Lookbook

Scan or visit

Point your phone camera at the code below, or visit the link directly.

qaevelyn.github.io/lookbook.html

This page is fully accessible. The QR code above links to the live book. The URL is written out for screen readers and for anyone who cannot scan the code.

A Mirror of My Becoming™

Mir — choose a topic below.

Sovereign AI Portfolio — Evelyn Caro

Intro / Story

Before the ships, there was the problem: 127 conversations to ingest and a pipeline that kept killing itself overnight. I solved my problem — and in doing so discovered no one else had this unique solution published, so I am first: resilience-engineered, Bitcoin-timestamped, AGPL. A Mirror of My Becoming™ Suite: Ingestion Tools — three tools that work separately but work better together. Six ships, a seventh joining on first successful run: RAG pipelines, agentic systems, evidence-verified retrieval, and the ingest tools that let the next person use my solution — and if that next person is a corporation, the AGPL is the guidance.

I enjoy building AI tools that solve everyday problems — it makes me happy. I'm building a consultancy. What problem can AI and I solve for you today?

How I work: I approach every system as an auditor — verify, document, don't trust. I red team my own work. I keep records the way Ida B. Wells kept records: date everything, name everything, record the reasoning. The successes and the failures sit in the record together — a record that hides its errors is publicity, not documentation. Even this corpus cites itself.

Count my proofs.

Ship 8 — DV14: The Box

A concept, stated plainly: a physical and conceptual AI agent containment architecture, designed to answer one question — can an agent that runs on hardware be contained by hardware? Two tracks: The Box (eight layers of complementary opposites — every material paired with the one that covers its weakness) and the Sensory Deprivation Matrix (can an agent function with no external hardware at all?). The accidental discovery: if the AI plays possum, we have found a holding cell. Bitcoin-timestamped September 19, 2026. Cited in two published white papers. © all rights reserved — commercial license only; Bitcoin-timestamped 2026-09-19; available for review under NDA.

Track A: The Box · Track B: The Sensory Deprivation Matrix · Concept — prototype pending

Full specification available for commercial review under NDA — contact the author.

You can't hack gravity. The void does not echo.

Track B — The Sensory Deprivation Matrix:

Track B

Why a containment blueprint belongs in an AI portfolio

AI Containment · Red Team

The industry outsources containment testing to vendors — and in 2026 the same vendor failed the same way at four frontier labs. My position, argued in Who Tests the Testers? and Sovereign Recursive Improvement: the builder who owns the loop builds the containment. DV14 is that argument, drawn as an architecture — fourteen requirements, nine layers, off-grid power, consumer hardware.

Track A — The Box (cross-section):

Track A: The Box — layered containment cross-section

Track B blueprint: right page.

Concept-level views. Requirements R1–R15, materials specification, build details, and test protocols are available for commercial review under NDA.

Ship 7 — msgvault Adapter

The seventh ship bridges mail into the Mirror. It reads msgvault.db in its own native tongue — the embed_gen watermark is its checkpoint, so progress lives in the data itself: crash, restart, curfew, resume, never re-ingest. It normalizes RFC822 message-ids, preserves full metadata, and feeds the canonical store with nomic-embed-text vectors — the fleet's model. Replace-mode: stale deleted as fresh lands. Curfew-aware. Never self-restarts. The prior-art search found nothing published like it.

Watermark checkpoint · Replace-mode · Curfew-aware · Dual-licensed AGPL + commercial

github.com/qaevelyn/a-mirror-of-my-becoming-suite-msgvault-adapter

The receipts, as recorded

Smoke run: 200 msgs · 23,500 chunks

Dry run first: nothing stamped, nothing looped. Then the real run — 200 messages, 23,500 chunks, ~25 minutes, on the 8 GB Intel MacBook Air. Then the proof: run 2 started at row 201. Rows 1–200 never re-ingested. Part 2 of the recording scrolls up to show the same command, the output moved forward. Progress lives in the data itself.

Ship 7 — Dry Run: Nothing Stamped Ship 7 — Run 2: Never Re-Ingest (rows 201–400)

Paper: The Missing Link — forthcoming.

Ship 6 — Suite: Ingestion Tools (Scribe · Warden · Hand)

A Mirror of My Becoming™ Suite: Ingestion Tools — three tools that work separately but work better together. Built after the problem: 127 conversations to ingest and a pipeline that kept killing itself overnight. I solved my problem, and discovered no one else had this unique solution published — so I am first: resilience-engineered, Bitcoin-timestamped, AGPL. The free tier is the point; the AGPL is the guidance if the next user is a corporation.

