fde — the decision before the AI build

Evidence in, an architecture with receipts out, and a refusal when the evidence is not there. Built for Forward Deployed Engineers; useful to anyone who has to decide whether an AI system should exist.

Install · Quickstart · Architecture · Worked example · Contributing · PyPI · Website

fde is an open-source framework for Forward Deployed Engineers: it takes a client engagement from a problem statement to a runnable, deployable AI project — with every decision traced to a fact, and every fact traced to a source.

Not an agent framework — the thing that decides whether you need one, picks it as a stack when you do, and grades what it builds.

Forward deployed engineers arrive with incomplete information, a client who may not know what they need, and a deadline. This is the tooling for that: structured discovery and requirements intake, a decision engine that cites its evidence, and code generation that ends in something you can actually deploy — including on-premise, inside a customer VPC, or fully air-gapped.

pip install fde-framework
fde start acme --statement "Extract fields from supplier invoices."
fde ask acme --role admin        # role-scoped discovery interview
fde architect acme               # topology + chosen approaches (rationale lands in ARCHITECTURE.md)
fde report acme                  # the decision brief, for whoever has to act on it and is not here
fde build acme --out project     # refuses: eight gates guard the build
# ...verify data access, capture the baseline, name the eval owner (each gate prints its remedy), then:
fde build acme --out project     # code + evals + deploy assets + runbook

That refusal at the end is the product working: no baseline, no verified data access — no build. The remedies ship with every gate.

In thirty seconds

You give it the problem as the client put it, the client's own examples, the constraints, a baseline someone measured, and where it has to run. It gives back an architecture decision with its receipts, an exam the build has to pass, an implementation driven inside a fence, a deployment package, a scorecard of what the build can prove, and a record of what was never settled. It refuses to build until the evidence is there, and it can say stop.

pip install fde-framework
fde doctor               # the interpreter, the package, the registry, the extras: measured, not guessed
fde start acme --statement "Route each support message to one of 77 queues."
fde next acme            # the one thing to learn or do next, and what turns on it

One minute, nothing to bring: paste two problems and see which one is decidable yet. For what the whole loop looks like on a real dataset, including the one time the framework told itself to stop, read the banking run.

$ fde triage --statement "Route each inbound support message to one of 77 queues; customer data may not leave our cloud account." \
             --statement "Answer staff questions from 400 policy PDFs inside the firewall."
ranked by what the statement already settles:

  1. Route each inbound support message to one of 77 queues; customer...
     1 fact(s) read, 9% of what gets decided settled, 8 component(s) in scope
  2. Answer staff questions from 400 policy PDFs inside the firewall.
     1 fact(s) read, 6% of what gets decided settled, 5 component(s) in scope

This ranks decidability, not value: the business case comes from the baseline,
and the baseline comes after `fde start`. A candidate ranked low is
under-described, not unworthy.

Inside your coding agent. The same commands are an MCP server, so an engineer in Claude Code or Cursor can ask what to learn before building without leaving the editor:

pip install "fde-framework[mcp]"
fde mcp --check                      # what it would serve, and from where, without serving
claude mcp add fde -- fde mcp        # Claude Code

For Cursor or any other stdio client, the server is the command fde with the argument mcp:

{ "mcpServers": { "fde": { "command": "fde", "args": ["mcp"],
                           "env": { "FDE_WORKSPACE": "/path/to/your/engagements" } } } }

It serves triage, start, frame, status, next, stage, debt, stop_when and history, each returning exactly what the terminal prints, refusals included, beside a small envelope an agent can branch on without parsing prose: schema_version, status (ok, blocked, refused, failed), mutated, and the workspace. The status follows the command's exit, never a reading of its text; mutated is measured, the record's files hashed before and after, so a write that failed partway and left changes says so. Its contract is a JSON Schema, fde mcp --schema, that every reply is tested against; a client pins schema_version, which moves only when a field changes meaning or goes away, never when one is added (it moved to 2 when mutated became a measurement). An engagement is named as start named it (start(name="acme"), then status(engagement="acme")), or by its path. An agent is not a person, so every path it supplies must stay inside one workspace, the directory the server started in or FDE_WORKSPACE; a path outside it, a traversal, or a symlink that points out is refused in words, and engagement names are plain names. The two tools that write to the record, start and frame, say so and carry the protocol's hints so a client can ask first. Calls run one at a time inside the server, so an agent that asks for status and next in the same breath gets both. Exercised against two real clients: the official MCP Python client over stdio, in the test suite, with concurrent calls, an abandoned call and a record changed on disk between calls; and Claude Code 2.1, by hand, from the configuration above. Building, implementing and deploying are not served; they stay on the command line, where the person who runs them is accountable for them.

