ADR-036: Sync Measures Procedural Quality and Legibility, Not Values or Uniformity¶

Status: Accepted (philosophical foundation) Date: 2026-04-30 Context: A recurring philosophical question shadows Sync's design: what exactly is the score measuring, and what is a "good" decision-maker? Two adjacent traps surface whenever this is pressed:

The temptation to score substantive quality — "did the player reach the right answer" — which requires the platform to adopt contested values and ends up being ideology in twin-clothing.
The temptation to score consistency — reward sameness across scenarios — which punishes growth, context-sensitivity, and updating, and produces brittle twins that are wrong in the moments that matter.

Both temptations look reasonable from the outside (one preserves "good" outcomes, the other preserves predictability) but each is incompatible with Sync's stated philosophy that the score measures how you actually decide, not how you think you should decide. This ADR commits the project to the alternative on both axes.

At a glance

What it decides: Sync scores the quality of your reasoning system and the legibility of its variance — not your values, and not your consistency.

Procedural, not substantive — Sync measures whether you're actually reasoning (calibration, updating, framing-insensitivity, articulability), regardless of what you conclude. The score is orthogonal to your values.
Legibility, not uniformity — articulable variance is signal; opaque variance is noise. A consistent player and a context-sensitive player can both score high if the pattern is legible.
Rejected: scoring substantive correctness (makes Sync an ideological filter) and scoring uniformity (punishes growth, produces brittle twins). Also rejected: gatekeeping low-quality reasoners at entry.
Consequence: twins represent distributions, not point estimates — they know their own limits and fail safely; delegation calibrates to per-region legibility.
Watch: product copy drifting toward "Sync measures whether you're a good person" or "rewards consistent thinkers" — both signal the commitment is leaking.

flowchart LR
    subgraph IN["In scope — what Sync measures"]
        A["Procedural quality<br/>(calibration, updating,<br/>framing-insensitivity)"]
        B["Legibility of variance<br/>(articulable context-sensitivity)"]
    end
    subgraph OUT["Out of scope — what Sync does NOT measure"]
        C["Substantive correctness<br/>(the 'right' values)"]
        D["Uniformity<br/>(sameness across scenarios)"]
    end

Sync's measurement target: reasoning quality and legible variance are in scope; values and consistency are not.

Decision¶

Sync's measurement target is two-part:

Procedural quality, not substantive correctness. Sync measures whether the player is actually reasoning — calibration, updating, framing-insensitivity, articulability, considering relevant factors, treating disagreement as informative — regardless of what they conclude. The score is orthogonal to the player's values.
Legibility of variance, not uniformity. Sync rewards the player (and their twin) being able to articulate and predict their own variance, not eliminate it. Articulable variance is signal; opaque variance is noise. A consistent player and a context-sensitive player can both score high, as long as the context-sensitive one's pattern is legible.

Together: Sync measures the quality of the player's reasoning system and the legibility of its outputs, not conformity to specific values or to a flat point-estimate self.

Rationale¶

Why procedural, not substantive¶

There is no value-neutral platform from which to score whether someone reached the "right" answer on a contested question. A pro-life player and a pro-choice player can both be highly calibrated, well-updating, framing-insensitive thinkers; or both can be brain-stem reactors. A platform that scores substantive correctness has to declare contested values as ground truth — at which point Beacon is no longer a measurement instrument, it's an ideological filter wearing measurement clothes.

Procedural quality has a different status. It's not value-neutral in some absolute sense, but it's value-orthogonal: regardless of what you value, are you reasoning toward it coherently or sabotaging yourself? The 12 behavioral signals (ADR-024) are largely procedural. Calibration is procedural. Framing-sensitivity is procedural. Updating is procedural. None of them require the platform to take a position on contested substance, and they remain meaningful across the full distribution of human values.

This also resolves the social-media counter-product question: building a tool that rewards deliberation isn't ideologically loaded the way "rewarding the right values" would be — it's restoring the deliberative capacity that brain-stem-optimized media has actively degraded. You don't have to claim systems thinking is objectively correct to claim people should be able to reason if they want to.

