SHEET 03 / 06RESEARCH · ACTIVE THESIS

Cognitive Integration Intelligence.

The greenfield-AI era is ending. Ninety-plus percent of enterprise value lives inside legacy systems that cannot be rewritten — and the agents being built today execute without observing, reset between calls, and treat failure as terminal. CII closes all three gaps at once: AI that learns the way an experienced engineer does, inside systems that refuse to change.

FIG. 01 — WHERE BROWNFIELD REALITY, HUMAN EXPERTISE AND AGENTIC AI CONVERGE

diff legacy-agents/ cii/

Three gaps, closed at once.

SEC 02 / SHEET 03

GAP 01

Observe, don't read.

Real expertise is behavioural, not documented. The senior engineer checks a log timestamp before a config file — a sequence that appears in no SOP. CII agents shadow workflows and capture the implicit ordering that makes decisions correct.

REPLACES — DOCUMENT INGESTION + RAG OVER PDFS

GAP 02

Accumulate, don't reset.

Current agents start every session with amnesia. A Rust-backed persistent memory layer carries forward what the system learned yesterday — per-engineer, per-workflow, per-failure-mode — so expertise compounds instead of resetting per prompt.

REPLACES — STATELESS PER-CALL CONTEXT WINDOWS

GAP 03

Failure is data.

When an agent fails inside a 30-year-old stack, that failure is the most information-dense signal in the system. Meta-cognitive recovery reformulates strategy from the failure trace itself — what breaks production teaches the agent.

REPLACES — RETRY LOOPS & HUMAN ESCALATION

SUBSTRATE — RUST PERSISTENT MEMORY LAYER The accumulator runs below the model — a small, fast, typed store carrying observation traces, failure signatures, and engineer-specific decision patterns across sessions. This is the piece that makes the stack behave like an engineer with tenure rather than a prompt.

bench --target=edge --stack=cii

CII on constrained hardware.

SEC 03 / SHEET 03

The thesis only counts if it runs where the constraints live. Benchmarked across four hardware classes — automotive-grade edge to a training cluster.

TABLE 01 — HARDWARE MATRIX · GEMMA 4 · INT8 · OLLAMA
Hardware	Class	Role in the CII stack
NVIDIA Jetson Orin	AUTOMOTIVE EDGE	Full CII stack · INT8 · real-time throughput
Google Coral	TPU-ONLY	Distilled memory layer, quantised observer models
Raspberry Pi	CPU-ONLY	Failure-mode recorder + async sync fallback
6× RTX A6000	TRAINING CLUSTER	Training & distributed inference · observer fleet supervision

Edge AI Benchmark Report — Cognitive Integration across constrained hardware; +5pp accuracy across math/code/reasoning; ~20% INT8 latency reduction; 3/5 sites deployed with cognitive layer vs 0/5 standard

+5pp

Accuracy gain across math / code / reasoning with the cognitive layer

~20%

INT8 latency reduction on edge targets

3/5 vs 0/5

Sites deployable with the cognitive layer vs standard stack

SOURCE — INTERNAL BENCHMARK PUBLISHED IN THE CII FIELD-NOTE SERIES ON LINKEDIN. CLICK THE REPORT TO VIEW THE ORIGINAL POST.

cat field-notes/*.md

The thesis, told in three essays.

SEC 04 / SHEET 03

FIELD NOTE 01

The greenfield AI hype is over.

"Ninety percent of enterprise value is trapped inside legacy systems that cannot be rewritten. The next decade of AI is not about building on blank slates — it is about learning to operate inside the ones that already run the world."

READ ON LINKEDIN →

FIELD NOTE 02

Stop building AI that executes.

"An agent that executes without observing is a faster way to be wrong. Real expertise is behavioural — captured in what order a senior engineer checks things, not in any document they've written."

READ ON LINKEDIN →

FIELD NOTE 03

The problem with tribal knowledge.

"Tribal knowledge is not a documentation failure — it is a behavioural one. You cannot write down what you don't know you know. So we stopped trying to extract it and started observing it instead."

READ ON LINKEDIN →

query record --type=peer-reviewed

The record: the thesis isn't new.

SEC 05 / SHEET 03

PBL-McRBFN in 2014 was meta-cognition applied to biomedical signals and compressed-domain vision. The substrate changed — radial basis networks became language models — but the question is the same: how does a system know what it doesn't know?

JOURNAL · ELSEVIER