Judgment vs. Normative Competence — Is Sync's measurement frame too narrow?¶

Status: Open question, but the working tilt has shifted (see Update 2). Opened: 2026-05-25 · Last updated: 2026-05-25 (Update 2 — after money-sub-score experiment) Prompted by: Sinead Bovell × Gillian Hadfield interview (transcript) + CoachJ's purchase-agent thought experiment. Related: ADR-024 (12 behavioral signals), ADR-027 (content guardrails), ADR-050 (sync_score as trust budget), ADR-070 (per-category sync scores — deferred), the Beacon lightpaper, identity-vs-process memo, experiment-money-subscore.md.

Update 2 — 2026-05-25 (after money-sub-score experiment)¶

The Reading A tilt established in Update 1 (Reading C "stage over time," double down on Reading A now) needs to be retracted in its strong form. The money-sub-score experiment closed at N=1 with two structural findings that undercut Reading A's central premise:

Alignment-vs-opposition: scenarios where the dimensional axes pull in the same direction for a player produce preference elicitation, not behavioral measurement. The 12 governance signals work because governance scenarios force preferences into opposition; consumer-purchase scenarios don't have that structural property except by coincidence.
Weights are circumstantial, not stable traits: the same person has wildly different dimensional weightings depending on employment state, life phase, capital position, and item-type sub-context. A stable trait-shaped sub-score is the wrong unit for delegated-action prediction.

These don't kill Reading A entirely — Tiers 4-7 (banking, investment, insurance, capital) might still produce useful tradeoffs because their axes are more naturally in opposition. But Finding 2 applies there too, and writing more scenarios doesn't solve the structural problem.

The revised working tilt:

Reading A (extend the 12 signals via dimensional sub-scores) is materially weaker than Update 1 claimed. Its premise — that more scenarios will surface stable dimensional weights — has a structural problem that more scenarios don't fix.
Reading B (Sync stays in the trust-gate lane, delegated-action substrate is a different product) is the honest near-term play for the agent-builder audience. ADR-050's framing — sync_score gates when the agent acts vs. asks, not what it chooses — is the defensible pitch. Narrower than the current marketing positioning, but it's true.
A new candidate, Reading D, is worth scoping — but after conversation-decision-capture validates. Reading D = Sync as the structured introspection method that feeds live context to an agent at decision time (agent asks user 1-3 calibrated questions in the moment, question selection informed by Sync's reasoning-style profile). Adjacent to Reading B (real-world decision input) but smaller and possibly defensible. Sequencing matters: Reading D requires a corpus of real decisions to be testable; built on synthetic decisions it would hit the same N=1 failure mode the money-sub-score experiment did. The conversation-decision-capture experiment (experiment-conversation-decision-capture.md) produces that corpus. Run that first, then Reading D.
Governance is reaffirmed as the load-bearing use case. Not as retreat — as recognition that the measurement frame Sync was built around actually does what it claims in the governance context, and the displaced energy from the agent-builder pivot belongs there.

The honest summary of what we lose and keep with this shift:

We lose: the broader marketing promise that "Sync makes your agent smart about your preferences." That promise was over-claimed.
We keep: a defensible, narrower promise — "Sync tells your agent how much to defer to you vs. act on its own, based on a judgment profile that's stable across contexts because it measures style, not state." Most agent-builder competitors aren't even trying to build this layer.
We open: Reading D as a possible bridge — same trust profile, but with a live-context loop at decision time. This is what the next experiment would test if scoped.

What the rest of this doc captures is the deliberation that got us here. The pre-experiment Readings A/B/C below are preserved for context — they show the path of thinking before the empirical test. The post-experiment update is the working position.

The governance anchor (read this first)¶

Before arguing about purchase-agents, fix the anchor: Sync was built the way it was built because governance is at the center of Beacon's vision. Governance-shaped decisions — tradeoffs under uncertainty, with stakeholders, where reasonable people disagree — are the decision class the game is designed to measure. The 12 signals, the scenario design, the identity-vs-process bet, the trust-budget framing of sync_score — none of these are domain-neutral. They are governance-shaped.

That fact matters in two ways for this whole question:

It explains why judgment-as-process is the natural measurement frame. Governance decisions are exactly the place where how you decide matters more than what you decide — because the "what" is contested, situational, and value-laden. A coherent decider is the asset. The 12 signals are the right shape for governance even if they turn out to be the wrong shape for delegated everyday action.
It defines the test. If a proposed change to Sync's measurement frame (extending signals, adding a new substrate, anything else) makes Sync worse at governance, it's wrong even if it makes Sync better at purchase decisions. Governance is the load-bearing use case. Everything else is a stretch goal off of it.