Why legibility, not uniformity¶

The "consistency = quality" assumption is doing more work than it deserves. It treats variance as noise and the consistent player as the gold standard. But variance has multiple sources, and they are not equivalent:

Drift over time → learning rate (high-quality signal)
Same-day flip on identical scenario → occasion noise (often acceptable; sometimes diagnostic)
Flip across similar-but-not-identical scenarios → framing sensitivity (sometimes a feature, sometimes a flaw — depends on whether the framing change should matter)
Articulable context-sensitivity → legitimate context-dependence (high-quality)
Random un-articulable flip → opaque variance (the actual problem)

Scoring uniformity collapses all of these into "bad." Scoring legibility distinguishes them. A player who says "I flip on team-vs-individual scenarios when I'm tired, here's why" is producing a better twin than a player whose scenarios always look the same but who can't say why.

Legibility also has the right scaling property for delegation. A trustworthy delegate isn't one who pretends consistency it doesn't have — it's one that can say "in this region of decisions I'm reliable; in this region I'm noisy; here's the variable that flips me." That's a usable instrument. Pretended consistency is a liar at scale.

Reconciliation with ADR-013 ¶

ADR-013 ("Deliberative baseline — training the DTA on your higher self") established that the DTA is trained on the player's deliberative state, not their reactive state. That decision is preserved and reinforced here — but its language gets refined.

ADR-013 framed the deliberative state as the "higher self." Under ADR-036, the deliberative state is more accurately understood as the condition under which procedural quality can be measured cleanly, not as a substantively superior version of the player. The DTA isn't capturing who the player wishes they were; it's capturing how the player reasons when reasoning is possible. That's a procedural distinction, not a substantive one.

ADR-013 also stated: "session-to-session variance is noise; the persistent signal is who they are when they think clearly. EMA smoothing naturally filters for this persistent signal." This claim is partially superseded. Under ADR-036, even within the deliberative baseline, articulable variance is signal, not noise. EMA smoothing remains useful for damping random fluctuation, but the model should additionally surface and represent legible variance — it shouldn't flatten it.

Why this is the only honest design at scale¶

A product that scores substantive correctness will inevitably encode the designers' values, and at scale will impose them on every player. A product that scores uniformity will reward people for performing stability they don't have and punish people for the kind of context-sensitivity that's often appropriate. Both produce twins that are easier to predict and easier to misuse — confidently wrong in exactly the regions where the player's real judgment is most fragile.

A product that scores procedural quality and legibility produces twins that know their own limits. They fail safely. They flag low-confidence regions. They can be paired with scaffolding (ADR-016 delegation, ADR-031 orientation) to get higher-quality outcomes than the player alone. This is the only scoring philosophy that's compatible with Beacon's longer-term claim that twins can be trusted instruments rather than performative simulacra.

Alternatives Considered¶

Score substantive quality (capture the "higher self" as values). Reward players for reaching answers the platform considers correct. Rejected: requires the platform to adopt contested values, makes Sync an ideological filter, fails the identity vs. process test, and produces twins that flatter rather than represent.
Score uniformity (consistency as the quality target). Reward players for the same answer across scenarios. Penalize variance. Rejected: punishes growth and updating, produces brittle twins, conflates random flip with legitimate context-sensitivity, and is incompatible with the occasion-noise signal already in ADR-024.
Gatekeep at entry — admit only deliberators. Filter out players with low procedural quality before they can produce twins. Rejected: builds a parallel governance class, abandons the population most in need of the tool, and is structurally identical to the elite/populist failure modes Beacon is meant to counter. The procedural score itself does the differentiation more honestly — low-quality reasoners get low Sync scores and aren't delegated to, without being excluded from playing.
Score procedural quality but flatten variance via EMA only (status quo from ADR-013). Capture deliberative reasoning, then smooth variance away as noise. Rejected as insufficient: even within the deliberative baseline, articulable variance is signal. The model should represent the player's distribution, not just its mean.

Discussion¶

This ADR was written in response to a sustained philosophical conversation about whether Sync's design risks either (a) imposing the designers' values on players or (b) demanding a kind of consistency that real humans don't have. Both worries surfaced at once and the resolution required separating them.

On the "higher self" framing. ADR-013 used the language of a player's "higher self" to describe the deliberative baseline. In conversation it became clear that "higher self" can mean three different things: (1) aspirational self — who the player thinks they should be, which is performance not signal; (2) best-case self — the player's good days, which is selection-biased; (3) reflective self — the player when they have time and conditions to think, which is what ADR-013 actually captured. Only the third is honestly capturable, and it's procedural (a state of the system) rather than substantive (a better version of the person). This ADR preserves ADR-013's substance but disambiguates the language.