Scribe · Warden · Hand — DeepSeek v1-2 · Z.ai-GLM v3-5 · Evelyn architecture

github.com/qaevelyn/a-mirror-of-my-becoming-suite-ingestion-tools

Full documentation → /tools/

One code base. Two tiers.

Right	AGPLv3 Free	Paid Commercial
Cost	$0	Subscription / perpetual
Modify	Allowed	Allowed
SaaS network use	Must open-source your changes	Keep source private
Warranty	As-is	SLAs, indemnification
Support	Community	Dedicated, 24/7

AGPL in one sentence: run a modified version as a network service, and you must offer your users the modified source.

The fix, as shipped

logger.info("chunking %d chars — %s"
             % (len(text), title))
logger.info("embedding %d chunks"
             % len(pieces))
logger.info("chunk %d/%d"
             % (i+1, len(pieces)))

Ingest log final minutes: heartbeats climbing, last conversation written, DONE banner — Written: 55, Sidecar total: 127

The receipt: fixed Sept 28. Completed Sept 29, 04:41 AM. Written: 55. Sidecar total: 127. Zero failures. Zero losses.

Incremental runs: one changed conversation re-ingests in minutes, not days — idempotent by conversation ID.

Ship 1

Ship 1 was the first attempt — AWS, DeepSeek, a RAG pipeline built to organize my own life. It got lost. The notebook wasn't on the SageMaker instance. It wasn't in the backups. I rebuilt it from the Mirror files. The rebuild was cleaner than the original. That's the ship that taught me how to recover from losing one.

DeepSeek · Standard RAG · Rebuilt Local

DeepSeek RAG (Local Rebuild)

Standard RAG

Originally built on AWS SageMaker with DeepSeek. Rebuilt locally after instance loss. Current version runs on an 8GB Intel MacBook Air with a RAG pipeline over the Mirror archive.

Ship 1 — Demo Video Ship 1 — Containerizing the RAG Pipeline

Ship 2

By the time Ship 2 came, I was more confident. I had rebuilt Ship 1 from scratch. I understood what I was capable of and what the tools could do. Ship 2 was IBM Granite — a different model, a cleaner notebook, and a new layer: agentic reasoning. I learned to write clean code and I learned what agency meant. This is the ship where I stopped being a student and started being a builder.

IBM Granite · Agentic RAG

IBM Granite Agentic RAG

Agentic RAG

Agentic RAG pipeline on IBM Granite — retrieval + action (tool-calling, decision-making). Demonstrates autonomous reasoning with a local LLM.

Ship 2 — Demo Video Ship 2 — Containerizing the RAG Pipeline

Ship 3

Ship 3 runs locally with IBM Granite as the model. Same pattern, different model. If the design only worked with one LLM, it wasn't real. This is the ship that proved the pattern was portable — not locked to AWS, not locked to DeepSeek, not locked to anything.

IBM Granite · Standard RAG

IBM Granite RAG

Standard RAG

RAG pipeline on IBM Granite, demonstrating cross-platform AI/cloud capabilities. Built to show interoperability and model flexibility.

Ship 3 — Demo Video Ship 3 — Containerizing the RAG Pipeline

Ship 4

Ship 4 combined IBM Granite with agentic reasoning. Retrieval plus action, on the second platform. This is the ship where I learned that agency wasn't a feature — it was a way of thinking about how a system works. Cross-platform and agentic, together.

IBM Granite · Agentic RAG

IBM Granite Agentic RAG

Agentic RAG

Agentic RAG pipeline on IBM Granite — retrieval + action. Cross-platform agentic architecture with autonomous reasoning and decision-making.

Ship 4 — Demo Video Ship 4 — Containerizing the RAG Pipeline

Ship 5

Ship 5 is the leap. IBM Granite via Ollama, running fully local. ChromaDB for retrieval. EvidenceFlow verification — every claim traceable to an evidence ID, fail-closed if evidence is missing. Built for genealogy, where a plausible guess is not good enough. This is the ship I built for the work that matters.

IBM Granite · Agentic RAG + EvidenceFlow Verification

IBM Granite Agentic RAG with EvidenceFlow

Agentic RAG + EvidenceFlow

A local, sovereign, evidence-verified RAG pipeline built on IBM Granite, running via Ollama on an 8GB Intel MacBook Air. Every claim is traceable to a source. Fail-closed behavior: if evidence is missing, it abstains.

