Engineering Intelligence Framework

Engineering
Intelligence
Framework

A governed methodology that lets engineering experience compound.

EIF turns validated engineering experience into reusable operational intelligence: one shared operating layer, project memory, session discipline and invokable skills that outlive any single session.

01 · The problem

Chat history is not engineering memory

AI coding agents are powerful, but most workflows still treat a session as if it were durable engineering knowledge.

  1. Previous decisions live only inside one session and are lost when its context is compacted or the session ends.
  2. The agent re-discovers the same files and re-derives the same conclusions every session.
  3. Local lessons never resurface before a new, similar task.
  4. Command output floods the context window without filtering.
  5. Architecture decisions and scope get mixed up with unverified assumptions.
  6. A merged PR gets treated as proof of working behavior, even when behavioral evidence is missing.
  7. Quality, privacy and token-efficiency are governed by separate, sometimes contradictory rules.
02 · The methodology

Engineering Intelligence makes experience reusable

At closeout, a session records a Knowledge Delta. This EIF artifact captures the change in knowledge. It is checked, kept in the project, or explicitly promoted into a reusable rule, playbook, template, or skill.

In EIF, Engineering Intelligence means versioned engineering experience plus the methods that govern its use. It shapes how the next task is prepared, executed, and verified. It is more than a place to store knowledge.

“Learning” happens at the system level. Validated experience becomes a reviewed, versioned artifact that the next session can retrieve. EIF does not train model weights, fine-tune an LLM, or let an agent rewrite its own safeguards.

Methodology
Principles, ontology, authority, playbooks and evidence rules define how work is performed.
Experience
Validated facts, decisions, incidents, failure patterns and rejected approaches become retrievable before implementation.
Capability
Templates make good structure repeatable; skills make complete workflows invokable by an agent.
03 · Intelligence layers

Three layers: how to work, what a project knows, what is happening now

EIF separates reusable framework methodology, durable project knowledge and ephemeral session context into three layers, with Knowledge Delta governing promotion between them.

L1 Reusable

Framework layer

The reusable operating layer for agent work: ontology, the authority model, global rules, playbooks, templates and invokable skills. More than method, and shared across every project.

How to work

L2 Durable

Project layer

Architecture, decisions, domain facts, incidents, risks and local conventions that belong to one project.

An optional private workspace can coordinate several projects inside L2. It selects and pins shared rules, skills, playbooks and templates into each project without becoming another layer or owning project content.

What this system knows

L3 Ephemeral

Session layer

Current scope, progress, observations, assumptions, open questions and a draft Knowledge Delta.

What is happening now

Flows down: the whole operating layer together with the project's own prior experience. Ontology, the authority model, rules, playbooks, templates and skills all reach the session before implementation starts.

Flows up: only validated knowledge, and only by a separate decision recorded in Knowledge Delta. It usually stays on the layer where it was found. Promotion is the exception, not the required closing step of every session.

Traced: two sessions of one packet

Experience arrives before implementation, not written up after it

One packet, two sessions, three questions: what each session is given before it writes anything, what survives its closeout, and what session 2 finds already waiting for it.

  1. Before implementation

    Retrieval, not recall. A session opens by searching durable knowledge for prior lessons, rejected approaches and repeat failures. Work under EIF begins with this step, even when a task looks simple.

  2. Between sessions

    Packet memory. A task spanning several sessions gets a packet: charter, facts, decisions and roadmap written before implementation. That is what carries across the gap, so session 2 does not reopen what session 1 settled.

  3. At closeout

    Retro, then Knowledge Delta. Closeout compares intent against evidence and names the failure patterns. Raw reasoning is dropped on purpose; only structured, sourced results become project memory.

  4. Rarely, upward

    Promotion is the exception. A lesson leaves its project only if it applies beyond it, passes a lint on source and type, and an owner accepts it. Most validated knowledge is meant to stay local.

One framework, many projects

The framework is shared. Every project keeps its own memory.