Five founding engagements

One real workflow, your real examples, a measured baseline, and someone who owns the number. You get a written answer -- build it, change the design, or don't -- with the receipts, and the system where the evidence earns one. Your data stays in your environment. No fee for the first five. How it goes · Bring a workflow

What it does

How it fits together

Discovery narrows an answer space; gates decide whether building is honest yet; the decision engine picks the simplest applicable approach per component and cites why; emit writes a project whose exam fails until it is truly implemented. The full design is in ARCHITECTURE.md.

Bring an engagement

The claim ladder below ends on a rung only engagements can move. The first five that run to the protocol are free, your data never leaves your environment, and a stop is a legitimate outcome. ENGAGE.md says what each side gives and gets, with the agreement in plain language; an engagement proposal is the front door, and fde triage --statement "..." is the first step you can run yourself.

Who this is for

• Forward deployed engineers and solutions engineers delivering AI systems inside client environments, where discovery, deployment constraints and handover matter as much as the model.
• Consultancies and AI delivery teams who want engagement knowledge to compound — every retrospective can enter a shared corpus as an anonymised case.
• Platform teams shipping LLM systems into regulated, on-premise, or air-gapped environments, where "call a hosted API" is not an option and the evaluation has to run where the system runs.

Status: built, demonstrated, unproven

Three different claims, and the differences are the point. Read as a ladder, so nothing here can be over-read:

Demonstrated: four complete engagements have run end to end on real data, all public with every refusal preserved.

• fde-demo-receipts -- 626 scanned receipts: the holdout refused an overfit implementation, and a measured plateau flipped the design from rules to a model, reason on the record.
• fde-demo-complaints -- 2,034 real consumer complaints through the decision shape: the exam refused ambiguous ground truth before it existed, and the loop finished green with the holdout passing, the framework's first.
• fde-demo-rfc-qa -- freeform QA over 58 real RFCs: 100% recall@10 on real queries, then the calibration gate refusing its own judge exactly as the prediction published beforehand said it would (73.7% agreement, refused; the judge's 89.5% was 26 points of flattery over the hand-graded 63.2%).
• fde-demo-banking -- an industry use case: support intent routing for a retail bank over 13,083 real messages and seventy-seven queues. The finished deliverable is an assist-mode router that abstains and says why, at 90.2% on what it routes with 18% handed to a person, scored on a 3,036-case holdout and the vendor's own test split.

Fourteen outside readings have shaped the releases since, each finding pinned as a check before it was fixed; the story, pass by pass, is in HISTORY.md.

Built: the pipeline exists end to end — intake (prose, documents, sample pairs, role-scoped interview, hardware scan) → fact log with provenance → permutation space → eight gates → decide → architect → build (code, evals, deploy and ops assets, RISKS.md, COMPLIANCE.md) → retro and case capture. Overrides are honoured on the next run, trigger observations feed calibration, and a reviewed case can enter the corpus. The test count is stated once, in the claim ladder above, and updated at every release; eleven fresh-eyes audit passes, every finding resolved and the fix pinned as a regression test; CI gates on the suite, lint, and a sanitisation scan of the tree and its history; the evidence corpus is anchored to publicly documented production deployments; every decision is reproducible from its inputs.

Unproven means exactly one thing: no production engagement has yet run through it start to finish. The proof loop is wired and waiting — fde retro captures measured outcomes, human-reviewed cases grow the corpus, and rule revision begins when there are retrospectives to revise against, not before. Pretending earlier would be borrowing rigour rather than having it.

What that means for you today: the generated projects are real and the decisions are defensible, but you are an early adopter, not a reference customer — and the first measured retrospectives will be worth more to this framework than any feature.

Side by side. BENCH.md reads the same figures off all four records with fde bench: stage, the out-of-sample rows on each card, the gap, incidents. The banking run is the one where the whole operating loop has run in public -- two field streams through the delivered build, a drift incident opened and closed on the record, a value document -- and its README says what the deployment was (a laptop) and what the campaign was (a drill).

