Direct answer
The digital Four Fs translate adaptive pressure into an engineering loop: Feed evidence, Fork bounded variants, Fight with independent evaluation, and Flee from variants that fail cost, safety, or quality gates. No-op is always allowed when change does not improve net viability.
Why this cycle matters
The Digital Four Fs translate a biological survival shorthand into software operations that can be measured, tested, and governed. Feed acquires useful evidence and resources. Fork generates controlled variants. Fight evaluates candidates under selection pressure. Flee withdraws from losing, unsafe, or wasteful states.
This cycle is not a branding layer over ordinary retraining. It is an operational discipline for deciding when an adaptive system should change and when it should hold still. The important move is to make Flee and no-op first-class outcomes. Without them, a breeding system becomes an accumulation engine that loads more modules, keeps stale descendants, and eventually collapses under memory, latency, or maintenance cost.
The four operations
| Operation | System question | Code-breeding example | Model-breeding example | Failure when absent |
|---|---|---|---|---|
| Feed | What evidence or energy keeps the ecology useful? | collect failing traces, tests, profiles, incidents | collect examples, feedback, retrieval evidence, compute | starvation and stale competence |
| Fork | What bounded variants should be explored? | branch a routing policy, synthesize a patch | train an adapter, distill a specialist | stagnation and poor adaptation |
| Fight | Which variants survive independent selection? | run tests, fuzzers, security scans, latency targets | run task, calibration, drift, and robustness suites | arbitrary variation and metric gaming |
| Flee | What should be rolled back, quarantined, unloaded, or retired? | disable a feature, delete a broken branch | unload a failing model, prune a coalition | bloat, fragility, and escalating cost |
The loop as governance
Treat the cycle as a release protocol, not a metaphor. Every transition needs evidence. A fork without feed is random. A fight without an independent evaluator is self-reference. A flee without records becomes silent degradation. Feed is therefore not just data collection; it is the evidence foundation for all later decisions.
PROCEDURE digital_four_fs_cycle(ecology, evidence_stream, policy)
feed_packet <- CURATE_EVIDENCE(evidence_stream, policy.data_rules)
candidate_set <- []
IF feed_packet.justifies_exploration
candidate_set <- FORK_BOUNDED_VARIANTS(ecology, feed_packet, policy.mutation_budget)
END IF
scorecards <- []
FOR each candidate IN candidate_set
scorecards.ADD(FIGHT_WITH_INDEPENDENT_EVALUATORS(candidate, policy.evaluation_suite))
END FOR
decision <- SELECT_OR_NOOP(ecology, scorecards, policy.viability_threshold)
IF decision.action == "promote"
PROMOTE_THROUGH_RELEASE_STAGES(decision.candidate)
ELSE IF decision.action == "flee"
RETIRE_OR_QUARANTINE(decision.target, decision.reason)
ELSE
RECORD_NOOP(decision.reason)
END IF
END PROCEDUREPractical interpretation
Use the cycle initially for offline experiments. Collect a batch of failures, fork candidate specialists, evaluate them with frozen tests, and either promote one through shadow mode or do nothing. Online systems can use the same vocabulary later, but the governance burden rises sharply because feed, fork, fight, and flee occur near live users.
Links into the curriculum
After this page, read The two Four-F frameworks, What model breeding means, and The viability controller.
Source reports used for this guide
These reports are preserved verbatim in the site archive. The guide above is an editorial synthesis and may narrow, qualify, or reorganize claims from the source material.