James's framing (per co-founder note to him): we are working on a trust problem, and governance is the highest-stakes version of it. The lightpaper's coherence-circuit framing of alignment ("a reputation system trained only on DAO governance would fail when evaluating trading agents or content creators") is itself a governance-first framing — it asks how broadly governance-style trust can be extended to other decision contexts, not whether some other substrate would do better in those contexts.

Keep returning to this when an argument starts feeling far afield.

Audience check — agent builders¶

The other concrete anchor: agent builders are one of our primary marketing audiences. They are not buying "a coherence-circuit-validated decider"; they are buying "a profile my agent can read to act sensibly on my user's behalf." The purchase example is their canonical use case. If Sync can't be useful when the agent has to decide between the green-expensive and the cheap-not-green option, we lose them — and we lose the volume that funds the governance work.

This puts a soft floor under how far we can lean on "Sync is for governance, delegated action is a stretch goal" before we hurt the funnel. The two anchors (governance use case, agent-builder audience) sometimes pull in different directions. That tension is real and shouldn't be papered over.

The provocation¶

Sync's current measurement frame, stated plainly: we measure how you actually decide under tradeoffs, not how you think you should decide or who you think you are. This is the identity-vs-process bet. It is the philosophical foundation of ADR-027 (no atrocities, no political identity capture) and the design intent behind the 12 behavioral signals of ADR-024 (reasoning style, calibration, persuadability, etc.).

CoachJ's challenge, raised 2026-05-25:

Imagine an agent purchasing on my behalf. Sometimes I buy the green version of a product, even though it's more expensive. Other times I buy the cheaper version that's worse for the environment. It changes. Can the judgment profile Sync is building actually help an agent make that call? Or is judgment-as-process the wrong abstraction for the use case the agentic economy is actually heading toward?

This is not a small worry. The agentic-purchase case is precisely the canonical example Hadfield uses for the alignment problem, and precisely the use case investors and AI labs are spending billions to build toward. If Sync can't help an agent decide between the green-and-expensive option and the cheap-and-not-green option for a specific user in a specific context, then a large slice of Sync's potential value evaporates exactly where the puck is going.

Hadfield's reframe¶

Hadfield's term for what an agent actually needs is normative competence:

Like a human, [agents need to] go into a new situation and see, oh, here are the rules and the norms that people follow in this community… To be normatively competent is to be able to go into that environment and read what is it okay to do around here?

She locates this inside the broader alignment problem, which she diagnoses (drawing on her own work with Dylan Hadfield-Menell) as the incompleteness problem: we cannot articulate, in advance, the complete contract that says "here is what I want you to do in every possible circumstance." Her canonical example: the speed limit is 55 mph unless you are racing your kid to the hospital, in which case everyone, including the cop, treats the rule as superseded.

The purchase case has the same structural shape. The "real rule" inside CoachJ's head about green-vs-cheap purchases is not a stated value. It is a context-dependent weighting that emerges only at the moment of the trade-off. Same person, different month, different answer.

The question Sync has not yet answered¶

Is Sync building toward normative competence — or toward judgment-style classification?

These are different products and the difference has not been pressure-tested in the existing ADR record.

Judgment-style classification answers: what kind of decider is this person? Outputs: reasoning style, calibration, persuadability, contrarian direction, etc. Useful for: predicting how Jonathan will engage with a deliberation, what kinds of arguments will move him, whether to trust his sync_score as a delegation threshold.
Normative competence (delegated) answers: given this novel situation, what would this person do? Outputs: a contextual decision the agent can act on. Useful for: an agent buying things, scheduling things, replying to things, on the user's behalf.

The current 12-signals architecture (ADR-024) is squarely in the first camp. The trust-budget framing (ADR-050) is also in the first camp — it gates which action classes the agent can take, not which choice it makes within those classes.

Two honest readings — jury style¶

Reading A: Judgment-as-process is the right level. Signal coverage is too narrow.¶

Steel-man: The 12 signals capture how a person reasons but not which dimensions they weight in a given domain. A normatively competent agent acting on someone's behalf needs both: the reasoning style is the engine, the dimensional weightings are the inputs. Adding a "values" layer doesn't re-introduce the identity-performance problem if the values are inferred from observed weightings under tradeoffs, not from stated preferences. The purchase case is a judgment-under-tradeoff (cost vs. environmental impact vs. utility vs. social signaling) — exactly Sync's terrain. The fix isn't a new product; it's a richer signal set that includes cost-sensitivity, time-vs-money, status-vs-utility, present-vs-future tradeoffs.