One shared operating layer at the centre, and every project on its own orbit around it. The centre holds what is reusable everywhere: ontology and the authority model, rules and constraints, playbooks, templates and skills. Each project keeps its own repository and its own memory, and retrieves from the centre before it starts. Under P1 a packet of three sessions is open, because the task was too large for one of them. Method and the project's own history descend into those sessions before their first line of implementation; when the packet closes, a validated lesson can climb back to the centre, and from there reach another project only where it applies.

  • The centre is the framework layer. Everything reusable lives there, and every project orbits it with its own repository, its own memory and its own packet rules.
  • Method and experience go down first. Both arrive in the session before it implements anything, retrieved rather than recalled.
  • One packet, three sessions. They run in order under P1, and the packet closes before anything it learned is offered upward.
  • The climb back is the exception. A lesson reaches the centre only by an explicit decision, and the dashed relay carries it to another project only where it applies.
  • The curator never stops. The amber dot circling the centre is the one continuous motion here: it watches the framework layer's own health on its own rhythm, and merges nothing itself.
04 · The control plane

One governed loop, from intent to durable knowledge

EIF is a control plane that coordinates tools and agents without replacing them. It connects knowledge, authority, execution and evidence into a single, observable sequence.

  1. 01 · User intent

    A task or question enters the loop.

  2. 02 · Experience retrieval

    Prior validated knowledge is pulled into context before new work starts.

  3. 03 · Authority and scope

    What should happen, what actually happens, and what may be cited are kept as separate questions.

  4. 04 · Structural navigation

    An optional code graph may help find things faster. It is navigation evidence, not authority.

  5. 05 · Controlled execution

    Work proceeds under explicit scope, budgets and stop conditions.

  6. 06 · Tests and evidence

    A merged change is not treated as proof of working behavior on its own.

  7. 07 · Knowledge Delta

    What was learned is written down in the same session, not left in chat.

  8. 08 · Durable engineering intelligence

    Validated knowledge is available to the next task and the next agent.

Source files remain authority throughout. An optional structural graph, when present, is navigation evidence: it does not replace source verification.

05 · The session

A session is the unit of controlled work and learning

EIF requires retrieval of relevant prior knowledge before non-trivial work and defines preparation, controlled execution, verification and closeout as the session lifecycle.

  1. 01

    Retrieve and orient

    Read active instructions and source, then search durable knowledge for relevant lessons, failures and rejected approaches.

    L1 + L2
  2. 02

    Scope and prepare

    Choose a single session or an execution packet. Name the goal, evidence, constraints, budgets and stop conditions.

    Task scope
  3. 03

    Execute and verify

    Keep observations separate from assumptions, work within scope and use declared evaluators instead of completion language.

    L3
  4. 04

    Close out and promote

    Record actual evidence, open work and Knowledge Delta. Retain validated learning; remove or discard ephemeral session state.

    Durable delta

The session layer is a checkpoint, not memory: it may contain progress, evidence labels, open questions and a draft delta, but no raw AI reasoning. It is never promoted wholesale.

06 · The evidence loop

Bounded iteration, not open-ended autonomy

EIF includes a bounded evidence-loop contract for iterative agent work, with explicit evaluators, budgets, evidence and stop conditions. This is a methodology contract. It does not by itself prove fewer retries, higher quality, lower cost, or reliable autonomous completion.

  1. 01 · Orient

    Read current source and relevant retrieved knowledge. Record facts separately from assumptions.

  2. 02 · Define

    State the next smallest hypothesis and the observation that would confirm or reject it.

  3. 03 · Act

    Make one reversible change or run one diagnostic experiment.

  4. 04 · Observe

    Run the declared evaluator and record its actual output.

  5. 05 · Study

    Compare expected and actual results. If unsuccessful, record a concise failure signature comparable across attempts.

  6. 06 · Adapt

    Select the next action because of the observation. An unchanged failing action with no new evidence is not repeated.

