Strategy Module: Trendslop Mitigation¶

Structural mitigation against LLM tendency to recommend trendy, context-insensitive strategies ("trendslop") for strategic agent roles.

Module: src/synthorg/engine/strategy/

Covers the core models, config, and prompt integration documented on this page.

Background¶

Industry research shows LLMs systematically recommend trendy, context-insensitive strategies across 7 core business tensions. Prompt-level fixes produce only marginal bias reduction. SynthOrg mitigates this structurally through constitutional principles, multi-lens analysis, a confidence-calibration prompt instruction, and output mode control.

Strategic Output Modes¶

Controls how strategic agents frame their recommendations. Set per-agent via AgentIdentity.strategic_output_mode or company-wide via strategy.output_mode.

Resolution: agent override > config default. context_dependent resolves to decision_maker for C-suite/VP, advisor otherwise.

Strategic Lenses¶

8 analysis perspectives forced on strategic agents:

Default (always active)¶

Optional (enabled via config)¶

Constitutional Principles¶

Anti-trendslop rules loaded from YAML packs and injected into system prompts. Each principle has an ID, text, category, and severity level (informational, warning, critical).

Built-in Packs¶

Pack Schema¶

name: "pack-name"
version: "1.0.0"
description: "Pack description"
principles:
  - id: "principle_id"
    text: "Rule text injected into prompts"
    category: "category_name"
    severity: "critical"  # informational | warning | critical

User packs: ~/.synthorg/strategy-packs/<name>.yaml (override builtins by name).

Confidence Calibration¶

prompt_injection.py injects a fixed instruction asking strategic agents to state, in their own recommendation text, a confidence level, an upside/downside range, key assumptions, and what would change the recommendation. This is the whole mechanism: it does not vary with StrategyConfig.confidence.format (the structured / narrative / both / probability enum), and nothing parses the agent's stated confidence back into a structured record. impact.py and confidence.py -- the scorer and formatter that would have turned a recommendation's risk profile and stated confidence into ImpactScore / ConfidenceMetadata and attached them to a DecisionRecord -- had no production caller and were removed. StrategyConfig.cost_tier, StrategyConfig.confidence.format, and StrategyConfig.progressive are consequently unconsumed: they parse and validate but select nothing.

Prompt Injection¶

Strategic sections are injected into the system prompt after autonomy instructions, before the task section. Injection occurs when:

1. Agent has explicit strategic_output_mode, OR
2. The agent's role sits in the executive tier: reporting depth <= 1 (the CEO and its direct reports), via role_depth(agent.role)

Injected Sections¶

1. Strategic Analysis Framework: maturity stage, industry, competitive position
2. Constitutional Principles: anti-trendslop rules from active pack
3. Contrarian Analysis: forced opposite-case consideration
4. Confidence Calibration: fixed instruction to state confidence, range, and assumptions
5. Assumption Surfacing: explicit assumption listing
6. Output Requirements: mode-specific output instructions

The strategy section is trimmable (removed first when over token budget).

Config Shape¶

strategy:
  output_mode: "advisor"
  cost_tier: "moderate"
  default_lenses:
    - contrarian
    - risk_focused
    - cost_focused
    - status_quo
  constitutional_principles:
    pack: "default"
    custom: []
  confidence:
    format: "structured"
  conflict_detection:
    strategy: "auto"
  context:
    source: "config"
    maturity_stage: "growth"
    industry: "technology"
    competitive_position: "challenger"
  progressive:
    weights:
      budget_impact: 0.2
      authority_level: 0.15
      decision_type: 0.15
      reversibility: 0.2
      blast_radius: 0.1
      time_horizon: 0.1
      strategic_alignment: 0.1
    thresholds:
      moderate: 0.4
      generous: 0.7

Decision Records¶

DecisionRecord.risk_card is an optional RiskCard field (decision type, reversibility, blast radius, time horizon). It is nullable and defaults to None; no production path constructs a DecisionRecord with it populated today. ConfidenceMetadata and LensAttribution -- the structured capture of a recommendation's stated confidence and per-lens attribution -- had no production caller and were removed along with impact.py / confidence.py.

Architecture¶

Protocol Pattern¶

The surviving major component is pluggable behind @runtime_checkable Protocol:

ImpactScorer (CompositeImpactScorer, ExplicitImpactScorer, HybridImpactScorer) and ConfidenceFormatter (StructuredFormatter, NarrativeFormatter, BothFormatter, ProbabilityFormatter) had no production caller and were removed with impact.py / confidence.py.

Module Layout¶

engine/strategy/
  __init__.py                    -- Public exports
  models.py                      -- Config + domain models (frozen Pydantic)
  lenses.py                      -- StrategicLens enum + definitions
  principles.py                  -- Pack loading service
  active_principle.py            -- Active-principle resolution
  active_principle_provider.py   -- Active-principle context provider
  principle_override_provider.py -- Per-scope principle overrides
  context.py                     -- Context providers
  strategic_context_provider.py  -- Context provider protocol
  adapter.py                     -- Strategy adapter for the engine
  scoping.py                     -- Scope resolution
  output.py                      -- Output mode handler
  prompt_injection.py            -- Prompt section builder
  packs/                         -- Built-in YAML principle packs
    default.yaml
    startup.yaml
    enterprise.yaml
    cost_sensitive.yaml

References¶

• Prompt injection entry point: src/synthorg/engine/strategy/prompt_injection.py