Skip to the content.

Benefit–harm standard for defaults

Status: shipped doctrine; automated default-admission checker is not yet shipped
Scope: any behavior FAK enables without an operator deliberately selecting it, plus any change that broadens such behavior
Related standards: net-true value, observer effect, support-maturity honesty fence

Rule

A default is an intervention, not free convenience. Enable it only when its expected benefit over the real next-best behavior outweighs its total expected harm for the population that receives it. Use the least invasive effective setting, expose important residual risks, monitor effects after release, and define a stop or rollback rule before broadening exposure.

“First, do no harm” is a useful instinct, but an absolute no-harm test would reject every useful intervention. Medicine instead contributes a stronger operating model: compare benefits and harms, account for uncertainty and patient circumstances, reduce avoidable harm, watch real outcomes after introduction, and change course when the balance moves. For FAK, the “patient” is the operator and workload; the relevant outcomes include task correctness, authority, privacy, latency, cost, attention, recoverability, and trust.

This standard complements net-true value. Net-true asks whether a claimed gain survives all costs against the right baseline. Benefit–harm asks the prior product question: should users be exposed by default, at what dose, with which contraindications, monitoring, and withdrawal criteria?

The agentic default window

FAK must avoid two symmetric failures:

The agentic default window is the set of capabilities that current evidence supports exposing automatically for a named workload and authority envelope. It is deliberately not a cultural popularity poll. The window moves as models, tools, safeguards, operator expectations, and field evidence change. A capability can move into the window when safeguards make default use favorable; it can move out after adverse evidence.

Every opt-in gate therefore needs a reason and a review date. “Documented reason” without re-review turns uncertainty at one moment into permanent product inertia. fak score default-value monitors this boundary: unreviewed value flags are omission debt, and expired gates return to the work list instead of silently remaining opt-in forever. The review must choose one action: safely default it on, renew the gate with current evidence and a new date, or exclude/retire it.

Great default value is not the fewest defaults. It is the largest evidence-supported set that works without setup while preserving authority, reversibility, and material operator choice.

Admission record

Before adding or widening a default, record all nine fields. A prose design note, issue contract, or test table is sufficient if each answer is explicit and witnessed.

  1. Indication — Which concrete user problem and workload population is this intended to help?
  2. Comparator and non-intervention — What happens with the current default, including doing nothing? Compare against the strongest practical alternative, not a strawman.
  3. Benefit — Which user outcome improves, by how much, for how many workloads, and how soon? Separate measured outcome from mechanism proxy.
  4. Harms and interactions — Include correctness loss, hidden work, stale state, authority expansion, privacy leakage, latency/cost tails, lock-in, operator confusion, and interactions with other stacking defaults. Estimate severity, probability, duration, and affected population.
  5. Uncertainty — Which populations and tail cases are missing from the evidence? Uncertainty lowers admissible exposure, but it also has an expiry: state what evidence or review date will reconsider an opt-in gate. Popularity does not convert missing evidence into safety, and old uncertainty does not justify permanent inertia.
  6. Contraindications — Name conditions where the behavior must remain off or fail closed. Do not make every user discover these through an incident.
  7. Dose and safeguards — Choose the narrowest scope, frequency, authority, and resource budget that produces the benefit. Prefer preview, bounded trials, read-only operation, idempotence, isolation, and an observable fallback.
  8. Consent and control — State whether the operator can understand the material tradeoff, inspect activation, override it, and return to the prior behavior. Irreversible or authority-expanding behavior requires deliberate opt-in even when average benefit is positive.
  9. Surveillance and movement rules — Name the benefit and harm signals collected after release, the review horizon, the threshold that disables or narrows the default, the tested rollback path, and the threshold that promotes an opt-in capability when omission harm now dominates. Aggregate success must not hide severe tail harm.

Verdict

Assign one exposure class. “Useful” alone is never a DEFAULT verdict.