Preconditions

Declared before the first iteration. Without them an agent is not iterating, it is retrying.

Goal
One observable outcome.
Success evidence
What proves the outcome behaviorally.
Evaluator
Test, tool, source comparison, rubric or named human decision.
Iteration limit
Positive integer, default 5.
Remote budget
Default 0 unless explicitly authorized.
Stop conditions
Safety, ambiguity, budget or owner-only boundaries.

Exits

Three ways out and no fourth. Each one names the evidence or the boundary that ended the loop.

Pass
Only when success evidence exists.
Blocked
A stop condition is reached, the evaluator is unavailable with no adequate substitute, the iteration budget is exhausted, or the same unchanged failure signature would be attempted a third time.
Deferred
Only when the remaining work is explicitly owner-gated or outside ratified scope, with a named destination.
07 · Governed learning

From experience to reusable knowledge and skills

Learning becomes durable only when a session's validated Knowledge Delta is classified, linted, retained locally or explicitly promoted into a versioned rule, playbook, template or skill.

Two loops, not one. The inner loop runs inside a session and asks what this session learned. The outer loop runs across many sessions and asks a question no single closeout can answer: what keeps happening? A lesson seen once is an incident. The same failure signature seen repeatedly is a missing rule.

  1. Observe

    Real work produces evidence: a fact, incident, rejected approach, risk or repeatable method.

  2. Capture

    The session records a concise Knowledge Delta, including what remains unverified.

  3. Retro

    The outer loop, across a packet or several weeks: derive the period from the repository, then find what repeats. Running it over one session finds nothing, by design.

  4. Classify + validate

    Knowledge Ingest and Lint check artifact type, source, status, language, duplication and links.

  5. Promote

    Project-specific knowledge stays in the project layer; a general lesson reaches the framework layer only by explicit decision.

  6. Retrieve

    The next relevant session finds the artifact before acting, closing the learning loop.

What that knowledge actually lives in

A linked, indexed, wiki-like base - not a folder of notes

Every artifact is a plain file with structured front matter: type, scope, status, confidence, source, and related: links to its neighbours. Those links are what make it a graph instead of a pile - a decision points at the incident that caused it, a lesson at the rule it became.

  • Each circle is one artifact; each line between two of them is a related: link.
  • A search enters from the index lane below and walks those links, instead of reading every file.
  • The dashed node is rejected or superseded: still readable, kept out of results by default, and flagged by the curator.
  • Indexed. An index is generated over every artifact under the configured knowledge root, and retrieval searches that index offline: keyword and substring scoring, no embedding service, nothing leaving the machine.
  • Lifecycle-aware. Each node carries a status. Rejected and superseded artifacts stay readable but are excluded from retrieval by default, so a closed question does not quietly reopen itself.
  • Honest about gaps. Unparseable, schema-invalid and not-found are reported as three distinct outcomes. An empty result never gets to look like a clean one.
  • Maintained. The knowledge curator reads the same base and ranks what has rotted in it - stale artifacts, orphans, broken links, findings that were closed and came back - into one list with an owner for each. It never merges anything.
Experience

What the system has learned

Facts, decisions, incidents, edge cases, failure patterns and lessons with provenance and lifecycle status.

Skill

How the system can act

A reusable, invokable workflow for planning, execution, review, retrieval or curation. It is versioned instructions, not hidden model behavior.

This is governed system-level learning, not autonomous self-modification. A lesson can be rejected, remain local, expire or be superseded; a local skill may specialize workflow but cannot lower framework safety rules.

08 · Optional capabilities

Three tools help. None of them is required.

Graphify, RTK and Context7 each add one narrow capability. None is the methodology itself, and the framework installs none of the three.

What happens when one is missing

Every tool has a fallback path when it is unavailable

EIF takes the fallback path and reports the tool as unavailable instead of quietly substituting a guess: source search instead of the graph, unfiltered output instead of compression, local documentation instead of a live lookup.

