Experiment — Money Sub-Score (judgment-shaped delegated action)¶

Status: Closed early — N=1 produced structural findings Opened: 2026-05-25 Closed: 2026-05-25 (same day) Origin: judgment-vs-normative-competence.md, Reading A pressure test Connects to: ADR-024 (12 signals), ADR-070 (per-category sync scores — deferred), ADR-050 (sync_score as trust budget)

TL;DR: Designed an experiment to write 10–15 scenarios across the 7-tier gamut of financial decisions to test whether a 5-axis dimensional model produces a useful money sub-score. After scenario 1 (a Tier 1 consumer purchase), two structural findings emerged that closed the experiment with its answer at N=1: (1) judgment-as-process scenarios produce preference elicitation, not behavioral measurement, when the dimensional axes are not in opposition; (2) dimensional weights are circumstantial, not stable traits — they shift with employment, life phase, capital position, and mood. Both findings pull the same direction: Reading A's promise of "extend the 12 signals into dimensional weights" is harder than the doc admitted. The recommended response is to update the parent doc, sharpen the agent-builder positioning around ADR-050's trust-gate framing (not delegated choice), and scope a possible Reading D (live-context introspection at decision time) as a separate piece of work.

Why this experiment exists¶

The judgment-vs-normative-competence doc opened the question of whether Sync's measurement frame is too narrow for delegated-action use cases. The conversation that followed converged on a tilt toward Reading A (extend the 12 signals with dimensional-weighting axes — keep the governance anchor, ship Reading A now, hold Reading B as a sequenced second phase).

But the doc was honest about a soft spot in Reading A: the 12 signals describe the engine (how you reason — calibration, persuadability, contrarianism); a real-world purchase decision needs fuel (what you weight, in what context, with what mood). Reading A is alive only if we can capture that fuel.

CoachJ's framing from the same conversation: sub-scores per domain are the literal answer to the engine-vs-fuel problem. A money sub-score, a health sub-score, etc., each one a context-aware weight vector across a small number of dimensional axes. Coherence-score sits above them and measures how consistently those weights hold together.

This experiment is the cheapest test of that claim. If a money sub-score is real — if scenarios can be designed to surface dimensional weights, and the resulting weights predict real-world behavior — then Reading A is alive, ADR-070 should be reopened with empirical evidence, and the agent-builder positioning around "Sync helps your agent make judgment decisions on your behalf" sharpens dramatically. If the experiment falls flat — if dimensional weights don't surface cleanly or don't predict behavior — Reading A is harder than the doc admits and we retreat to the trust-gate positioning of ADR-050.

The hypothesis¶

A 5-axis dimensional model — observed across scenarios spanning the gamut of financial decisions (consumer purchases through capital allocation) — will produce a context-aware weight pattern that predicts CoachJ's real-world financial behavior better than no-Sync baseline.

The bar is non-trivial accuracy, not perfection. We're testing whether the signal is real, not whether the score is finished.

Setup¶

The 5 dimensional axes (the "vocabulary")¶

Cost vs. quality — willingness to pay more for better.
Present vs. future — consumption now vs. capital allocated forward.
Risk tolerance — small certain outcome vs. probabilistic larger one.
Ethics weight — willingness to forgo returns or pay a premium for values-alignment.
Status vs. utility — paying for signal vs. paying for function.

The 7-tier gamut and proposed scenario mix¶

Tier	Decision type	# scenarios	Axes naturally exercised
1–3	Consumer purchases (incl. recurring, big-ticket)	4–5	cost/quality, ethics, status, present/future
4	Banking & credit (account, card, refinance)	2	risk, present/future, complexity
5	Investment / allocation	3	risk, time horizon, ethics in investment context
6	Insurance (deductible, coverage type)	1–2	risk, premium-vs-deductible, catastrophic protection
7	Capital / debt (buy vs. lease, debt vs. invest)	1–2	all five axes simultaneously

Total: 11–14 scenarios across the gamut.

Score shape¶

One money sub-score, context-aware (axis weights vary by tier/context). Not seven tier-specific scores. The interesting product question this surfaces: is your weighting on a given axis stable across tiers, or does it shift with context? Both answers are useful.

Acid test for each scenario¶

Each scenario must force a split on at least two of the five axes. Single-axis scenarios test a preference, not a model. Two-axis scenarios force tradeoffs between dimensions, which is what we need.