Stage is the record's lifecycle stage, computed from its evidence by fde stage. production means the build was deployed and attested by name and the field has been journalled through it -- here on a laptop, as the record says -- not that production reliability is proven; the claim ladder above keeps that unverified until a client's production says otherwise.

Two rules

Every intake surface — prose, interviews, scans, sample pairs, the client's existing stack — emits facts into one profile, and two rules make that safe. Arrival order never decides anything — provenance does, and it is dimension-dependent: a measurement outranks anything said about the environment, while a stated requirement outranks a measurement, since you cannot detect a latency budget. And two people disagreeing is a finding, not a conflict to resolve; the dimension is left unresolved and reported, because the gap between what a sponsor believes and what a user experiences is usually the most valuable thing discovery produces.

The full pipeline those rules feed — space pruning, decomposition, the seven gates, decision, emission — is in ARCHITECTURE.md.

What it will not do

Recommend a tool because it is fashionable. plain-python is a first-class option in every pattern, and the schema rejects any pattern that omits it. A two-step linear workflow should not get a graph framework, and the framework has to be able to say so.

Reach for a container by default. Deployment substrate is a ladder from a systemd unit through to Kubernetes. For a single-node on-prem deployment serving one model to a team with no container competence, rung zero is the right answer.

Assume a jurisdiction, sector, topology or stack. Everything is a discovered parameter with a sensible default. A locale pack may pre-set values on dimensions that already exist and attach obligations to the build; it may never introduce a new dimension, because geography changes what you must produce, not how you decide.

Guess. Every claim carries evidence, a date, and a re-derivation rule. Where the framework has no evidence, it says so — undecidable components ship as modules that raise with the reason attached, never as silent gaps.

Install

Prerequisites: Python 3.11+.

pip install fde-framework          # the registry ships inside the wheel
fde start acme --statement "..."   # works from any directory

Extras: pip install "fde-framework[documents]" for PDF/docx intake, [llm] for the hosted-model reader path.

Or from source (contributors — a local ./framework outranks the packaged copy):

git clone https://github.com/atulkapoor/fde-framework.git
cd fde-framework
python3 --version   # must say 3.11+; an older python3 makes pip backtrack for ages instead of failing fast
python3 -m venv .venv && .venv/bin/pip install -e ".[dev,documents]"

Or with uv: uv venv && uv pip install -e ".[dev,documents]"

Optional extras

The core needs none of them: plain-text intake, the decision engine and the build work with zero optional dependencies, which is deliberate — an air-gapped install should not drag in what it will not use. A missing reader refuses by name and says exactly what to install.

File types

Try it

fde start acme --statement "Extract fields from supplier invoices."
fde ask acme --role admin        # bare names resolve to ./engagements/acme
fde status acme                  # gates, gaps, disagreements
fde architect acme               # the design, with rationale
fde build acme --out project     # refuses until gates clear

fde scan acme                    # what this hardware runs
fde cost --requests-per-day 500000 --model-b 70   # dated fleet sizing

fde kb validate   # parse and cross-link the registry
fde kb gaps                        # what the corpus is missing
fde kb sweep                       # profiles no approach can serve

kb validate is strict, because CI runs it and a warning nobody reads is not a check. --lenient exists for the hour when you are mid-way through authoring content and the links do not resolve yet.

A complete worked engagement — real transcript, synthetic client — lives in examples/invoice-extraction.