  • Solid: the route with the tool available.
  • Dashed: the fallback path used when the tool is unavailable. Both routes arrive.
  • Amber: the two boundaries. RTK drops output before it costs context; only Context7 leaves this computer.
  • The box on the right is the context the agent has to spend.
  • Boundary: local for AST-based extraction. It reaches an external API only if semantic labeling is turned on deliberately.

    Without it: the agent falls back to source search and manual browsing, with no graph-level path or impact hints.

    Verified: D-08 metadata binds the graph digest to its source commit, the Graphify version, an explicit repository identity and the reviewed scope. Only fresh reports healthy; a code, semantic or history change asks for the matching update, and an invalid or deliberately suppressed artifact stays unavailable. Bounded to one local version and one small synthetic fixture, not a navigation outcome.

  • Boundary: local, with nothing sent anywhere.

    Without it: shell commands run unfiltered and no reduction is recorded.

    Verified: RTK 0.43.0 passes all 24 checks on the tested Windows host through the native rtk rg route. A failed required canary still reports degraded, and a raw or failed proxy route records zero savings rather than a positive number. Not a cross-platform or efficiency claim.

  • Boundary: the query text goes to the network to reach a hosted documentation service. The framework sends no source code.

    Without it: the agent relies on local docs or training-data recall, which may describe a different version than the one installed. An unavailable provider is reported, never silently replaced by recall.

    EIF does not install it. An MCP server is configured at the user level, outside the project instance, and the framework never changes user-level configuration. It ships generated routing guidance and a reviewable per-adapter template whose provider command stays owner-supplied.

    Verified: the declaration, the routing guidance and the template shape are covered by tests; live provider behavior is not. EIF generates configuration for the agent to load and does not speak MCP itself, so reachability is reported by the agent and stays marked unverified.

09 · Capabilities and evidence

What runs today, and what each claim rests on

Eight things this framework does now, each with the check behind it and the scope that check covers. The scope line belongs to the row it qualifies, not to a disclaimer at the foot of the page.

  • The wheel attached to GitHub Releases, or pip install . from a clone, gives you a working eifctl. Its eleven commands cover private-workspace and project creation, connection, guarded two-axis updates, verification, retrieval, rendering and validation. The installed wheel is exercised through the workspace lifecycle and a synthetic task, not by an import check.

    Scope: the artifact is available on GitHub Releases, but it is not published to PyPI. Each creation command makes a local repository, not a GitHub remote. Multi-project updates are preflighted but not atomic across repositories.

  • The adoption preflight reads the repository before it writes anything. If the project is already governed it stops and asks for an explicit decision instead of taking over, and in coexist mode the generated block defers to the rules already there rather than declaring itself the only authority.

    Scope: existing governance is recognised by the size and shape of the files, not by reading what they mean. Verified against a synthetic fixture and repeated runs against one real private repository.

  • Search runs offline over the project's own index, aware of each artifact's schema and its place in the lifecycle. Rejected and superseded artifacts stay readable but are kept out of results by default, and unparseable, schema-invalid and not-found come back as three different answers rather than one empty one. Ukrainian artifacts are indexed and returned like any other.

    Scope: keyword and substring scoring, not search by meaning. No embedding service, and nothing leaves the machine.

  • A unit of work opens as a session or as an execution packet with the evaluator named up front, budgets for iterations and remote runs, and a stop condition that does not depend on the agent deciding it is finished (see the evidence loop).

    Scope: a contract the workflow holds itself to. It sets how work is bounded, not how many attempts a given task will take.

  • The bundled verifier recomputes the manifest digests from the files on disk, validates config and lock against their schemas, and checks that the active adapter matches the configuration, that the generated entrypoint has not drifted from the config that produced it, and that the block markers are intact.

    Scope: it verifies the EIF bundle it manages. It is not a general file-integrity tool for the rest of the repository.

  • Claude Code, Cursor, Codex and Hermes. All twelve ordered pairs are exercised: after every switch there is exactly one active EIF block, the new adapter's entrypoint resolves to the right source, and the project's own content is left as it was.