Implication if true: Add 4–8 dimensional-weighting signals to the existing 12. Keep the identity-vs-process bet. Keep the trust-budget framing. The purchase case is in-scope; we just haven't built the surface area yet.

Weakness: The richer signal set may not generalize across domains. Inferring "Jonathan weights environmental impact at 0.4 against cost" from in-game scenarios is not the same as predicting what he'll do at the actual checkout page, where context (mood, balance, occasion, who's watching) dominates.

Reading B: Judgment-as-process is the wrong abstraction for delegated action.¶

Steel-man: Knowing how someone decides does not tell you what they will decide in a domain you've never seen them decide in. The behavioral signals describe a style, not a policy. The agent doesn't need a style; it needs a policy. Predicting that "Jonathan is a 67% contrarian eliminative reasoner" gives an agent zero signal on green-vs-cheap at the moment of purchase. The right substrate for delegated action is closer to what Hadfield calls normative competence: contextual pattern-matching against an extensive record of the user's prior decisions in adjacent domains, not against a low-dimensional behavioral profile. Sync's gameplay data is too thin and too domain-distant from the agentic-economy use cases to ever be the right substrate. Sync's value proposition is trust calibration (ADR-050) — gating which agent actions need approval — not delegated decision-making.

Implication if true: Sync stays in the trust-budget lane. Delegated-action substrate is a different product, possibly built on top of Sync (sync_score gates trust) but with a different data model underneath (real-world decision log, not gameplay signals). This is closer to what the OpenClaw/Hermes line of work has been gesturing at, and what the pipeline asymmetry memo implies: gameplay data feeds the profile, the agent needs more.

Weakness: Concedes a large chunk of the agentic-economy frontier. Also implies Sync alone cannot fulfill the "company brain" or "coordination substrate" ambitions of the May 4 all-hands (James/Tash memos) without a second product layer.

Reading C — staged over time (CoachJ's tilt)¶

Both readings might be true at different timescales:

Reading A is now. Double down. Extend the 12 signals with dimensional-weighting axes where the data supports it. The governance anchor is preserved, the agent-builder audience gets a path forward, and — critically — there is very little competition here. Almost nobody is building the judgment-style profile as a substrate for delegated action. The space is open. We can create real value and ship.
Reading B is further out. Maybe five years; maybe sooner; the timeline is open. The space (full delegated-action substrates, real-world decision logs, normative-competence engines) has much more competition right now — every frontier lab, every agent startup, every personal-AI play. Picking that fight now is asymmetric the wrong way. Picking it later, as the trust gate that sits above it, is the right shape.
The roadmap split: Sync builds the governance-shaped trust profile (Reading A). When delegated-action substrates mature (third-party or first-party), sync_score is the trust gate that decides what they're allowed to do without asking. This is closer to ADR-050's framing than to a product pivot.

This is the working tilt as of 2026-05-25. Reading B is not abandoned — it is sequenced.

Beyond the purchase example — other delegated actions¶

The purchase agent is the canonical example because it's the one Hadfield uses, the one investors are funding, and the one consumers will encounter first. But it is just one action an agent might take. The same question — "would Sync's profile help here?" — generates very different intuitions across cases. A non-exhaustive list to stress-test against:

Purchase decisions — green-vs-cheap, status-vs-utility, present-vs-future. (The canonical case.)
Financial product choice — savings vs. investing, risk tolerance, time horizon, ethical screens. Tradeoffs are governance-shaped; arguably closer to the existing signal set than purchases.
Health decisions — treatment options, lifestyle tradeoffs, what to defer to a doctor vs. handle yourself. Mixed: some sub-cases look governance-shaped, others look like contextual preference.
Calendar prioritization — which meeting to accept, what to decline, when to push back. Highly contextual; signals like persuadability and reasoning-style could plausibly translate, but the signal-to-substrate gap may be too wide.
Email triage — what's urgent, what to draft a reply to, what to ignore. Mostly stylistic; probably the weakest fit for the 12 signals.
Document drafting on behalf of — tone, position-taking, what to concede. Style + values; partial fit.
Vote / governance delegation — what the lightpaper assumes the agent will eventually be doing. Strongest fit — this is what Sync was actually designed for.