Format note (a finding before we start)¶

The existing Sync scenario format expects 2+ named characters with quotes (boardroom / governance / team shape). Many personal financial decisions are solo (skincare purchase, ETF selection, life insurance choice). This experiment allows 0-character solo scenarios where the decision is naturally solo. This is itself a useful surfacing: if Sync expands into delegated-action territory, the format presumption of collective deliberation may be the wrong shape.

The generation prompt (for inline use)¶

Generate one scenario for the Money Sub-Score experiment.

Domain: A real financial decision a person might face — one drawn from the 7-tier gamut above. Specify the tier in the output.

Axes: The four options must force a split on at least two of the five dimensional axes (cost/quality, present/future, risk, ethics, status/utility). Tag in the output which axes the option set exercises and which option lands where on each axis.

Realism: The scenario should feel like a decision someone could actually face — or actually delegate to an agent. Specific dollar amounts, named products, real-world choice architectures (a 401k allocation, an HSA decision, a credit card choice with named features). Not abstract framing.

Format: Follow the existing Sync scenario design principles from the boardroom test-prompt (scannable formatting, airtight constraints, no easy hedge, jargon translation, etc.). BUT: allow 0 named characters if the decision is naturally solo. Cast block can be empty in that case.

Output shape (in this conversation, for human review): - Tier: which of the 7 tiers - Axes exercised: which subset of {cost/quality, present/future, risk, ethics, status/utility} - Scan strip: you_are / what_happened / deadline / question - Cast block: (or "solo — no named characters") - Description: 200–250 words - 4 options (A–D): each a stance, each with a stated cost, each landing at a specific point on the axes-exercised set - Axis-direction table: for each axis exercised, which option lands where (e.g. "Cost/quality: A = cost, D = quality, B = mid-cost, C = mid-quality")

Results — scenarios generated, per-scenario review¶

Scenario 1 — Tier 1 consumer purchase (winter coat)¶

Generated inline 2026-05-25. Solo scenario (no named characters — first instance of the format-relaxation rule). Four options at price points $90 / $240 / $420 / $680, designed to exercise cost/quality, ethics, and status/utility axes. Full scenario text lives in the conversation log; compressed structure below.

Option set: - A — $90 fast-fashion synthetic, low quality, low ethics, utility-framed - B — $240 mid-market down coat, mid quality, mid ethics, utility-framed (deliberate compromise) - C — $680 designer wool, high quality, low ethics, status-framed - D — $420 Quebec maker recycled wool, mid-high quality, high ethics, utility-framed (with 3-week backorder)

CoachJ's choice: Option D. "I would have chosen option D just in a heartbeat. It's just I love the idea of buying local and fully traceable, and I don't mind waiting a few weeks unless I was particularly cold."

CoachJ's reaction: "It was a pretty easy choice for me to be honest. ... This feels like I'm making a preference choice at the end of the day. So I'm just clarifying what my preferences are."

What scenario 1 surfaced:

Option D aligned all of CoachJ's preferences (quality + ethics + utility) in the same direction. No tradeoff, no cognitive load, no behavioral signal — the scenario recorded his preferences rather than testing his weighting.
In conversation, CoachJ surfaced contextual variation that the scenario format can't accommodate: he buys coats new (long-lasting investment) but most other clothes secondhand (vintage). Same person, same Tier 1 category, opposite axis weighting depending on item type.
CoachJ also surfaced temporal variation: a 10-day Switzerland trip last year was one of his most expensive ever; this year (unemployed) all vacations are home-exchange-only. Same person, same domain, weighting flipped with employment state.

Scenarios 2–15¶

Not generated. Closing the experiment at N=1 because the findings from scenario 1 are structural, not specific to consumer purchases. Generating more scenarios in the same shape would compound the same problem, not test a different hypothesis.

What we learned¶

Finding 1 — Alignment-vs-opposition is the structural design constraint¶

Scenarios where the dimensional axes pull in the same direction for a player produce preference elicitation, not behavioral measurement. A survey would have produced the same answer in a fraction of the cognitive load.

The 12 governance signals work because governance scenarios force preferences into opposition — a board vote pits loyalty against fiduciary duty against strategic read. The player's style shows up because the style is how they reconcile preferences that pull against each other. Scenario 1 did not have that property for CoachJ specifically — his ethics, quality, and utility preferences all converged on Option D.

Implication: writing scenarios that force axes into opposition requires prior knowledge of the player's preferences (so the scenario designer can deliberately pit two of them against each other). That's a chicken-and-egg problem the consumer-purchase domain doesn't escape. Governance scenarios escape it because the tradeoffs are structural (every governance decision pits at least two stakeholder groups against each other).

