About FDE training

Role and accountability first

AFDE works in customer contexts, implements systems, handles ambiguity and owns verifiable production outcomes. Titles cannot replace responsibility agreements. Engineers define access, business boundaries, acceptance criteria and escalation. Code generation is an implementation tool.

The super intern metaphor

Models read and generate quickly, but do not know project history or own outcomes. The metaphor reminds practitioners to guide and verify, rather than describe every model capability. Supply real context, explicit constraints and reproducible criteria, then inspect implementation.

Three traps

Self-consistent hallucination: plausible output need not be factual, and compilation need not establish acceptance.

Context entropy: repeated failures and discarded logic can crowd out relevant facts. Stop when constraints are lost, save verified state and restore context. A long conversation does not inevitably fail.

Local optimization: a repair can break module contracts elsewhere. Inspect impact, dependencies and regression; decide rebuilding from evidence and total cost.

Distinct evidence

Software tests verify code behavior; evaluations assess probabilistic outputs. Handover, real adoption and business impact require separate evidence. Method benefits are project hypotheses, rather than speed guarantees.

What cognition is

Cognition is a progressive, incomplete mapping of the possibility base.
The world offers far more possibilities than any single observation.
Observation does not exhaust the world; it narrows a corner of it.

Observational convergence

Observe first, then judge.
Each observation eliminates a share of the possibilities.
Repeated observations converge the remaining set to an actionable range.
Convergence is not the end; it is the starting point for the next step.

Observational locking

The causal chain grows node by node.
Each new node is verified before it is trusted.
Verified nodes lock into the chain and become the ground for further reasoning.
Unverified nodes remain suspect.

Controlled generation

Blueprints and acceptance criteria make intent explicit. They constrain implementation and provide a basis for verification. They do not make model output deterministic: matching context does not guarantee matching output or correctness.

Applicability

RC is a philosophical lens for observation and decision-making, rather than a measured guarantee about models. The method uses these principles to organize hypotheses and evidence. Software tests, model evaluations and customer observations must still establish outcomes independently.

Asynchronous work and sharing boundaries

Tasks and documents record goals, evidence, decisions and next actions. Default openness applies inside authorized spaces. Customer data, private information and confidential records do not become shareable merely because a channel is internal. Use synchronous interaction for observation, urgent coordination and difficult conversations, then record conclusions.

Trust through predictable delivery

Trust rests on reliable responses, prompt disclosure of bad news and verifiable delivery. Every task has an owner, status and escalation route. Autonomy and review depend on risk; sensitive authentication, data and production changes require appropriate human review.

Scoped autonomy

Clear blueprints and authorization enable independent work. Skills advance defined tasks. Authorized people decide business trade-offs, new data domains, releases and customer commitments. Do not lower standards to pass failures.

Task lifecycle and knowledge feedback

Retain a trail from clarification and execution through acceptance and closure. Commits, blueprints and changes enable future owners to reconstruct decisions. Field notes separate facts from hypotheses, sharing calibrates boundaries and reusable assets have versions and maintainers before field revalidation.

Outcome accountability

Team measures support internal improvement; verify customer outcomes separately. Faster delivery, passing tests, successful deployment and sustained adoption are distinct evidence.