Measurement schema · v0 stable

Pulse Measurement Schema · data flow & emergence

Four causal layers feed forward: what a player does becomes what the system infers, rolls up, and exposes. Measured outcomes sit outside the chain — they emerge from player × system interaction and are never written directly. Click any primitive for its definition, source, mode, and what it feeds. Canonical IDs match docs/research/measurement-schema.md.

The pipeline at a glance click a stage to jump in · use the carets (or these buttons) to collapse detail

⬡ Cube substrate The scenario's design inputs that feed the extraction step — set by the author, not measured by the system. ⓘ

L114

CaptureWhat the player does — raw moves recorded each round.

→

L222

ExtractionWhat the system infers — computed from captures.

→

L313

AggregationRolled up per player across their whole history.

→

L411

ExposureWhat actually reaches users & agents.

⇢

Measured outcomesEmergent — revealed by play, never written directly.

L1 Capture L2 Extraction L3 Aggregation L4 Exposure Measured outcome (emergent) Agent-side signal Cube substrate (input) SOLOMULTI = mode-specific

L1 Capture — per-round records, frozen at write 14 primitives · written when a round is played

Game-side

C1Option choice

A/B/C/D / none-of-the-above

C2Rationale text

free-text

C3Conviction

1–5

C4Decision time

C5Post-reveal reaction

gets_me / off_track / interesting / null

C6Post-reveal note

free-text

C7Milestone self-prediction

at game 10 / 25 / 50

SOLO

C8Peer prediction

MULTI

C9Predictor confidence

MULTI

C10Post-discussion action

stand / revise / delegate

MULTI

C14Session metadata

timestamps · mode · device

Agent-side

C11Agent profile read event

AGENT

C12"Why" request flag

AGENT

C13Agent response feedback

rating + free-text

AGENT

▼computed from captures▼

⬡ Cube substrate The scenario's design inputs — set by the author before anyone plays, not measured by Pulse. Several extractions read a player's choice against them (e.g. which driver an option stood for, which trigger fired). cube-C1 / cube-C2 are families of those design axes. Click for the full explanation. ⓘ

└─ feeds ─┐

L2 Extraction — computations on capture 22 primitives · computed from captures

Write-time — computed & stored at round time (frozen)

E_ai_predAI twin prediction + confidence

E_rule_predRule-engine prediction

E_classify_driverDriver / reasoning-mode / attention class.

Claude call

SPOF

E_chosen_driverChosen driver

substrate × option

⬡ cube

E_chosen_triggerChosen trigger

substrate × option

⬡ cube

E_sync_matchSync match

prediction = choice?

Read-time — computed on demand (always current)

E5Confidence calibration

E6Decision timing signature

E7Persuadability profile

domain-conditional

MULTI

E8Peer read confidence

MULTI

E9Dissent profile

E10Rationale-choice expression match

was: "authenticity"

SPOF-dep

E11Learning rate

E12Occasion noise

same rationale + diff choice

SPOF-dep

E13Pragmatic routing

decline-the-binary pattern

E14Ideal-actual reference frequency

was: "aspiration gap"

SPOF-dep

E15Reasoning style aggregate

SPOF-dep

E16Attention model aggregate

SPOF-dep

E17Arc trajectory

per campaign

E18Self-awareness

SOLOSPOF-dep

E19Drift indicator

E20Domain-confidence inference

E21Cross-domain signal consistency

needs multi-domain corpus

low-N

E22Context-bleed classification

real-world parallel mentions

SPOF-dep

▼rolled up per player▼

L3 Aggregation — per-player roll-ups, all carry an N denominator A1–A14 (A3 retired) · per-player roll-ups

Game-side roll-ups

A1Sync score

lifetime

A2Sync dimensions

Accuracy / Coverage / Readability

A4Category patterns

SPOF-dep

A5Context rules

A6Option preferences

A7Signal aggregates

JSONB

SPOF-dep

A8Resonance stats

reactions + note semantics

SPOF-dep

A9Calibration stats

A10Behavioral trajectory

per campaign

Agent-side roll-ups

A11A_why_rate

AGENT

A12A_feedback_engagement

AGENT

A13A_feedback_positivity

classifies free-text too

AGENTSPOF-dep

A14A_context_bleed_rate

AGENT

▼surfaced to users & agents▼

L4 Exposure — external surfaces: UI, agent context, MCP, share cards 11 surfaces · what reaches users & agents

X1Sync score headline

X2Three Sync Dimensions

Play screen

X3Behavioral attention signals

insights page

X4Category patterns cards

X5Agent context block

prose summary · mainstream paste-user path

AGENT

X6Personal context field

X7MCP tools

get_profile / get_recent_decisions / submit_feedback

AGENT

X8Drift detection nudge

X9Activity log

X10Share cards / Twin pages / rankings

X11"Why" expansion text

Measured outcomes EMERGENT · player × system · not directly written

M1Identity load

M2Subjective stakes threshold

M3Domain confidence asymmetry

M4Calibration drift

M5User trust state with agent

M6Aspiration gap magnitude

M7Frame-restructuring tendency

M8Cross-domain signal stability

M9Play-time context bleed

Agent-side chain parallel signal path

L1 · read event · "why" flag · response feedbackcaptured agent-side (C11–C13)

▼

L3 · A_why_rate · A_feedback_engagement · A_feedback_positivityagent-side roll-ups (A11–A14)

▼

→ M5 · User trust state with agent (emergent)

⚠ Reliability overlays — three orthogonal dimensions

C6Sample-size convention. Every aggregation carries an explicit N denominator. Extractions needing a minimum corpus (e.g. cross-domain signal consistency, domain-conditional persuadability) are unreliable until threshold is met. Flag low-n outputs; don't surface them as settled.
C7Staleness convention. State-sensitive measured outcomes (calibration drift, trust state) carry an explicit "as of [date]". Read-time extractions and aggregates degrade as underlying play data ages — timestamp every read-time value.
SPOFCentral classification SPOF. E_classify_driver routes through one Claude call feeding the whole driver/reasoning/attention line — and unlike the ground-truth-checked predictions, it fails silently, so it's the one dependency worth watching. Defended per ADR-080 (verbal-style-normalized prompt, prompt versioning, keyword-fallback cross-check). Click for the dependency graph.