What a build emits

project/
├── app/                  # components, pipeline, controls, boundary check
│   ├── components/       #   implementations or honest scaffolds — never silent gaps
│   ├── pipeline.py       #   topological order; approval gates before anything mutative
│   ├── controls.py       #   fail-closed gates & critics — when anything mutative was decided
│   ├── boundary.py       #   imported at startup when data may not leave
│   ├── contract.py       #   RefusedInput: forbidden input is refused, never guessed at
│   └── llm.py            #   the one model touchpoint — when a decision needs a model (boundary-gated)
├── evals/                # golden / edge / adversarial sets from the client's own pairs
│   ├── harness.py        #   fails CI until implemented; judge-based when the evaluation decided judged
│   ├── retrieval.py      #   recall@10/50 of the retrieval layer alone — when retrieval answers ranked queries
│   ├── acceptance.md     #   blind UAT protocol for the client's own judges, with the exam record
│   ├── manifest.json     #   split seed, holdout share, SHA-256 of every eval file and the holdout
│   └── load.py           #   p95 against the stated budget (when one was stated)
│   ├── shapes.py         #   the one envelope every step reads and writes; refusals at the door
│   ├── service.py        #   the HTTP edge: identity, request ids, framing, readiness, drain
│   └── ledger.py         #   append-only audit + idempotency keys under STATE_DIR — when anything is outward
├── tests/                # the deliverable's own model-free smoke: contract, fence, empty-exam refusal
├── train/                # when fine-tuning was decided: recorded split, LoRA recipe, before/after on the holdout
├── deploy/               # the substrate that was earned, its full install path, + TEARDOWN.md for all of it
├── ops/                  # runbook (first-five-minutes commands up top), diagnosis walk, SLOs, rollback
├── ARCHITECTURE.md       # scope read-out, decisions, tools & alternatives, agent posture
├── RISKS.md              # every waived gate and overridden recommendation
└── COMPLIANCE.md         # jurisdiction obligations, when a locale pack was applied

Emitted quality is a pinned property, not a promise: the framework's own test suite emits projects across representative architecture shapes and holds every emission to the operational contract, from every environment variable the code reads being documented to a classifier that refuses to serve a constant; the full list of what is held, and the audit passes that put each check there, is in HISTORY.md. A quality finding lands there as a check before it lands anywhere as a fix (tests/test_acceptance.py, tests/test_finetune.py). And the claim is measurable per build: fde scorecard <project> --holdout <file> --external <file> runs what the deliverable can prove about itself and writes SCORECARD.md with a verdict that is a count of rows, never an adjective, and says which rows measure fitness and which measure self-consistency.

After the build: the engagement's operating loop

fde build used to be where the framework stopped. An engagement does not stop there, so the record now carries the rest of it, and every part of it is either computed from evidence or attested by a named person -- never declared by the tool.

