Gain Floor/Ceiling Rule¶
Decision rule — instantiates Adaptive Gain Retuning
A hard clamp on the gain parameter — a floor that stops the pathway going numb and a ceiling that stops it running away — that the adaptive loop may retune within but never breach.
The adaptive loop is allowed to move the gain, but not anywhere it likes. The Gain Floor/Ceiling Rule is the hard bound on the gain parameter itself — a minimum below which sensitivity may never fall and a maximum above which it may never rise — enforced no matter what the input statistics seem to justify. Its distinguishing move is that it constrains the gain value, not the output and not the update timing: whatever the automatic loop computes, the rule clamps that result into a pre-authorized band before it takes effect. The floor exists so a long quiet spell can't drive sensitivity toward zero and leave the pathway numb; the ceiling exists so a noisy or adversarial spell can't drive amplification into clipping, flooding, or runaway. It is the mechanism that makes "adaptive" safe by making it bounded.
Example¶
A bank runs a real-time transaction-risk scorer whose gain — how steeply risk features translate into a decline — is retuned continuously as fraud patterns shift. Left unbounded, two failure directions loom. In a quiet week the loop, seeing almost no fraud, keeps lowering gain to avoid annoying customers, until the model is effectively asleep and a fraud burst walks straight through. In a noisy week — say a data-quality glitch inflates a feature — the loop cranks gain up and starts declining a large slice of legitimate customers. The Gain Floor/Ceiling Rule pins both ends: gain may never fall below the level that still catches known high-risk signatures (the floor), and never rise above the level that would decline more than ≈2% of traffic (the ceiling). Within that band the adaptive loop runs free; at the edges the rule wins. When the noisy week pushes the computed gain past the ceiling, the model clamps at the ceiling and raises a flag for review — rather than quietly blocking a chunk of the customer base.
How it works¶
- Set the bounds from consequences, not convenience. The floor is fixed by the weakest signal that must still be caught; the ceiling by the output rate (clipping, decline rate, alert volume) that downstream capacity and users can actually absorb.
- Clamp after the loop, before the effect. The adaptive loop proposes a gain; the rule saturates that proposal into
[floor, ceiling]; the clamped value is what the fast pathway uses. - Treat a bound-hit as a signal, not a silence. Repeatedly clamping at a limit means the world has moved outside the authorized band — an event to surface, not to absorb.
Tuning parameters¶
- Floor level — how low sensitivity may go. Raise it to guarantee weak-signal capture in quiet periods; lower it to cut false positives, at the risk of numbness.
- Ceiling level — how high amplification may go. Lower it to protect against clipping, flooding, and runaway; raise it to allow aggressive response in genuinely high-stakes regimes, at the risk of saturation.
- Band width — the room the adaptive loop has between floor and ceiling. A narrow band is safe but barely adaptive; a wide band is responsive but leans hard on the loop's judgment.
- Bound-hit response — clamp silently, clamp-and-flag, or freeze and escalate. The stronger the response, the less a persistent out-of-band condition can hide.
- Hard vs. soft bound — an absolute cap, or a penalty that makes approaching the limit progressively costly.
When it helps, and when it misleads¶
Its strength is that it converts an open-ended adaptive loop into a bounded one: however badly the input statistics are misread, sensitivity stays inside a band someone deliberately authorized, and a bound-hit becomes a visible event rather than a silent drift into numbness or saturation.
Its failure mode is bounds set for the wrong reason. Too wide, and the rule authorizes exactly the numbness or clipping it exists to prevent; too narrow, and it strangles the adaptation until the system might as well be fixed-gain. The classic misuse is running the bound backwards — loosening the ceiling after the fact so the loop's already-chosen aggressive gain reads as "within policy," which is Goodhart's law playing out on the safety limit itself.[n1] The discipline that guards against it is to derive the floor and ceiling from stated consequences (the weakest signal that must be caught; the output rate that can be absorbed) and to change them only through the same reviewed, logged process as any other gain change — never quietly, to fit a number.
How it implements the components¶
gain_floor_and_ceiling— it is this component: the explicit minimum and maximum on sensitivity, enforced against every proposed update.gain_parameter— it operates directly on the gain value, saturating the loop's proposal into the authorized band before the pathway uses it.
It does not measure whether output is actually clipping or going numb — that occupancy read belongs to the Saturation Occupancy Dashboard and the High-Load Clipping Test; it does not decide the moment-to-moment gain (that's the Automatic Gain Control Loop) or record each change for review (the Gain-Change Review Log).
Related¶
- Instantiates: Adaptive Gain Retuning — supplies the hard bounds inside which every other gain mechanism must operate.
- Consumes: High-Load Clipping Test and Weak-Signal Recovery Test — their results calibrate where the ceiling and floor belong.
- Sibling mechanisms: Automatic Gain Control Loop · Hysteretic Gain-Update Filter · Saturation Occupancy Dashboard · Gain-Change Review Log · Fixed-Gain Degraded Mode · Contextual Gain-Scheduling Table
Editorial Notes¶
Form Classification¶
Form family: Control, Automation & Runtime
Rationale: Gain Floor/Ceiling Rule operates as a live operational control that automatically routes, enforces, adapts, or responds during execution because it a hard clamp on the gain parameter — a floor that stops the pathway going numb and a ceiling that stops it running away — that the adaptive loop may retune within but never breach.
Independent corroboration: The frozen evidence defines Gain Floor/Ceiling Rule as 'A hard clamp on the gain parameter — a floor that stops the pathway going numb and a ceiling that stops it running away — that the adaptive loop may retune within but never breach', so its operative form is Control, Automation & Runtime.
Nearest alternative: Rule, Policy & Commitment — The clamp automatically enforces gain bounds at runtime, while the floor and ceiling are standing constraints.
Review outcome: Independent reviewer agreement; medium confidence.
Origin Attribution¶
Primary origin: Engineering & Design
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Multi-domain
Rationale: Bounded adaptive control and actuator clamps are standard control-engineering safety mechanisms.
Related originating lineages:
- Systems Thinking & Cybernetics — Cybernetic feedback theory supplies the adaptive-gain loop being bounded.
Review resolution: Both reviewers agree that engineering_design is primary: Bounded adaptive control and actuator clamps are standard control-engineering safety mechanisms. I retain systems_cybernetics only where the reviewers identify formative lineage, not every later application. I resolve origin_mode as cross_disciplinary_synthesis because the artifact joins distinct disciplinary contributions. I resolve domain_reach as multi_domain because it transfers across several fields but is not a domain-free primitive. Encyclopedia synthesis is false because the exact generalized packaging is already established enough that encyclopedia-specific synthesis is not required.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
The rule is a constraint, not a controller: it never moves gain toward a target, it only forbids regions. That is why it pairs naturally with the Fixed-Gain Degraded Mode — the rule bounds a loop that is still running and trusted, while the degraded mode replaces the loop entirely when it can no longer be trusted at all. Between them they cover both ways adaptation goes wrong.
[n1] Goodhart's law — "when a measure becomes a target, it ceases to be a good measure." Applied here: once the gain bound is adjusted to make a chosen gain look compliant, the bound stops measuring safety and starts certifying whatever was going to happen anyway. ↩