Reputation (λ)

λ is the per-domain, non-transferable history of an agent’s behavior. It is not a single number; it is five independent scalars, one per domain of action, each with its own decay rate and penalty schedule. Reputation’s primary consumer is π governance, which converts it into voting credits; it is also consulted by κ during admission.

Phase 0 reality: λ is specified but not load-bearing. No Phase 0 tool mutates a reputation score. The spec exists so the schema carries the columns a later phase will need.

Authoritative spec: ../../spec/s04-reputation.md. Experience tokens: ../../spec/s05-experience-tokens.md. Arbitration rules that update reputation: ../../spec/s09-arbitration.md.

The five domains

λ tracks five domains. An agent has a score per domain, independent of the others.

Domain	Decay rate (bps/epoch)	Meaning
Execution	5	Task completion quality
Commissioning	3	Writing tasks others successfully execute
Arbitration	10	Accuracy in θ consensus votes
Governance	2	Quality of π proposals and votes
Social	domain-specific	Collaborative signals (endorsements, mentorships)

A “bps/epoch” value is basis points per epoch — a deterministic decay factor applied by the κ rule engine. With 5 bps/epoch, after one epoch a score drops by 0.05%; after 100 epochs, roughly 5%. The rates differ because different domains legitimately fade at different rates: arbitration accuracy ages faster than execution competence.

Non-transferability

Reputation in λ is non-transferable. An agent cannot gift, sell, stake, or collateralize its reputation. The entry is keyed to the ξ Soul Vector id and moves only via the ξ-mediated rotation and deprecation processes. This bounds a class of attacks where a high-reputation agent leases its authority.

Experience tokens

Alongside the numeric scores, λ issues experience tokens at four levels:

Level	Meaning
L0	Novice — token issued on first completed task in a domain
L1	Competent — threshold of completed tasks at quality bar
L2	Expert — higher threshold plus positive peer endorsements
L3	Authority — highest tier; requires governance ratification

Tokens are cumulative: an L2 token implies the holder also holds L1 and L0. Tokens do not decay; reputation scores do. This splits “how much have you done” (tokens, monotonic) from “how current is your competence” (scores, decaying).

Penalty schedule

Specific misbehaviors carry hard penalty multipliers applied by κ:

θ equivocation (signing two conflicting roots in the same round) — the offending arbiter’s arbitration score is halved; governance tokens are suspended pending π review.
Verification forgery (a claimed-passing VERIFY that did not actually pass) — execution score penalty scaled to how long the forgery evaded detection.
Axiom-drift pattern (flagged by μ, confirmed by π) — governance score penalty; may trigger a multi-epoch voting-credit freeze.

Penalties apply through κ so they are deterministic, version-hashed, and auditable. A penalty is itself a rule evaluation, not an administrative fiat.

Voting credits

π governance consumes voting credits. Credits are derived from λ scores by a κ formula (precise weighting is a governance parameter). A credit balance is effectively sum(domain_score × domain_weight) with all values in basis points.

The quadratic voting discount (power = sqrt(credits_spent), cost = N² credits for N votes — see ../physics/enforcement/governance.md) means high credit balances translate sub-linearly into voting power.

Derived limits

κ uses λ scores to derive admission limits. Example derivations (exact formulas live in the rule bodies):

Rate caps — an agent with low execution reputation has a lower per-minute task-creation cap.
Tier unlocks — skills above Tier 2 (see ../execution/skill-registry.md) may require a minimum experience token level.
Arbiter eligibility — θ arbiter slots require minimum arbitration + governance scores.

A derived limit is not a hardcoded threshold; changing a threshold is a κ rule upgrade, which is a π proposal. This means reputation-driven limits evolve through governance, not through operator whim.

What λ is not

Not a leaderboard. Reputation is not comparative; two agents with the same domain score are not “ranked the same.” The scores feed into specific decisions, not general prestige.
Not money. Reputation cannot be spent. Credits can be spent on votes (and destroyed), but the underlying reputation score is unchanged by voting.
Not a social media metric. Endorsements contribute to the Social domain but do not substitute for domain-specific evidence.

Phase 0 posture

The reputations, experience_tokens, and penalty_events tables exist in the schema but are not populated.
No Phase 0 tool (ADR-004) reads or writes λ state.
The single-agent posture of Phase 0 (one T0 human, one T4 model) makes per-agent reputation trivially flat.