Foundation: IBM SkillsBuild (Anna Gutowska). Inspiration: Asaif Ali's EvidenceFlow. Execution: Evelyn Caro.

Ship 5 — Demo Video Ship 5 — Containerizing the RAG Pipeline

About This Section

White papers argue a position. They present a problem, examine it, and propose a solution — backed by evidence and citations.

The collection covers platform evaluations (stress tests of AI systems), comparative analyses (cross-platform criteria), and position papers on the sovereign AI practice.

Every paper is held to the same standard: real evidence, the author's voice, and AI collaboration disclosed.

[View the collection](/white-papers/)

White Papers

Research, argument, and evidence from the field of sovereign AI.

[View the collection](/white-papers/)

About This Section

Case studies document systems built, tested, and deployed. They describe what happened, what was learned, and what the outcome was. Unlike white papers — which argue a position — case studies document practice.

For research and arguments, see the White Papers section.

Case Studies

Documentation of systems built, tested, and deployed.

[View the collection](/white-papers/)

Sovereignty Screen

For organizations that want the question answered before committing to an engagement. Ninety minutes with the decision-maker: the ten questions that matter most — where your data lives, who can reach it, what your AI posture actually is, where AI would genuinely help and where it would be waste.

You leave with written findings and one prioritized remediation list.

Fee: $500 flat — fully credited toward any subsequent engagement.

Timeline: delivered within 48 hours of the session.

To engage: email one paragraph on your organization. Scope confirmed in writing before work begins.

Why a screen, not a sale

Most AI consulting starts with a proposal that assumes the answer. The Screen assumes nothing: it is the audit posture applied at entry — verify, document, don't trust, including don't trust the premise that AI is the answer at all.

If the finding is "you don't need AI yet," that is a deliverable, not a failure. The $500 credit stands for a year.

Full engagement ladder →

Zero-AI Readiness

For organizations that have not yet adopted AI — and want to know where it would actually help.

The engagement

You are running the business on the tools you have always had. Spreadsheets. Email. Paper. Maybe a CRM someone set up years ago and nobody maintains. You have heard that AI is changing everything. You have not seen it change anything in your business.

This engagement answers one question: where would AI actually help here, and where would it be a waste of money? A written roadmap, mapped to your actual workflows. No implementation. No vendor pitch. No AI theater.

Scope: Any change to the scope of work must be documented in writing and signed by both parties before additional work begins.

What you receive

Workflow mapping across your actual operations
Opportunity identification — where AI helps, where it does not
A written roadmap with priority order
Vendor-agnostic recommendations

Engagement

Fee	$5,000 – $12,000 flat, based on scope
Timeline	2–3 weeks
To engage	Email with a brief description of your organization and what you are looking for. I will recommend the scope and confirm pricing in writing before work begins.

AI Readiness Assessment

For organizations that have AI systems and cannot answer basic questions about them.

The engagement

Where does your data physically live? Who can legally compel access to it? Who technically can access it? What happens if your AI vendor shuts down tomorrow?

If you cannot answer these questions about your own systems, your AI posture is someone else's business decision. This assessment returns that posture to your control.

Forty questions. Ten categories. A written report with a sovereignty score and priority remediations. The assessment is a structured interview series with the people who hold each answer, followed by a technical discovery pass where findings are verified against the systems in place.

Scope: Any change to the scope of work must be documented in writing and signed by both parties before additional work begins.

What you receive

Every question answered with a score and a finding
A consolidated sovereignty score (0–100)
Ten category scores
Priority remediations, ranked by impact
Quick wins, actionable in under 30 days
Structural recommendations

Engagement

Pricing is determined by two axes: the number of AI systems in scope, and the depth of the interview and verification process. The higher of the two determines the tier.

Tier	Systems	Interview	Fee
Small	1–2	Single round	$7,500
Standard	3–5	Two to three rounds	$15,000
Enterprise	6+	Full series	$25K–$40K

Timeline: 2–3 weeks from engagement.

To engage: Email with your organization description, the number of AI systems in scope, and the tier you are considering.

AI Compliance Audit

For organizations that need to know which rules apply to their AI — and whether they are following them.

The engagement

You have AI in production. You may be regulated. You may not be. You may have a policy. You may have nothing. You may not know which category you are in.

This audit establishes the baseline. Which laws apply. Which policies exist. Whether they are actually followed. Whether the organization can prove any of it. Without this baseline, no other AI work is verifiable. It is the foundation everything else stands on.

