Adaptability without unpredictability

One week the Meal Companion composes leisurely, exploring variety. The next week it snaps instantly to a compressed busy-week strategy. A household member asks: "Did something break? It feels random."

Nothing broke. The guard passed and Busy Week Pattern ran. Because the strategy switch was invisible, adaptation felt like unpredictability. Users want systems that meet changing context. They also want to predict how behaviour will change when context changes.

Adaptability without unpredictability is the design goal: legitimate compression and context response that follow visible rules, not mood swings between sessions.

Helpful context: Cognitive posture and trust across time set expectations. Applicability conditions govern legitimate strategy switches. Deliberation and automaticity explains two speeds.

What makes adaptation unpredictable

Unpredictability is change without notice or without governed triggers, and it erodes trust faster than outright errors because users cannot form mental models of behaviour. Speed differences alone are fine when users understand why the path changed. Different standards week to week with no exposed assumptions feel capricious. Random plan style each run feels capricious even when meals are edible. Silent promotion of chat preferences into policy feels capricious because the user never authorized durable change. Expertise activating with no user-visible guard result feels capricious because compression appeared from nowhere.

Adaptation should follow visible rules of conduct and bounded strategies.

Introduction to Thoughtware · Ch. 13

The distinction sharpens with examples. Adaptive and predictable: the system shows "Using Busy Week Pattern," busy evenings are anchored, leftovers are included when the guard passes. When the guard fails because an unfamiliar dietary mention appeared, the system shows "Full planning, input required on medical frontier." The user connects label to behaviour. Adaptive and unpredictable: the same inputs produce wildly different meal philosophy because temperature changed or prompt drifted. The user cannot reproduce or contest behaviour. Support cannot replay the decision.

Switching to Busy Week Pattern after guard pass is adaptive and predictable when the conduct spec says so and UI or summary reflects it. Random plan style each run is neither adaptive nor trustworthy.

Guards license compression

The Meal Companion Busy Week Pattern illustrates governed adaptation. Before reuse, the guard checks: at least two low-time evenings identified, busy evenings clearly marked, no unfamiliar medical restriction present, leftovers acceptable if strategy requires them, temporary request not conflicting with skeleton. Guard passing licenses compressed strategy: anchor busy evenings, place spinach before Wednesday, use leftover subplan, compose remaining evenings, run checks, patch locally if one weakness remains. Guard failing returns work to full deliberation. That return is also adaptation, and it is legible, not silent.

When users see only output style change without guard outcome, they infer model mood rather than architecture. Visible labels such as "Using busy-week strategy" versus "Full planning mode" are cheap trust investments for material switches. Not every internal optimisation needs UI, but material switches do. Material switches change effort assumptions, safety checks applied, authority posture, or promotion rules. Immaterial switches might include candidate ordering or internal retrieval depth. When strategy class changes materially, user-visible labels tied to conduct spec language keep the trust contract intact. Silence on material switches is how adaptive systems acquire reputations for moodiness even when engineering logs show clean guard logic.

Recovery is part of trust interacts here. Patch-first recovery feels like the same product correcting itself. Regeneration without explanation feels like a new product arrived. Unpredictability often starts in architecture with unguarded shortcuts and finishes in UX with no explanation. Fixing it requires both: versioning expertise with applicability guards documented beside strategy skeletons and writing conduct spec language for when compression activates and what users see.

Memory, speed, and confidence

When memory changes behaviour, users see what changed and whether it is durable endorsed knowledge versus this-run context. Silent promotion makes adaptation look like unpredictability. The system behaves differently next month because chat accumulated bias, not because the household changed policy. Endorsement events are visible when they alter standards. Knowledge is endorsed, not retrieved owns promotion logic. This page states the trust requirement: adaptation driven by durable change is legible.

Deliberation and automaticity explains why two speeds exist. Compression is valuable when licensed. The trust requirement is that fast paths remain inspectable and revert cleanly when guards fail. Users accept speed when they understand triggers. They reject speed when it feels like the system stopped caring on random days.

Systems that perform certainty while switching strategies silently combine two trust failures. Confidence theatre plus capricious adaptation teaches users to ignore both labels and numbers. Grounded adaptation exposes assumptions and guard outcomes instead of effusive praise when compression runs. Memory discipline in conduct enforces the boundary so context does not silently reshape durable behaviour across runs.

What this looks like in practice

The most visible strategy switches in a product map directly to this framework. Busy week compression: the trigger is guard pass on calendar and dietary inputs, the signal is a strategy label, and the test asserts that guard false prevents pattern activation. Medical mention: the trigger is an unfamiliar restriction detector, the signal is a deferral template, and the test asserts full deliberation path. Temporary pasta avoidance: the trigger is a contextual instruction only, no signal is required if the instruction is not promoted, and the test asserts no durable store write. Temperature or prompt-only variance: this trigger has no legitimate place in material behaviour and belongs as versioned expertise or explicit user mode selection.

When Busy Week Pattern updates, version number and changelog appear in internal review and optionally in user-facing release notes when standards change materially. Users who trusted last month's compression deserve notice when effort thresholds shift. Expertise versioning pairs with guard tests, because a v0.3 pattern activating under v0.2 guard assumptions produces unpredictability even when both files exist.

Household life has seasons: exam weeks, holidays, guests. Adaptation may legitimately change without new endorsed knowledge when context flags those seasons. Conduct still labels the mode: "guest week planning," "exam week compression." Labels tie adaptation to inspectable context rather than mystery. Adaptation follows declared policies: promotion rules, posture limits, substitution gates. Households tolerate change when diffs and sources explain what changed. They churn when behaviour shifts without visible cause. Versioning cognitive unit packages with semantic versioning and substitution rules keeps improvements from feeling like personality drift. Unpredictable adaptation is often hidden prompt edits. Predictable adaptation is sealed package promotion with suite evidence.

What to do next

The most visible strategy switch in a product is the starting point. Documenting the guard, user signal, and fallback path, and shipping the signal before adding new compressed paths, makes adaptation governed rather than capricious. Reviewing whether any behavioural variance remains explained only by model sampling and converting material variance to named modes users can recognize closes the gap between engineering intent and user experience.

See applicability conditions, memory discipline in conduct, and confidence theatre.

Read next: Memory discipline in conduct.