• fde stage computes where the engagement stands. Each stage is a set of criteria the record shows or does not: a statement; gates passing or waived, pairs seeded, a holdout drawn, data access attested; a build with its exam; a scorecard whose out-of-sample rows hold and whose edge answered a valid request; a deployment on record with no open incident; an adoption figure measured in the field; a retrospective captured as a case. Every transition is appended to lifecycle.jsonl with its evidence, so time-to-pilot and every reversal can be read back.
• fde drift reads the deployed service's journal -- the same answered lines the emitted service writes -- and compares abstention, decision mix, errors and margins against the exam and the last scorecard. Past a threshold it opens an incident on the record and exits 1. An open incident pulls production back to pilot; fde incident close needs a note saying what was done.
• fde value writes VALUE.md: the automated share and its accuracy from the holdout row, hours and money from the recorded baseline, build and run costs from the caller, payback, and a Wilson interval on the accuracy. Every line says what it rests on, so the number is argued row by row.
• fde deployed and fde outcome are attestations: where it runs, who put it there and which build it is (--project, which every command that reads a project by default then reads); adoption, time to first value, whatever the client measured. fde outcomes prints what the record shows without anyone's opinion -- transitions, days to pilot, loop rounds, reversals, incidents.
• fde stakeholders maps the engagement's people as the record shows them: which of the five roles has been heard (every session carries the role and, when given, the name), who signed what (--by on data-access, security-review, waive, deployed, outcome and incident close), which roles were never asked, and what is on the record with nobody's name on it. fde stakeholder add names the people who have not spoken yet. It is a map, not a contact list.
• fde import turns a client export -- .csv, .tsv, .jsonl, .json, with the client's column names -- into the pairs the intake reads, and reports what it kept, skipped for an empty side, dropped as a duplicate, and counts as verified. Nothing is verified unless the caller says which column and value means a person checked it.
• fde bench reads the same figures off every engagement record, side by side: stage, the out-of-sample rows on its card, the gap, incidents, days to pilot. BENCH.md is the four public demos, computed this way.
• fde history prints every dated entry on the record in order, the undated ones above it, one line each -- the page to read when picking an engagement up.
• fde outcome-contract is the eighth gate: owner, metric, its value today, its target, how it is measured and over what window. Nobody builds until somebody has agreed which number the system exists to move, or waived that with a reason that ships in RISKS.md. The adoption stage then reads the contracted metric back from what was measured in the field.
• fde debt is the decision debt: everything the engagement rests on that nobody has settled -- a gate still failing, a waiver standing in for a condition, a fact the framework guessed or a person merely said where a measurement was possible, two people disagreeing, an attestation with no name on it, a role never asked, an incident open, a component nothing serves -- each with an owner, what blocks on it, and its age.
• fde next now says what hangs on the question it asks: the evidence already on record for that dimension, every candidate answer tried, and the decisions that turn on it -- so the question is asked knowing why.
• fde predict records what the engagement expects before it measures -- holdout_accuracy >= 0.82, field_abstain_rate <= 0.22 -- in the stop-condition grammar, with a confidence if the forecaster has one, and what was in hand when it was made: which figures the record had already measured, and digests of the card and the profile. A forecast about a figure already measured is kept and marked, per figure, because a forecast made after the number is not a forecast. After the score each is held, missed or not measured, with the signed error; once five or more carry a confidence, the held rate is set beside the mean confidence.
• fde implement --sandbox docker runs the coding agent in a container closed by construction: only the project mounted, root read-only with a scratch /tmp, every capability dropped and none gainable, your own user rather than root, bounded memory, CPU and processes, the environment a default allowlist, and no network unless someone with a name and a reason opened it at the terminal (--allow-network "who: why", --env-allow NAME). The emitted ops/agent-policy.yaml sits inside the directory the agent writes to, so it can only narrow: it names the image and bounded limits, and anything wider it asks for is a request the log shows ungranted. The file is part of the fence -- an agent that rewrites it, or plants one, has the change reverted and the round says so. The implementation log records the image digest the agent actually ran on; without the sandbox it says the agent had the host.
• fde experiment runs the protocol in EXPERIMENT.md on the record: start assigns the arm by a seeded draw balanced within the shape, records the engineer, the order and a difficulty vector read off the profile, and freezes the forecasts, the contract and the stop conditions before the build; close reads the measures off the record, or off the control arm's log kept in the template it writes, and lists what is missing; packet renders the blind review in one plain form for both arms; review records the reviewer's fixed form and their guess at the arm; withdraw records an engagement leaving with the arm it drew, so the report can count withdrawals by arm; report pairs within shape by the nearest frozen difficulty and prints the distance, sets the primary endpoint beside the secondaries and beside a floor line for any score under three, breaks the figures out by engineer and by half of the series, gives the blinding accuracy per packet version, and names the rung of evidence the count has reached.
• fde stop-when records what evidence would stop the engagement -- answered_accuracy < 0.88, abstain_rate > 0.25, adoption < 0.4 -- over figures the record measures: the scorecard's out-of-sample rows, the field journal, the outcomes recorded in the field. A triggered condition makes STOP the engagement's stage, on the record with the trigger and the threshold, until the condition is restated with a reason, the build is changed and scored again, or the case is captured. A figure the record has not measured leaves its condition unjudged and says so.

What this still is not: a live connector to a ticketing or data system (an export is the interface, and the only one testable without a client's credentials); a benchmark corpus (four public demos are four rows, and the bench says so); collaboration beyond a git-native record with names on it. Each waits for engagements that have not happened yet.

The full lifecycle, copy-paste

Everything below runs from an empty directory after pip install fde-framework:

fde start acme --statement "Extract fields from scanned supplier invoices; \
data cannot leave; 200,000 documents, 8,000 verified; a person is waiting."
# plays back the typed facts it read, and the three questions worth asking next

fde ask acme --role eval_owner       # answer what discovery still needs
fde status acme                      # facts by scope, gates, % settled

cat > baseline.yaml <<'YAML'
volume: {value: 20000, unit: docs/month, definition: invoices received by AP}
cycle_time_per_unit_seconds: {value: 300, unit: s, definition: arrival to posted}
labour_hours_per_week: {value: 35, unit: h/week, definition: AP team keying time}
rework_rate: {value: 0.1, unit: ratio, definition: entries corrected after post}
exception_rate: {value: 0.07, unit: ratio, definition: routed to a human queue}
error_rate: {value: 0.04, unit: ratio, definition: wrong amount or vendor posted}
business_metric: {value: 9, unit: days, definition: mean days payable outstanding}
sampled: {n: 40, method: random invoices across two months}
YAML
fde baseline acme --file baseline.yaml
fde data-access acme --note "read replica returned 14 real rows"
fde security-review acme --note "client infosec walked the data paths"
fde ask acme --role eval_owner       # or: fde waive acme client_readiness --reason "..."