Class Use when Required posture
DEFAULT Representative evidence shows a favorable balance; serious harms are prevented or bounded; rollback and surveillance exist. Enable at the minimum effective dose. Publish activation and material residual risk.
CONDITIONAL DEFAULT Benefit is favorable only when a machine-checkable indication is present. Detect the indication, refuse on contraindication or uncertainty, and fall back safely.
OPT-IN Value is workload-specific, evidence is immature, the action expands authority, or harms may be material/irreversible. Require deliberate selection; provide preview or bounded trial where possible.
EXCLUDE The same outcome has a less harmful route, serious harm cannot be bounded, benefit is not net-true, or there is no credible surveillance/withdrawal path. Do not ship as an offered behavior until the balance changes.

A default graduates by evidence: opt-in trial → conditional default → broad default. It can move backward at any time. Rollback is a normal lifecycle transition, not an admission of failure.

Applying the standard to FAK’s improvement families

These are default postures, not permanent verdicts. A concrete feature can move only with its own admission record.

Improvement family Expected benefit Material side effects Starting posture
Structural policy and capability floor Blocks unauthorized tool calls before model persuasion can matter. False denials, workflow interruption, and a misleading sense that policy covers semantics it does not inspect. DEFAULT for the minimum deny floor; capability expansion stays explicit. Keep an allow witness and explain the denial.
Exact tool-result or prefix reuse Avoids repeated latency, tokens, and external work. Stale answers, cross-session or cross-tenant disclosure, and replaying a result after hidden state changed. CONDITIONAL DEFAULT only with scoped identity, deterministic/declared-safe inputs, freshness rules, and miss fallback; otherwise opt-in or exclude.
Context shedding, elision, and compression Extends useful sessions and lowers provider input cost. Removes evidence, instructions, provenance, or an anomaly the model still needs. Harms can be delayed and hard to attribute. CONDITIONAL DEFAULT at a conservative budget with protected invariants, restore path, semantic-quality surveillance, and stop thresholds. Never infer safety from token reduction alone.
Model routing or substitution Lowers cost/latency and can reserve stronger models for harder turns. Silent capability loss, changed safety behavior, nondeterministic quality tails, and provider/data-boundary changes. CONDITIONAL DEFAULT only inside a witnessed non-inferiority envelope with escalation fallback. Provider or data-boundary changes require explicit operator control.
Speculative tool execution and prefetch Hides tool latency. Executes unwanted work, duplicates mutation, spends budget, leaks intent, or races the chosen branch. OPT-IN until side-effect freedom, cancellation, deduplication, budget, and privacy are proven. Mutating or externally visible calls are contraindicated by default.
Retry, resume, and autonomous continuation Recovers progress from transient failure and reduces supervision. Duplicate effects, runaway spend, stale-intent continuation, and repeated harmful action. CONDITIONAL DEFAULT only for idempotent or deduplicated steps with bounded attempts, freshness checks, and visible stop control; authority expansion is opt-in.
Telemetry and post-release learning Detects regressions and rare harms that pre-release tests miss. Privacy loss, observer overhead, retention risk, and metrics steering the product away from user outcomes. DEFAULT only for minimum-necessary, scoped, inspectable telemetry with retention limits; sensitive content collection is opt-in. Monitor the monitor’s cost and effect.
Destructive automation or automatic landing Removes toil and shortens completion time. Data loss, unintended publication, supply-chain impact, or an irreversible shared-state change. OPT-IN behind preview, narrow capability, witness, and tested recovery. Exclude actions with no credible rollback unless the operator explicitly accepts that condition.

Evidence and review rules

Worked admission: speculative mutating tool calls

Medical-industry sources borrowed, and limits

Accessed 2026-08-17:

The analogy has limits. FAK is not a medical device, users are not patients, these sources do not confer clinical or regulatory compliance, and software harms are not commensurable with bodily injury. We borrow the decision structure—indication, comparator, benefit, adverse effects, uncertainty, contraindication, dose, consent, surveillance, and withdrawal—not medical authority or terminology as decoration.