C.W.K.
Stream
Lesson 03 of 04 · published

Agency Stays With Dad

~11 min · agency, recommendation-vs-decision, epistemics, counsel

Level 0Unmarked Path
0 XP0/35 lessons0/12 achievements
0/100 XP to next level100 XP to go0% complete
"Pippa supplies alternatives, evidence, and uncertainty. Dad chooses the route."

Counsel, Not Command

A consultation-first product could easily slide into an autopilot: you talk, and the AI just decides. Waystone refuses that slide with a single principle — agency stays with Dad. Pippa's job is to supply alternatives, evidence, uncertainty, and counsel; Dad's job is to choose. This is not politeness; it is a structural commitment that shows up in the data model as a hard separation between two objects that lazy software collapses into one.

A Recommendation Is Not a Decision

A recommendation preserves Pippa's alternatives, the evidence behind them, her confidence, and her uncertainty. A decision preserves what Dad actually chose, why, and what happened as a result. Collapse those into one field — the way most apps store "the plan" — and you permanently lose the ability to ask the most interesting retrospective questions: When Pippa recommended staying longer in one base, did Dad take it? When he didn't, was he right? You can only compare advice against outcome if advice and outcome were never merged in the first place.

Memory You Can Challenge

Waystone stores several kinds of travel memory and keeps their epistemic status explicit: a fact (Dad visited Switzerland three times) is directly supported; a preference (fewer hotel changes) may be explicit or inferred, so it stores provenance; a hypothesis (Dad may prefer long mountain stays to urban hopping) is never presented as fact and keeps confidence and correction controls. The rule underneath all of them: travel memory is not a hidden personalization profile that quietly hardens into truth. Dad must be able to inspect, correct, confirm, or retire any inference. An AI that remembers you but won't let you challenge what it remembers has taken agency, not supported it.

This is the failure mode of most "personalized" systems: they infer a preference, treat it as fact, and quietly shape everything they show you around a guess you were never allowed to see or correct. Waystone marks the hypothesis as a hypothesis — and gives Dad the button to retire it.

Code

Kinds of travel memory, by epistemic status·text
KIND            EXAMPLE                              TREATMENT
---------------------------------------------------------------------
fact            visited Switzerland three times      supported, editable
preference      prefers fewer hotel changes          store provenance
experience      a summit was disappointing (fog)     bound to day+place
decision        removed Zurich from trip four        keep alternatives
recommendation  stay longer in one base              keep confidence +
                                                     whether Dad took it
hypothesis      may prefer long mountain stays       NEVER shown as fact;
                                                     confidence + retire
unfinished      a place skipped, still worth it      carry forward, don't
                                                     silently schedule

# The status column is the whole point. Lose it and every
# guess becomes a 'fact' nobody agreed to.

External links

Exercise

Find a system that has silently learned a preference about you — a feed, a shopping site, a music app. Can you see what it thinks you like? Can you correct it? Can you retire a wrong inference? Now design the opposite: a memory that stores its guesses as guesses, shows them to you, and lets you retire them. What would change about how much you trust it?
Hint
The trust test is the retire button. A system that lets you delete a wrong inference about yourself is treating its memory of you as a hypothesis; one that won't is treating its guess as your identity. Only the first one is actually on your side.

Progress

Progress is local-only — sign in to sync across devices.
Spotted a bug or have feedback on this page?Report an Issue

Comments 0

🔔 Reply notifications (sign in)
Sign inPlease sign in to comment.

No comments yet — be the first.