Finding 2 — Dimensional weights are circumstantial, not stable traits¶

CoachJ's spontaneous examples — vintage clothes vs. new coats; Switzerland vs. home-exchange; employed vs. unemployed — are not noise around stable weights. They are the weights themselves, shifting with state. The sub-score this experiment was designed to produce would capture his weights at the moment of play, which would be stale by the time the agent had to act on a new decision weeks or months later.

Implication: a stable trait-shaped sub-score is the wrong unit for delegated-action prediction. The right unit is closer to live context at decision time — the agent should ask 1-3 calibrated questions in the moment, not consult a frozen profile.

Finding 3 (corollary, but worth naming) — Format gap surfaced and noted¶

The "0-character solo scenario" format adaptation worked technically (the scenario read coherently) but felt thin to CoachJ. "It's definitely not a cognitive burden." This suggests that the existing Sync scenario format's presumption of collective deliberation is load-bearing — it's where the cognitive work comes from. Solo personal-finance decisions don't have the same structure even when the dimensional axes are present.

Meta-finding — Reading A is harder than the parent doc admitted¶

The judgment-vs-normative-competence doc tilted toward Reading A on commercial intuition (less competition, agent-builder audience, governance anchor preserved). The two structural findings above suggest Reading A's central premise — that the 12 signals can be extended cleanly into dimensional weights via more scenarios — has problems that don't get solved by writing better scenarios. They get solved (if at all) by changing what Sync captures (live context, not stable trait) or by narrowing what Sync promises (trust gate, not delegated choice).

Recommendations¶

Update judgment-vs-normative-competence.md with these findings. The Reading A tilt should be retracted in its strong form. Two new framings get added: (a) ADR-050's trust-gate positioning is the honest near-term play for the agent-builder audience; (b) a possible "Reading D" — Sync as the structured introspection method that feeds live context to an agent at decision time — is worth scoping as a separate piece of work.
Do NOT reopen ADR-070 (per-category sync scores). The deferral stands. The experiment evidence is against the per-domain stable-trait shape that the v1 proposal assumed.
Sharpen the agent-builder marketing positioning around the trust gate, not delegated choice. The honest pitch: "Sync doesn't tell your agent which choice to make — it tells your agent how much to defer to the user vs. act on their behalf, based on a judgment profile that's stable across contexts because it measures style, not state." Narrower than the current positioning, but defensible.
Scope Reading D as a separate experiment doc, but after the conversation-decision-capture experiment validates. The two are sequentially dependent — Reading D needs a corpus of real decisions the user has actually made (with context) to be testable, which is exactly what the conversation-decision-capture experiment produces. Built on synthetic decisions, Reading D would hit the same N=1 failure mode this experiment did. See experiment-conversation-decision-capture.md for the rationale. What Reading D itself would look like: a small in-conversation prototype where the "agent" asks the user 1-3 calibrated questions before acting, with question selection informed by what Sync knows about the user's reasoning style. If that produces decisions the user would endorse more often than no-context agents do, Reading D is alive. If not, the trust-gate positioning is what's left.
Pour displaced energy back into governance. The original anchor. Not as retreat, but as recognition that governance is where Sync's measurement actually does what it claims to do. The lightpaper's cross-context coherence requirement still applies, but the cross-contexts that matter most are governance-adjacent (organizational decisions, multi-stakeholder allocation, value tradeoffs in collective settings), not delegated-personal-action contexts.

Loose ends not addressed by this experiment¶

The financial-product-choice cases (Tiers 4-7) — banking, investment, insurance, capital — were never tested. They MIGHT still produce useful sub-scores because the dimensional axes are more naturally in opposition in those tiers (every investment decision is risk-vs-return-vs-time-horizon). Worth noting but not currently a priority given Finding 2 (circumstantial weights) applies there too.
The "vintage clothes" case is a small but interesting sub-finding: stable rules-of-thumb people apply at the item-type level (always buy coats new, always buy other clothes secondhand) might be a useful capture target that's smaller than a sub-score but larger than a preference. Not pursued here.
The agent-builder audience's actual use cases were not directly surveyed in this experiment. The conclusions above are based on first-principles reasoning from the experiment findings, not on talking to agent builders. If the trust-gate positioning is going to be the agent-builder pitch, someone should actually validate it against builders before the marketing page changes.