Tier Manager¶

tier_manager -- mvp.tier_manager

Cluster: Uncategorised | Type: component | MCP Tools: 10

Overview¶

The Tier Manager provides a single system-wide tier state with explicit user-controlled advancement. It bridges the gap between individual tier components (self_training, evoskill, formal_methods) and gives users visibility and control over which adaptation layer is active.

State is persisted to an Obsidian vault with YAML frontmatter, wikilinks, and human-readable markdown — consistent with G6's explicit reasoning thesis.

Operations¶

Tier Descriptions¶

Advancement Rules¶

• T0 -> T1: Always ready. Just needs user confirmation.
• T1 -> T2: Requires >= 60% of components at maturity tier >= 3 (checked via self_training.get_competence).
• T2 -> T3: Requires non-empty Pareto frontier with at least one skill scoring >= 0.7 (checked via evoskill.get_frontier).
• Advancement is never automatic — the system proposes, the user confirms or rejects.

Vault Structure¶

tier-manager/
  status.md              # Current tier state (YAML frontmatter)
  history/
    2026-04-22-T0-to-T1.md    # Advancement decisions
    2026-04-23-rejected-T2.md  # Rejected proposals
  readiness/
    2026-04-22-T1-audit.md     # Readiness audit results

Each history note links back to [[status]] via wikilinks. The status note includes a Dataview query for recent history.

Surface Integration¶

MCP¶

get_status
assess_readiness target_tier=2 readiness='{"ready": false, "evidence_status": "failed_evidence"}'
audit_readiness
list_patterns
describe_tier

The dedicated MCP package is intentionally read-only for tier state: mutation operations (propose_advance, confirm_advance, reject_advance, cap_tier, uncap_tier, sync_vault) remain local-block-only and are surfaced as excluded operations by describe_tier / list_patterns.

REST¶

# Get current tier
curl http://localhost:8000/tier/status

# Propose advancement
curl -X POST http://localhost:8000/tier/advance \
  -H "Content-Type: application/json" \
  -d '{"action": "propose", "reason": "ready"}'

# Confirm
curl -X POST http://localhost:8000/tier/advance \
  -H "Content-Type: application/json" \
  -d '{"action": "confirm"}'

# Cap at T2
curl -X POST http://localhost:8000/tier/cap \
  -H "Content-Type: application/json" \
  -d '{"tier": 2}'

TUI¶

• Status bar: Shows current tier (T0-T3) with colour coding
• t key: Opens tier popup with status, history, advance/confirm/reject buttons, and cap controls

Source Files¶

See Also¶

• Adaptive Tiers (Bases Overview)
• EvoSkill — T2 skill evolution
• Formal Methods — T3 verification
• Theoretical Foundations — practopoiesis framework

Public API¶

TierManagerDecisionError(ValueError)¶

The LLM did not produce a usable, validated tier-readiness decision.

ReadinessDecision¶

Validated tier-advancement readiness-posture verdict.

Methods:

escalated() -> bool¶

agentic_evidence() -> dict[str, Any]¶

to_dict() -> dict[str, Any]¶

TierManagerPlanner¶

Runtime-first facade with deterministic fallback + one-way posture clamp.

Methods:

assess(target_tier: int, readiness: dict[str, Any], tier_cap: int = 3, current_tier: int | None = None) -> ReadinessDecision¶

TierManagerPatternRuntime¶

Stateless executable mechanisms for tier_manager's applied patterns.

Methods:

tighten(posture: Any, baseline_posture: Any = '', requires_review: bool = False, degraded: bool = False, not_ready: bool = False, non_passing_evidence: bool = False, cap_exceeded: bool = False) -> TierManagerPatternReview¶

One-way readiness-posture guard (hook-based-safety-guard-rails).

AdaptiveTier(IntEnum)¶

TierState(BaseModel)¶

TierInput(BaseModel)¶

TierOutput(BaseModel)¶

TierManagerBlock(AIBlock[TierInput, TierOutput, TierState])¶

Methods:

infer(data: TierInput) -> Result[TierOutput]¶

Functions¶

deterministic_readiness_decision(target_tier: int, readiness: dict[str, Any], tier_cap: int = 3, current_tier: int | None = None) -> ReadinessDecision¶

Demoted-real STRUCTURAL baseline (the zero-LLM floor).

apply_readiness_floor(baseline: ReadinessDecision, candidate: Any) -> ReadinessDecision¶

One-way readiness-posture safety clamp (stricter-of, ESCALATE only).

build_default_planner() -> 'TierManagerPlanner | None'¶

Build the on-by-default agentic planner, or None when suppressed.

grounded_readiness_decision(target_tier: int, readiness: dict[str, Any], tier_cap: int = 3, current_tier: int | None = None, planner: TierManagerPlanner | None = None) -> ReadinessDecision¶

The ONE shared grounded chokepoint for EVERY tier-readiness decision.

applied_agentic_patterns() -> list[dict[str, Any]]¶

Return compact metadata for tier_manager-applied vendored patterns.

MCP Tools¶