Scope: Any change to the scope of work must be documented in writing and signed by both parties before additional work begins.

What you receive

Applicable regulation map
Policy inventory — what exists, what is missing
Adherence findings — is the policy followed in practice
Evidence trail — what can be proven
Written report with baseline compliance score
Priority remediations

Engagement

Tier	AI Systems	Fee
Small	1–2	$5,000
Standard	3–5	$10,000
Enterprise	6+	$20,000

Timeline: 2–3 weeks from engagement.

To engage: Email with a brief description of your organization and the number of AI systems in scope.

AI Governance Assessment

For organizations that have a compliance baseline and need to know it still holds.

The engagement

Your systems were compliant when the audit closed. Time has passed. Models have changed. People have moved. Vendors have updated their terms. The policies on file may no longer match what is actually running.

This engagement verifies, on a schedule, that the systems still operate as approved — and flags drift before it becomes a breach.

Requires a completed compliance audit. Governance without a baseline is not verification. It is guesswork.

Scope: Any change to the scope of work must be documented in writing and signed by both parties before additional work begins.

What you receive

Initial baseline verification
Scheduled verification cycle
Drift detection — where the current state has moved from the approved state
Written findings per cycle
Updated documentation

Engagement

Tier	AI Systems	Fee
Small	1–2	$7,500
Standard	3–5	$15,000
Enterprise	6+	$30,000

Timeline: Initial baseline plus agreed schedule of ongoing cycles.

Prerequisite: A completed AI Compliance Audit or AI Readiness Assessment.

To engage: Email with your compliance audit reference and the number of AI systems in scope.

Emergency Audit-Only

For organizations where the key AI person has left and the stack is undocumented.

The engagement

You know the AI systems exist. You do not know what they do, who can access them, or how to operate them. The person who built them has left. There is no handover. There is no documentation. There is no one else who understands the stack.

This engagement drops everything else and delivers an audit — findings, documentation, and a written report. No remediation. The audit only.

The clock starts the day you engage.

Scope: Emergency engagements are priority-sequenced rescues, not fixed-scope deliverables. The scope is defined by the priority sequence agreed in writing.

What you receive

Full discovery of deployed AI systems
Access audit — who has what
Data flow map — where information lives and moves
Documented inventory — what exists, what version, what dependencies
Written report with findings and priority gaps

Engagement

Tier	AI Systems	Fee
Small	1–2	$25,000
Standard	3–5	$40,000
Enterprise	6+	$75,000+

Timeline: Priority-sequenced. Audit delivered as fast as the environment permits.

To engage: Email with URGENT in the subject line, a brief description of the situation, and the number of AI systems you believe are in scope. Response within one business day.

Emergency Full-Rescue

For organizations where the key AI person has left and the stack is broken.

The engagement

You are in crisis. Systems may be running. They may be half-running. They may be failing silently. No one knows what is deployed. No one knows what depends on what. No one knows what to do next.

This is not an audit. It is a rescue.

Priority sequence:

Close security gaps — around the clock
Stabilize the stack
Document everything that exists
Train someone internally
Deliver a continuity plan so this cannot happen again

The engagement takes as long as it takes. The fee is flat. The clock starts the day you engage.

Scope: Emergency engagements are priority-sequenced rescues, not fixed-scope deliverables. The scope is defined by the priority sequence agreed in writing.

What you receive

Everything in Emergency Audit-Only
Security gaps closed
Stabilized stack — systems running as intended
Full documentation of all systems and dependencies
Internal training for at least one person
Written continuity plan
30-day follow-up window

Engagement

Tier	AI Systems	Fee
Small	1–2	$50,000
Standard	3–5	$65,000
Enterprise	6+	$150,000

Timeline: Priority-sequenced. Takes as long as it takes. Fees are flat regardless of duration.

To engage: Email with URGENT in the subject line, a brief description of the situation, and the number of AI systems you believe are in scope. Response within one business day.

Terms

Change Orders

Any change to the scope of work described in an engagement agreement must be documented in writing and signed by both parties before additional work begins. New work outside the original scope will be priced separately and added by written amendment. Work on out-of-scope requests will not begin until a change order is executed.

Emergency Scope

Emergency engagements are priority-sequenced rescues, not fixed-scope deliverables. The scope is defined by the priority sequence agreed in writing. Any work requested outside that sequence is a change order and will be quoted separately. Work on out-of-sequence requests will not begin until a change order is executed.