ADR-074: Pulse as a multi-domain judgment layer¶

Status: Accepted Date: 2026-05-31 Context: Pulse started as a governance-focused decision-revelation game. The light-paper positioning has shifted to "the judgment layer for any AI chatbot" — a multi-domain claim. The question this ADR resolves: does the current architecture support that positioning, or does multi-domain require a parallel/rebuilt product? The scenario dimension cube (docs/research/foundations/scenario-dimension-cube.md) was built during May 2026 as a domain-agnostic structural model and survived 12 swings, 29 corpus decompositions, governance-generator comparison, smoke-tested live integration (rolled back at v1.8.1, re-introduced at v1.9.0), and the design of a cross-domain consistency experiment that drafted 10 scenarios (5 vector α + 5 vector γ) across career / health / relational / financial / governance — proving the same cube vector lands in any domain. Earlier framing ("governance is a different breed of scenario") has been revised (see the project-governance-not-structurally-distinct working note): the corpus is governance-skewed because it has been more sampled there, not because governance is structurally distinct.

At a glance

What it decides: Pulse is positioned as a multi-domain judgment layer for any AI, because the existing architecture already supports non-governance domains — what's missing is content, not a rebuild.

The architecture is domain-agnostic by construction — the scenario cube, 14 signals, 8 drivers, 12 value tensions, sync mechanic, and agent layer all carry across domains
Remaining work is corpus + small taxonomy additions (~4–6 weeks): 50–100 non-governance scenarios, 1–2 categories, 2–3 triggers, UI copy refresh, signal baseline recheck
Rejected building a parallel product — duplication would share all the same machinery, so it isn't actually parallel
Watch for signal calibration drift on personal-stakes scenarios that doesn't recover with retuning; that would weaken the "no rebuild" claim

flowchart TB
    subgraph ARCH["Domain-agnostic architecture (already built)"]
        direction LR
        C["Scenario cube"]
        S["14 signals"]
        D["8 drivers"]
        V["12 value tensions"]
        M["Sync mechanic"]
        A["Agent layer"]
    end
    ARCH --> WORK["Remaining work: content, not rebuild"]
    WORK --> W1["50–100 non-governance scenarios"]
    WORK --> W2["1–2 categories, 2–3 triggers"]
    WORK --> W3["UI copy refresh"]
    WORK --> W4["Signal baseline recheck"]

The shared architecture already spans domains; the catch-up work is corpus and taxonomy, not a parallel product.

Decision¶

Position Pulse externally and internally as a multi-domain judgment layer. The architecture supports this; remaining work is corpus expansion + small taxonomy additions, not a product rebuild.

The light-paper claim that Pulse measures judgment across domains is structurally backed by the cube, the cross-domain signal architecture (ADR-024 + the two signals surfaced in experiment-new-domain-signals.md), and the domain-agnostic agent consumption layer (ADR-071). This is not aspirational positioning — it is a recognition of what the architecture already supports.

Rationale¶

What is already cross-domain by construction:

The scenario dimension cube (12 structural axes + 1 content axis + 2 measured outcomes) — domain-agnostic by definition. The cross-domain experiment design demonstrates that any cube vector can be realized in any of career / health / relational / financial / governance.
Behavioral signals (14 total) — the 12 in ADR-024 plus pragmatic-routing and self-aware-aspiration-gap from experiment-new-domain-signals.md. The experiment empirically validated that these signals fire consistently across domains for the same player.
Drivers (8) in scenario-generator.ts — reasoning styles, not domain shapes. "Team harmony" maps to "interpersonal harmony" in non-team contexts without modification.
Value tensions (12) — universally applicable (risk-tolerance, people-vs-outcomes, short-vs-long-term, etc.).
Variant D scenario format — works for any domain.
Sync score mechanic — predict → reveal → match → score is domain-agnostic.
Agent / Hermes consumption layer (ADR-071) — reads the profile, has no domain dependency.

What needs work for credible multi-domain serving:

Corpus expansion — write 50–100 non-governance scenarios across personal-choice, relational, health, financial, career-personal-stakes. The multiplayer corpus build already in progress doubles as multi-domain seeding (multiplayer scenarios tend toward shared-life-decision shapes).
Category taxonomy expansion — current categories (governance / resource-allocation / team-dynamics / values-culture) have varying transfer: values-culture and resource-allocation generalize cleanly; team-dynamics is narrow (work-only); governance is borderline. Add 1–2 new categories: personal-choice, possibly relational-decision.
Trigger taxonomy expansion — 6 of 10 triggers are universal; 4 are work-shaped (team_conflict, competitive_pressure, innovation_challenge, authority_threat). Add 2–3 non-work triggers: personal-crossroads, intimate-disclosure, health-decision-point or similar.
UI copy refresh — review onboarding / docs / settings for "team," "company," "governance," "board" language that locks the work framing. 1–2 day lift.
Signal baseline recalibration — if any of the 14 signals were tuned on governance-shaped scenarios (eliminative reasoning, overconfidence, persuadability), they may show drift on personal-stakes scenarios. Detectable post-corpus-expansion; fixable by retuning baselines.