On the "but isn't refusing to think anti-human" worry. Raised seriously: if Sync rewards procedural quality and someone refuses to engage, isn't that morally significant, and shouldn't the platform exclude them? The conclusion was no, for two reasons. First, "refusing to think" is almost always a state (exhaustion, threat, attention extraction) rather than a trait — treating it as identity is the same move Sync's philosophy already refuses. Second, gatekeeping at entry creates a parallel governance class, which is structurally the failure mode Beacon is supposed to counter. The procedural score itself differentiates without excluding. Low-quality reasoners get low trust; they don't get banned from the platform.

On the social-media counter-product framing. A concern was raised that procedural neutrality is itself a value — that in an information environment optimized for brain-stem reaction, "scoring deliberation" isn't neutral. Accepted, with the resolution: Sync isn't claiming procedural neutrality is value-free; it's claiming procedural quality is value-orthogonal. Players with very different values can both be high-quality reasoners. That's a much smaller and more defensible claim than scoring substantive correctness.

On EMA and variance. Tension surfaced between ADR-013's "session-to-session variance is noise" and the legibility framing. Resolution: EMA smoothing remains correct for damping random fluctuation, but the model should additionally represent legible variance as a feature of the player's distribution, not flatten it. ADR-013 is partially superseded on this point.

On where this lands. The product's measurement target is the quality of the player's reasoning system and the legibility of its outputs. Not their values. Not their consistency. This produces honest twins that fail safely, and it preserves the procedural/substantive separation that lets Beacon scale across players with different values without imposing on any of them.

Consequences¶

Scoring methodology. Sync's scoring should weight procedural signals (calibration, updating, framing-insensitivity, articulability) and legibility (predictability of variance, articulable context-sensitivity) as primary. Uniformity per se is not a quality dimension.
Repeated scenarios are not failures. A player who chooses differently on a repeated or near-repeated scenario is not penalized for inconsistency. They are scored on whether they (or their model) can articulate the variance. Articulable variance is full-marks; opaque variance surfaces as a low-legibility region.
Twin output shape. Twins should represent distributions, not point estimates. A delegation-ready twin can say "in this region I'm reliable; in this region I'm noisy; here's what flips me." Twins that pretend point-estimate certainty over high-variance regions are doing the wrong thing.
Delegation contracts. Per ADR-011 and ADR-016, delegation should be calibrated to the legibility of the relevant region, not just the player's overall Sync score. Low-legibility regions trigger human-in-loop; high-legibility regions (whether consistent or context-dependent-but-articulable) can auto-delegate.
No values gatekeeping. Beacon does not filter players at entry based on their values or current procedural quality. The score itself differentiates, and players can grow into higher trust over time. This is load-bearing for the project's structural defensibility.
Partial supersession of ADR-013. ADR-013's deliberative-baseline training is preserved. Its claim that "session-to-session variance is noise" is partially superseded: random variance is noise, but articulable variance is signal and should be represented by the model rather than smoothed away.
What to watch for. If product surfaces (UI, copy, marketing) start drifting toward "Sync measures whether you're a good person" or "Sync rewards consistent thinkers," that's evidence the philosophical commitment is leaking. Both framings should be actively resisted.

Key files: - docs/decisions/013-deliberative-baseline-training.md — partially superseded by this ADR on the variance-as-noise claim; the deliberative-state training premise is preserved - docs/decisions/024-judgment-layer-signals.md — the 12 procedural signals operationalize the procedural-quality side of this ADR; occasion_noise and self-awareness are the legibility-side signals - docs/decisions/004-ema-profile-updates.md — EMA remains correct for damping random fluctuation, but should not be the only mechanism for handling variance - docs/decisions/011-consistency-vs-sophistication.md — delegation map per region of decision; legibility refines what "high confidence in a region" means - docs/decisions/016-delegation-game-action.md — delegation contracts should be calibrated to legibility, not just overall score - docs/decisions/027-scenario-content-guardrails.md — content philosophy aligns with procedural-not-substantive (no scenarios that try to impose values via decision substrate)