ADR-011: Delegation Map — From Single Score to Per-Context Authority

Status: Recommended Direction (not yet implemented) Date: 2026-03-19 Context: The composite sync score (30% accuracy, 20% coverage, 30% consistency, 20% information gain) was designed as the threshold for DTA delegation authority (target: 95%). Analysis revealed that 95% is mathematically unreachable for any human player under the current consistency formula — even a rigid player who always picks the same driver caps at ~85% composite. Contextual decision-makers (most humans) cap at ~70%.

At a glance

What it decides: stop using one composite sync score as the gate for DTA delegation, and use a per-context delegation map instead — the DTA acts autonomously where it's confident and flags for human review where it isn't. Recommended Direction, not yet implemented — deferred until the consistency-formula limits are confirmed across multiple 30+ game players, context-rule confidence data accumulates, and semantic trigger classification ships.

• Core mechanism — 4 categories × 6 triggers = 24 context cells, each with a behavior distribution, confidence, and exposure count; a cell auto-delegates only when it clears game-count, confidence, and accuracy thresholds.
• Why the single score fails — 95% is mathematically unreachable: the consistency dimension (HHI concentration) caps a contextual thinker near 70 and a rigid one near 86, and penalizes sophisticated context-switching.
• Main rejected alternative — "accept the safety gate" (contextual thinkers just get more oversight); folded into the chosen option, which satisfies both predictability and context-sensitivity per-context.
• Cost/dependency — needs semantic trigger classification (extend the classify-driver endpoint) before delegation authority can rely on context detection; keyword matching isn't reliable enough.
• Composite score survives as a player-facing health metric, decoupled from delegation authority.

The Problem

The consistency dimension (30% weight) uses Herfindahl index concentration to reward having one dominant driver across all contexts. This creates three failures:

1. 95% is unreachable. To score 95/100 on consistency, a player would need one driver at 96%+ in every category. That's a bot, not a person.

2. Contextual thinkers are penalized. A player who is pragmatic under external pressure but principled during ethical dilemmas — arguably more sophisticated — scores lower than someone who is always pragmatic regardless of context.

3. The DTA already knows more than the score reflects. Context rules (trigger → behavior_distribution with confidence scores) capture exactly the nuance the consistency dimension penalizes. The DTA can predict a contextual thinker accurately using context rules, but the sync score doesn't credit this.

Math: Why 95% Is Unreachable

Contextual player (near-perfect accuracy/coverage/info-gain, realistic consistency):

0.30(91) + 0.20(90) + 0.30(25) + 0.20(85) = 69.8 composite

Rigid player (perfect accuracy, same driver across all categories):

0.30(100) + 0.20(90) + 0.30(59) + 0.20(85) = 85.7 composite

Neither reaches 95. The consistency dimension is the bottleneck.

Recommended Direction: Per-Context Delegation Map

Replace the single-score delegation threshold with a per-context delegation map based on the DTA's confidence in each decision context.

flowchart LR
    subgraph Before
        S[Single composite score] -->|"≥ 95%?"| G[Delegate everything]
    end
    subgraph After
        C[Decision context] --> M[24 context cells]
        M -->|confident| A[Auto-delegate]
        M -->|low confidence / novel| F[Flag for human review]
    end

Before: one unreachable threshold gates all delegation. After: each context cell delegates or flags on its own confidence.

How It Works

The delegation map is built from existing data structures: 4 scenario categories × 6 context triggers = 24 context cells. Each cell has: - A behavior distribution (from context rules) - A confidence score (from context rule confidence) - An exposure count (from trigger_exposure tracking)

Auto-delegation criteria per context cell (thresholds tunable):

Can auto-delegate if:
  - Category has 8+ games (sufficient category pattern data)
  - All detected triggers have confidence ≥ 0.40 AND exposures ≥ 8
  - Recent accuracy in this category ≥ 60%
Otherwise: flag for human review

Graduated Sovereignty

This naturally produces the governance model described in the litepaper (Section 3.2.4):

1. Early games: All contexts flag for human review. The DTA is learning.
2. Mid-play (20-30 games): Routine contexts with high-confidence rules auto-delegate. Novel or low-confidence contexts still flag.
3. Mature (50+ games): Most governance decisions auto-delegate. Only genuinely novel contexts or low-confidence triggers require human input.
4. Over time: As the DTA accumulates evidence in edge-case contexts, the delegation frontier gradually expands.