Total estimated effort: 4–6 weeks to credibly serve non-governance domains end-to-end. Compared to a product rebuild (which would be 6–12 months and require duplicating the cube, signals, scoring, agent layer for a parallel product), this is a small fraction of the cost.

Alternatives Considered¶

Build a parallel product for non-governance domains. Rejected. The architectural duplication would be expensive and the parallel product would necessarily have to share the cube, signals, drivers, value tensions, sync mechanic, and agent layer — which means it isn't actually a parallel product, it's the same product with different content. The "parallel" framing conflates corpus with architecture.
Stay governance-focused. Rejected. The light-paper positioning has already shifted externally. Reverting would create incoherence between marketing, agent-builder pitch, and the actual product capability. The cube enables multi-domain credibly; staying governance-focused would underuse the capability and undersell the product.
Wait for cross-domain experiment results (Phase 2 + 3) before deciding. Considered. Decided no: the experiment results would refine which signals hold cross-domain and which need recalibration, but the architectural-support finding is independent of the experiment outcome. The cube's ability to decompose 29 governance scenarios + design 10 cross-domain scenarios at constant vectors is already sufficient evidence. The experiment would inform signal-side calibration, not architecture-side feasibility.

Discussion¶

Key debate: is "the corpus is governance-heavy" a product property or just an artifact of sampling order? The answer concluded — artifact, based on the cube's ability to decompose any domain cleanly and the experiment's demonstration that constant cube vectors land in all five domains. This shifted the conversation from "Pulse is a governance product that could maybe expand" to "Pulse is a multi-domain judgment layer whose corpus happens to be governance-heavy because that's where sampling started."

Tension surfaced: light-paper positioning has shifted ahead of operational reality. Today's corpus is ~75% governance; today's onboarding flow uses governance-flavored examples; today's category/trigger taxonomy is work-shaped. The decision accepts that external positioning can lead operational reality by some weeks or months as long as the architecture supports the claim. The risk is that prospective users encounter the governance-flavored surface and conclude Pulse isn't actually multi-domain; mitigation is the 4–6 week corpus + taxonomy + UI catch-up work flagged above.

Open question on signal baseline drift: if signals were tuned on governance scenarios, they may not calibrate cleanly to personal-stakes scenarios. The trigger to revisit this ADR's "no rebuild needed" claim is: after 50+ non-governance scenarios are in the corpus and played, signals show calibration drift that doesn't recover with normal retuning. That would be evidence that the signal architecture is more governance-shaped than this ADR claims.

Cube as proof artifact: before the cube existed, multi-domain claims were assertion. Now they are structurally supported. The cube being a stable design tool that decomposes any domain is the load-bearing evidence for this ADR. If the cube turns out to be wrong (one of the falsification conditions in scenario-dimension-cube.md triggers), this ADR's foundation is also weakened.

Consequences¶

Light-paper revisions can lean on this structurally. The multi-domain positioning is no longer aspirational; it's architecturally backed.
Multiplayer corpus build doubles as multi-domain seeding. Multiplayer scenarios trend toward shared-life-decision rather than pure governance, which means the in-flight multiplayer corpus work is already advancing this ADR's roadmap.
Non-governance category/trigger expansion gets prioritized in the next quarter's roadmap. Specifically: add 1–2 categories, 2–3 triggers, refresh UI copy, audit existing tag taxonomy against the multi-domain frame.
Signal baseline drift may need correction after non-governance corpus expands. Detectable in routine signal-quality monitoring; budget 1–2 weeks of recalibration work as a follow-up after the corpus expansion lands.
Measurement schema audit becomes higher-priority. The cube systematized design primitives; no parallel exercise has systematized measurement primitives. Multi-domain expansion is a natural trigger to do this — see follow-up work tracked in next-conversation handoff.
Trigger to revisit: if after 50+ non-governance scenarios in the corpus, signal calibration shows drift that doesn't recover with retuning, or if the cube's falsification conditions trigger, this ADR's "no rebuild needed" claim should be re-examined.

Key files: - docs/research/foundations/scenario-dimension-cube.md — the structural model that proves multi-domain is architecturally supported - docs/research/experiments/experiment-new-domain-signals.md — empirical evidence that signals fire cross-domain - docs/research/experiments/cross-domain-consistency-experiment.md — current experiment validating cube vectors across domains - docs/decisions/024-judgment-layer-signals.md — the 12 signals that compose the measurement layer - docs/decisions/033-team-judgment-layer.md — earlier framing of the judgment layer as team-shaped; this ADR extends it - docs/decisions/045-five-sync-dimensions.md — sync dimensions; potentially affected by multi-domain calibration - docs/decisions/071-profile-embedding-across-agent-contexts.md — agent consumption layer that already works domain-agnostic - src/lib/ai/scenario-generator.ts — drivers (8), triggers (10), categories (4), value tensions (12) live here; this ADR informs which need additions