    Scope: adapter coverage is these four, deliberately, and the core is plain Markdown any other agent can be pointed at. What differs between them is mechanics, not status: Cursor has no tool-call hook to enforce through, so its adapter ships rules, and the Codex and Hermes hook contracts block rather than rewrite.

  • Retrieval, Knowledge Delta rendering, session closeout and the generated entrypoint content all render in the language the instance declares, verified in Ukrainian as well as English. The language is configuration, not a hardcoded default.

    Scope: the surfaces the framework itself renders. What an agent says in free-form conversation is the agent's own behavior, not something EIF produces.

  • The benchmark harness runs a real-agent A/B and records the outcome as a machine-checkable contract rather than a headline: one model, three fixtures, six attempts, every attempt completing in both modes. Modes C and D run as integrity contracts, so what they establish is that a run was well formed, not that it was faster.

    Command-level timing exists for specific mechanical operations and is recorded with the framework's benchmark evidence. Nothing measures how long a person spends writing or reviewing code.

    Scope: this page makes no efficiency or task-quality claim anywhere, and it will not until a study exists that carries one. Six attempts against no baseline and no control group is not that study.

10 · Quickstart

From a clone to a closed-out change

Cloning the repository gets you the ontology, the schemas, the governance docs, the eifctl package and one reproducible vertical slice.

  1. 01 · Install

    Put eifctl on your PATH, from a clone

    You need Python 3.11 or newer and Git. Every line below is typed into a terminal, from whichever directory you keep projects in. Pick the one that matches your shell.

    macOS, Linux, WSL

    git clone https://github.com/mike-arbuzov365/engineering-intelligence-framework
    cd engineering-intelligence-framework
    python3 -m venv .venv && source .venv/bin/activate
    pip install .

    Windows PowerShell

    git clone https://github.com/mike-arbuzov365/engineering-intelligence-framework
    cd engineering-intelligence-framework
    python -m venv .venv; .\.venv\Scripts\Activate.ps1
    pip install .

    The separator is ; and not &&: Windows PowerShell 5.1 parses && as an error.

    Windows CMD

    git clone https://github.com/mike-arbuzov365/engineering-intelligence-framework
    cd engineering-intelligence-framework
    python -m venv .venv && .venv\Scripts\activate.bat
    pip install .

    .venv is a throwaway Python environment that belongs to the clone, so nothing here touches the Python you already use. And the lone . on the last line is not a typo: it names the directory you just cloned into, which is where pip reads the package from.

  2. 02 · Initialize

    Turn a repository into a project instance

    One line, the same on every platform, with the environment from step 01 still active.

    eifctl init --project-name <name> --instance-path <path> --locale en
    --project-name
    What the instance calls itself. Any short name.
    --instance-path
    The repository you want EIF to govern. Usually your own project rather than this one, and . works if you are already standing in it.
    --locale
    Optional. The language the instance reports in. English if you leave it out, and uk is the only other language that ships.

In the reference walkthrough, the bundled demo seeds and retrieves real knowledge, makes one retrieval-informed change, and verifies it with a real test that fails before the change and passes after it. It then closes the session in whichever language the instance config declares. That run declares Ukrainian, because a closeout rendered in English would have proved nothing about the locale layer: the point is that the language is configuration, not a hardcoded default.

11 · Next

Where to go from here

This is the public extraction of a private framework I use every working day, on my own projects, across different agents and different models. The models change often. The governance around them does not.

Two kinds of experience accumulate at once. What is specific to one repository stays with that repository. What turns out to hold everywhere is promoted once and is there for every project after it. Neither is a chat log, and neither depends on the agent that produced it still being the agent you use.

It moves at the speed the field moves. A methodology that needs rewriting every time a model, an agent or a tool changes is already behind. Adapters absorb that churn, Graphify, RTK and Context7 are optional precisely so no vendor can break the method, and the core stays where it is while the surface around it keeps changing.