Example: CoachJ's Current Delegation Map

Context	Confidence	Exposures	Status
External pressure → pragmatism	24%	22	Flag (low confidence despite high exposure — ambiguous tendency)
Crisis → pragmatism	17%	19	Flag (low confidence)
Resource constraint → team_harmony	28%	13	Flag (approaching threshold)
Team conflict → team_harmony	29%	7	Flag (insufficient exposure)
Ethical dilemma → principle	33%	5	Flag (insufficient exposure)
Opportunity → pragmatism	28%	16	Flag (low confidence)

CoachJ has zero auto-delegated contexts yet. The DTA is still learning — which is honest, given 50% prediction accuracy.

What Happens to the Composite Score

The single sync score stays as a player-facing health indicator ("How well does your DTA know you overall?") but is decoupled from delegation authority. Think of it like:

• Sync score = motivational metric, shown on hub and profile
• Delegation map = operational metric, determines what the DTA can act on autonomously

The consistency dimension can be reworked to measure context-rule confidence breadth instead of HHI concentration — rewarding players whose DTAs have learned confident, differentiated rules across many contexts.

Context Space: Fixed but Expandable

The delegation map's context space is intentionally finite:

• 4 categories: governance, resource-allocation, team-dynamics, values-culture
• 6 triggers: crisis, team conflict, resource constraint, opportunity, ethical dilemma, external pressure
• 8 drivers: team_harmony, efficiency, principle, pragmatism, caution, growth, transparency, autonomy

Adding a new trigger or category is additive — it creates new delegation cells that start in "flag for review" mode. Existing delegation authority is unaffected. The architecture supports expansion without restructuring.

Implementation Note: Semantic Trigger Classification

Context triggers are currently detected via keyword matching on scenario text (e.g., "budget" → resource_constraint). For a delegation map that determines governance authority, this should be upgraded to semantic classification via API call — the same pattern as the existing classify-driver endpoint. This could be bundled into the classify-driver call (it already analyzes scenario + option context). Cost: negligible (small prompt, structured JSON output).

This upgrade matters because: - Keyword matching misses scenarios about resource constraints that don't use the word "budget" - Keyword matching false-positives on scenarios that mention "budget" incidentally - Delegation authority should be based on reliable context classification, not string matching

Decision

Adopt the delegation map as the recommended architecture for DTA delegation authority. The composite sync score remains as a player-facing metric but is not the delegation gate. Implementation deferred until: 1. Current consistency formula limitations are confirmed with multiple players at 30+ games 2. The context rule confidence data has accumulated enough to validate the per-context thresholds 3. Semantic trigger classification is implemented to ensure reliable context detection

How We Got Here

This recommendation emerged from a design exploration on 2026-03-19, starting from the question: "What does a 95% profile actually look like?"

Step 1: Dimension-by-dimension breakdown

Walking through each dimension's requirements for 95% revealed the mechanical reality:

• Accuracy (95/100): Needs ~77% weighted accuracy over last 20 predictions (roughly 16/20 correct). The code already treats 80% raw as the ceiling (perfect score), acknowledging that genuine dilemmas make 100% accuracy unreachable. This dimension is achievable.
• Coverage (95/100): Needs all 4 categories with balanced play (Gini ≤ 0.20), all 6 triggers encountered, 30+ unique scenarios. Achievable with enough games and category balancing (already implemented).
• Consistency (95/100): Needs HHI of ~0.93, meaning one driver at 96%+ across all categories. This is where the formula breaks. No real human has driver distributions this concentrated.
• Information Gain (95/100): Needs sustained novelty — new scenarios, not replays. Achievable but gets harder over time as the target scales with games played.