fde build acme --out project         # refuses until the gates truly pass
python project/evals/harness.py      # red: empty until pairs are seeded, then red until implemented -- that's the exam
fde implement project                # drive a coding agent until it's green
fde scorecard project --holdout engagements/acme/artifacts/holdout.jsonl   # measure it

fde stage acme --project project     # where it stands, computed off the record
fde deployed acme --note "runs in the client VPC; platform team put it there" --project project
fde drift acme --journal service.log --project project   # the field against the exam
fde incident acme close inc-001 --note "holdout re-scored; pairs redrawn"
fde outcome acme --metric adoption=0.62 --note "support lead's dashboard, week 3"
fde value acme --project project --hourly-cost 40        # VALUE.md, line by line

Python API

The CLI is a thin layer; everything is importable. The registry loads from the installed wheel, so this runs anywhere:

from fde.architect import architect
from fde.intake.prose import parse_prose
from fde.models.profile import Profile
from fde.registry import default_root, load_registry

registry = load_registry(default_root())

profile = Profile()
profile.ingest(parse_prose(
    "500,000 scanned invoices; data cannot leave; 10,000 verified; "
    "a person is waiting; structured records out.", registry))

design = architect(profile, registry)
print(design.topology)                    # customer-vpc
for component, decision in sorted(design.decisions.items()):
    if decision.approach:
        print(component, decision.approach, decision.rationale)

Every decision object carries its rationale and its rejected alternatives — the same receipts the emitted ARCHITECTURE.md prints.

Common commands

Why not RAGAS or TruLens?

Deliberately. Their headline metrics are judge-scored, and an uncalibrated judge is the failure mode this framework has now measured first-party: a local judge inflated results by 26 points before the calibration gate refused it (the run is public). The emitted evals are seeded from the client's own verified examples, stdlib-only — they run inside an air gap and hand over with zero dependencies — reference-based with discrete verdicts, and no judged number is quoted before the judge beats a human-agreement bar against the named eval owner. If your team wants RAGAS or TruLens dashboards alongside, point them at the same golden pairs — the JSONL is the same shape. The gate a delivery is graded on stays calibrated, or stays silent.

Troubleshooting

no registry here — you passed --registry/--root at a directory that holds no registry. Drop the flag (the corpus ships inside the package) or point it at the framework/ of a source checkout.

build refuses with gates listed — that is the point. fde status names each gate and its remedy; soft gates take fde waive <gate> --reason, data access takes only credentials that returned real rows.

A component module raises UndecidedComponent — nothing could be decided for it; the raise message names the unanswered question. Answer it and rebuild — holes are loud here, never silent.

fde kb sweep shows undecidable profiles — some are honest contradictions (data cannot leave + nobody to operate). fde architect on that profile names the conflicting facts.

The evaluation harness fails CI — it evaluates the emitted pipeline; it fails until the components are implemented end to end. A gate that cannot say no is not a gate.

Design

The registry under framework/ is data, not code. Adding a stack, a pattern, a locale or a case is a file, never a change to src/. That constraint is enforced by tests, and it is the thing that keeps the framework general rather than gradually becoming one consultancy's tooling.

Patterns are separated from stacks because patterns are stable for years and the libraries implementing them churn in months. A pattern says what; one realization per stack says how, as a template plus a claim to satisfy a typed interface. Swapping the stack changes the emitted code and not the architecture, and there is a test asserting exactly that.

FAQ

What is a Forward Deployed Engineer? An engineer who works inside a client's environment to deliver a working system — part solutions architect, part implementer, part translator between what a client asks for and what they need. The role is common in AI companies shipping into enterprises; this framework encodes the craft of running such an engagement well.

Does the framework itself call an LLM? No. Intake parsing, decision-making and code generation are deterministic — the same profile always produces the same project, so a diff between two builds means a decision changed. LLMs appear in the generated systems where the profile justifies one, behind interfaces that make them swappable.

Does it work air-gapped? Yes, by design. The framework runs from plain files with no server or network dependency, the registry knows which stacks can run inside an air gap, and the offline-evaluability gate refuses a design whose metric cannot run where the system runs.

