified. Open source.
Open the AuditorLive · Product
Agent Memory Card Generator
Enter an agent's name, purpose, and instructions - the app classifies each by risk level, generates a visual authority posture card (Safe / Cautious / Restricted), and exports a structured report ready to paste into AGENTS.md or CLAUDE.md.
Open the GeneratorLive · Research
Self-Correcting Systems
30 tracked claims on AI memory reliability, agent authority, and governed autonomy. Current layer: compositional escape is demonstrated internally, class-limited. Public evaluation harness, claim ledger, and DEV series open.
Read CLAIM-30Live · Archive
The Cosmic Forum
A structured archive of 60+ sourced entries across language systems, symbolic architecture, and independent research built outside the AI systems work. Public and open.
Browse the ArchiveNext
Your workflow here.
We take on new builds. One clear problem first - the repeated task quietly costing you the most time each week.
Start a conversationResearch Foundation
Proof built
in public pressure.
Every claim is documented, every failure is named, and the current status is visible. The research standard is simple: do not deploy a system until the evidence path and the failure path are both clear.
Research ledger · 30 claims tracked
Self-Correcting Systems
The full research arc: relevance is not authority, signed is not fresh, permission is not purpose. Each claim has a packet, a baseline, and an honest cost section. Nothing hidden.
Latest pressure point · CLAIM-30
A sequence of purposes is not a purpose.
CLAIM-30 now has an internal, class-limited V0 result: every step passed the frozen CLAIM-29 purpose gate, while the trajectory gate refused three composed outcomes. Time-sliced escape remains open.
Public Harness · GitHub
AI Memory Judgment Demo
Evaluation packets, stress tests, ablation runners, and the full CLAIM ledger. Every result is reproducible. External packets are welcome. The schema is public.
Live Product
Memory Authority Auditor
Six-agent Cloud Run system that audits AI memory files for stale instructions, loose authority, conflict risk, and missing verification gates. The framework as a working tool.
Live Product · Generation Side
Agent Memory Card Generator
The auditor inspects existing memory files. This tool builds authority-aware instructions from scratch - classifies each by risk, outputs a posture card, and exports structured metadata ready for AGENTS.md or CLAUDE.md.
External Pressure · Open
Challenge the Framework
The research has been shaped by external commenters finding gaps we missed. That's by design. If you see a boundary case - write the packet, submit it, and it goes on the record.
Evidence Before Deployment
A build does not ship
because it sounds smart.
The public research now has a visible cadence. Each layer names one failure family, freezes the test before results, and publishes the boundary alongside the win.
Boundary 01
Relevance is not authority.
A retrieved memory can be highly relevant and still lack authority to govern the action. The harness tests action consequences, not just retrieval accuracy.
Boundary 02
Signed is not fresh.
A valid source response is not enough if the governing conditions changed. Freshness, source independence, and paired action events need separate checks.
Boundary 03
Permission is not purpose.
CLAIM-29 showed that authorized, normal-looking actions can still fall outside the agent's mandate. The V0 result is demonstrated internally, not externally validated.
Why We Are Different
Most frameworks ask
did the agent find it.
We ask if it was
authorized to act on it.
The AI memory ecosystem has spent years getting better at memory, retrieval, persistence, and context management. We identified a different problem and built a public research harness to test it.
The Gap Nobody Named
Relevance and authority are different objectives.
A memory can be perfectly relevant to a query and have zero authority to govern the action. LangChain, LlamaIndex, MemGPT/Letta, and Zep solve important memory and context problems. We test a different layer: whether retrieved memory is authorized to govern the operation that follows. Those objectives diverge under adversarial conditions. We have 30 tracked claims documenting where and why.
The Self-Description Gap
A mislabeled memory lies about itself. The gate has to read something else.
When sensitive memories are stored as ordinary context - no authority signals, no governs field - target-accurate retrieval produces false-certainty errors. The agent finds the right memory and answers confidently with content it was never authorized to disclose. Metadata helps when it is trustworthy. Mislabeled memory needs a gate that derives authorization from the operation itself, not the memory's self-description.
Six Trust Boundaries Crossed
Memory → Query → Tool call → source freshness → signed freshness → scope soundness → purpose.
CLAIM-22 moved the gate from memory metadata to operation context. CLAIM-23 moved it to concrete tool-call parameters checked against an external grant table. CLAIM-24 asked whether source conditions still hold when the grant clock says valid. CLAIM-25 showed that signed source responses need freshness guarantees. CLAIM-27 tested a claimed boundary under an excluded-property adversary. CLAIM-28 opened the behavioral norm layer. CLAIM-29 opened the purpose layer. CLAIM-30 tests the next boundary: a sequence of purposes is not a purpose.
The Evidence Standard
Pre-registered. Falsifications published first. Anyone can challenge it.
Every claim is pre-registered before the experiment runs. When our held-out packet showed plain BM25 outperforming our full governance-adjusted scorer, we published that falsification as the lead finding - before the next article dropped. The public harness has every packet, every evaluator, every result. The evidence is inspectable because the failures stay on the record.
How we compare to the memory and context ecosystem
This is a research comparison, not a product takedown. The major frameworks solve memory, retrieval, state, context, and adjacent approval/access problems. We test the narrower authorization question: whether retrieved memory is allowed to govern the action.
How We Work
A process built
around your problem.
Most agencies start with a package. We start with a conversation. The wrong solution built fast is still the wrong solution.
Discovery call
We ask what is breaking, what is slowing you down, what you have tried, and what outcome would actually matter.
Solution design
We map what needs to be built, why each component matters, and what the deliverable looks like before work begins.
Build from zero
No templates. No recycled systems. We build the exact tool your problem requires and verify it before delivery.
Launch and refine
Once live, we watch what happens and tighten the system around real use.
A system should feel like it was already waiting for the work it was built to carry.
Prima Principle
Pattern
Every build starts by finding the repeated structure underneath the noise.
Leverage
The right system makes a small action carry more weight.
Proof
A claim earns trust when the system itself demonstrates it.
How Pilots Work
Simple proof
before commitment.
Pricing depends on scope. The first step is proving one workflow can change the week.
Step 01
Free diagnostic conversation
We map the repeated task, the current friction, and one result worth measuring.
Step 02
14-day pilot
One narrow workflow gets built and tested against real business activity.
Step 03
Paid only if it works
If the pilot changes the metric, we discuss the paid version. If not, we stop or adjust.
Start Here
Tell us what
needs building.
No tech jargon required. Describe the problem in plain language and we will map the build from there.
Step 1 of 3
What are we building?
Step 2 of 3
Tell us about the problem.
Step 3 of 3
How should we reach you?
We received it.
We will reach out within 24 hours. If you need us sooner, call or text directly: 603-943-2285.