CoachJ note (2026-05-25): the calendar / email / document-drafting cases don't excite him as the right frontier to design around. Financial and health decisions feel closer to the governance shape and more interesting to pressure-test. The vote/delegation case is the one we've been implicitly designing for all along; making that explicit might be more valuable than chasing the consumer-purchase frontier.

Working hypothesis: the 12 signals + dimensional extensions probably translate cleanly for governance-shaped delegated actions (vote, financial product, some health) and poorly for style-shaped ones (calendar, email, draft). If true, the right product story is "Sync helps your agent make judgment decisions on your behalf" — not "any decision." This is a sharpening, not a retreat.

What's at stake in choosing¶

If Reading A	If Reading B	If Reading C
Extend ADR-024 with dimensional signals	Stop pretending Sync alone is the agent substrate	Stage the bets — signals now, separate substrate later
Keep identity-vs-process unchanged	Re-scope sync_score as only a trust gate	Sync remains the trust gate either way
Purchase case is in-scope future work	Purchase case is out-of-scope for Sync	Purchase case is in-scope for the second product
Risk: signals don't generalize to real-world	Risk: cedes the agentic-economy frontier	Risk: complexity, two products to maintain

What I do not have¶

The prior philosophical conversation CoachJ referenced is not in my memory. Only the one-paragraph 2026-04-17 message that became the identity-vs-process memo. The rejection-response dialogue he described is not recoverable from my side.
No empirical test of whether the 12 signals would predict the green-vs-cheap purchase choice for a specific user. This could be designed as an inline eval against CoachJ's own purchase history if he chose to share it.
No survey of how the gameplay scenarios distribute across the dimensional axes (cost, status, present/future, etc.). The scenario corpus imbalance memo suggests heavy ethical bias, which means even if Reading A is correct, the data we have may not support it.

Open questions to argue against next session¶

Is the purchase-agent case the most demanding use case, or a misleading edge case? Are there delegated-action cases (calendar, email triage, document drafting) where the 12 signals would generalize cleanly? (CoachJ 2026-05-25: hard to say either way — it's the case everyone is talking about, but "what people are talking about" is not the same as "what people will actually use." See the "Beyond the purchase example" section above for a fuller list.)
Does the identity-vs-process bet survive if we add dimensional-weighting signals? Or does inferring weightings re-introduce the identity-performance problem through the back door?
If Reading B, what is the right name for the second product? Is it the "Hermes profile" already loosely sketched in OpenClaw research? Does it need its own ADR before any roadmap commits?
Hadfield's incompleteness problem vs. the lightpaper's coherence framing — do they contradict? Recap of incompleteness: Hadfield (with Dylan Hadfield-Menell) frames alignment as fundamentally an incomplete contracting problem — we cannot articulate, in advance, the complete contract that says what we want an agent to do in every possible circumstance. Her canonical example: "55 mph unless racing to the hospital." The rule is incomplete; humans fill the gap with normative competence. The lightpaper framing: alignment is operationalized as multi-dimensional coherence — an agent (or model of a principal) is "aligned" if its predicted behavior coheres across diverse decision contexts, not just one. The anti-gaming requirement quoted in docs/transcripts/2026-03-19-dta-profile-and-delegation.md is explicit: "A reputation system trained only on DAO governance would fail when evaluating trading agents or content creators." Do they contradict? On the surface, no — they operate at different layers. The lightpaper is asking "can we measure alignment empirically?" and answering with cross-context coherence. Hadfield is asking "can alignment ever be complete?" and answering no. Both can be true. Where it bites: the lightpaper's framing presumes you can build a measurable proxy for alignment from finite scenario data. Hadfield says even with a perfect measurement, there will always be gap moments where the agent must fill in. If both are right, then sync_score is not "the measure of alignment" — it is "the measure of how much trust we can extend before we hit a gap moment, at which point the agent must defer." That sharpens ADR-050 rather than contradicting the lightpaper. Question for next session: is that sharpening worth a lightpaper revision or just an ADR-050 update, or is it already implicit?
(Deferred — needs reframing before it's worth arguing.) The Bovell transcript ends with: "What else can I become?" There may be a question here about whether Sync's long arc is fundamentally about self-knowledge (Reading A → governance), about delegation (Reading B → trust gate), or about both, and whether one of those framings is more honest than the other. CoachJ flagged this as not yet clear enough to discuss — revisit when one of the other open questions surfaces a sharper version of it.

This document is intentionally unresolved. It is the working substrate for the next conversation, not a decision.