How is this different from a project template? A template gives everyone the same starting point. This decides — from discovered facts, with cited evidence and named rejected alternatives — and then generates. Two clients with different constraints get different architectures, and the document explains why.

How does it improve over time? Every engagement captures its overrides (when the FDE chose differently), trigger calibration (did predicted graduations fire?), and an anonymised case. Cases enter the corpus only after human sanitisation review; rules are revised only when a corpus of outcomes exists — capture now, revise later, never pretend.

What does "self-evolving" mean here, concretely? Three recorded signals — overrides, trigger observations, case outcomes — and a human-gated path from a retrospective into the shared knowledge base. Nothing in framework/ changes by itself; the corpus grows, and revision against it is a deliberate, evidenced act.

Privacy

Everything runs from plain files on your machine. The default path makes no network calls, has no telemetry, and never transmits engagement content anywhere — it works on a plane and inside an air gap, and a text editor is always a legal way into its state. Discovery, decisions, and builds never call an LLM.

Four commands are the deliberate exceptions — each opt-in, each governed by the boundary doctrine (hosted models refused unless the engagement states data may leave; local endpoints always allowed):

• fde frame --reader llm — a model proposes facts, at the weakest provenance
• fde kb suggest — mines a brief for recogniser gaps, proposing (never applying) vocabulary
• fde implement — drives a coding agent you name
• the judge-based eval harness in generated projects whose evaluation decided judged (LLM_ENDPOINT, hosted path refused inside a boundary)

Nothing calls a model silently, and fde scan recommends a local model sized to your hardware so none of it needs to leave the machine.

Engagement directories (client facts, baselines, gate state) are excluded from version control by construction and enforced in CI — along with credential patterns, personal-data patterns, and a check that no unreviewed case can ever be committed.

Team setup

The registry is the shared asset; engagements are private working state.

• Share framework/ — fork or clone it as your team's knowledge base. Every dimension, approach, stack and case is a markdown file; review registry changes like code, because they decide architectures.
• Never commit engagements/ — client facts stay local. The repository's own .gitignore and CI sanitisation gate enforce this shape; keep it in yours.
• Grow the corpus deliberately — fde retro captures a case, fde kb ingest-case lands it as sanitization: pending, a human reviews it for anything identifying, and only reviewed cases can be committed. One reviewed case per delivered engagement compounds fast.

Roadmap

• Rule revision from outcomes — capture is wired end to end; revision deliberately waits for a corpus of measured retrospectives rather than pretending a handful is evidence.
• More locale packs and stacks — both are data; contributions enter against CONTRIBUTING.md's contract (and the code of conduct).
• Language, channel, and device axes — six of twenty industry test statements named regional languages, low bandwidth, or basic devices; the honest wiring (per-language evaluation, SMS/IVR serving approaches, safeguarding governance) is a corpus milestone, not a checkbox.
• Scale words — "millions of applications", "tens of millions of players": refusing to guess a number from "millions" is doctrine, and a better answer than refusal is still owed.
• Capability-verb extraction — "update the claims system of record" implies an integration no regex can count; the LLM reader proposes facts today, and component hints are its natural next job. The honest gaps list lives in the tool itself: fde kb gaps and fde kb sweep report what the corpus is missing and which profile shapes no approach can serve yet.

Documentation

Development

Set up as in Install → from source (python3.11+), then:

.venv/bin/pip install -e ".[dev,documents]"
.venv/bin/pytest -q          # the full suite, about two minutes
.venv/bin/ruff check src tests

The registry is data: most contributions are a markdown file in framework/ plus a test that pins the behaviour. CI additionally runs a sanitisation sweep over the tree and history.

Community

Questions and engagement war stories → Discussions. Bugs and corpus gaps → issues (the forms ask for evidence, the way the framework does). Conduct → CODE_OF_CONDUCT.md.

License

Apache 2.0 — chosen for the explicit patent grant, because enterprise legal review is a real gate for the audience this is for.

Contributing

See CONTRIBUTING.md. The short version: contributions enter against a contract, and client material never enters this repository — only patterns re-expressed in the framework's own words. Sanitisation is enforced in CI: allowed paths only, history checked, credential and personal-data patterns, and no unreviewed case can be committed.