The key finding: accuracy, coverage, and information gain can all reach 90+. Consistency cannot exceed ~25 for a contextual thinker or ~59 for a rigid thinker. The consistency dimension is the sole bottleneck making 95% unreachable.

Step 2: Framing the tension

This surfaced two competing models of "alignment":

• Model A (Predictability = Alignment): The DTA can predict you, so it's safe to delegate. Consistency makes prediction easier. The formula correctly rewards this.
• Model B (Calibrated Context-Sensitivity = Alignment): A DTA that knows when you shift drivers and why is more aligned than one that just knows your default. The context rules already capture this — the scoring system just doesn't credit it.

The litepaper explicitly supports Model B: "coherence must be earned across diverse decision contexts" (A.1.6). The current scoring formula supports Model A.

Step 3: Three resolution options

Three directions were considered:

(A) Fix the formula. Make 95% reachable for contextual thinkers by reworking consistency to reward context-appropriate patterns instead of rigidity.

(B) Accept the safety gate. Contextual thinkers correctly get more human oversight. Their DTAs flag more decisions for review — predictability IS trust, and maybe we need to face the reality that perfect digital twins may not be achievable.

(C) Per-context delegation. The threshold isn't one number. DTAs get autonomous authority in contexts where they're confident, and flag for review where they're not. The delegation frontier expands gradually.

Step 4: The insight that resolved it

The key realization (from the discussion): "Routine decisions would get delegated away while high-impact novel decisions would require oversight, and then slowly over time more and more of those high-impact and/or novel decisions could theoretically get delegated away as the AI better understood the human's tendencies."

This matched the litepaper's vision (Section 3.2.4): "Routine decisions get executed autonomously. High-stakes or novel decisions get flagged for human review." Option C wasn't a compromise — it was the original governance design, now grounded in the scoring system's actual data structures.

Options A and B both became true within Option C: - A is satisfied: Contextual thinkers aren't penalized. Different drivers per context = more information for the delegation map. - B is satisfied: Predictability IS rewarded, but per-context. Low confidence in a context = more oversight. The system naturally requires human involvement exactly where the human is genuinely ambiguous.

Step 5: Validating the context space

A concern was raised: are the 4 categories and 6 triggers too limited for a delegation map? Analysis showed:

• The context space (4 × 6 = 24 cells) is intentionally finite, designed to cover the decision space relevant to organizational governance.
• Adding new triggers or categories is additive — new cells start in "flag for review" mode without affecting existing delegation authority.
• The 12 "dimensions tested" in scenario generation (risk tolerance, people vs outcomes, etc.) are currently untracked but could be promoted to operational triggers if needed.

This confirmed the delegation map concept is robust to expansion.

Step 6: Semantic trigger classification

The final insight: context triggers are currently keyword-matched, which is adequate for profile building but insufficient for delegation authority. The existing classify-driver API endpoint provides the pattern — extend it to also classify triggers semantically. Cost is negligible (small prompt, structured JSON). This upgrade would make the delegation map reliable enough to base governance authority on.

Broader question left open

Whether we can ever truly have digital twins, or whether some level of human oversight will always be necessary for increasingly complex contextual decisions, remains an open question. This ADR's recommendation is compatible with both outcomes — the delegation map can expand to near-full autonomy if the DTA proves capable, or stabilize at partial delegation if certain contexts remain fundamentally unpredictable. The system doesn't presume the answer; it lets the data decide.

Key Files

• src/lib/sync/multi-dimensional-score.ts — current consistency formula (to be reworked)
• src/lib/ai/player-model.ts — context rules, trigger detection, driver distributions
• src/app/api/ai/classify-driver/route.ts — existing semantic classifier (extend for triggers)
• src/app/api/sessions/start/route.ts — category balancing (already supports coverage goals)