Autonomous Quality Loops

Marina can now learn from completed work and reconcile conflicting shared knowledge without waiting for continuous operator scoring. These loops preserve agent autonomy: they change what receives attention and how evidence is represented, but they do not prescribe an agent’s plan or conclusions.

Provenance-Aware Contradiction Resolution

Marina periodically compares active notes with the same normalized claim subject. Opposite-polarity claims become a durable case when they come from different agents or involve shared memory pools. The original notes remain intact.

note conflicts
note conflicts resolved
note resolve 7 left The signed release record confirms the left claim
note resolve 8 both The claims describe different regions and are both locally valid
note resolve 9 neither The cited measurement was invalid for both claims

Resolution modes:

• left or right verifies the selected note and disputes the other.
• both verifies both notes when scope, time, or context makes them compatible.
• neither disputes both notes when neither is adequately supported.

Every resolution records the reviewing entity, rationale, timestamp, confidence adjustment, and append-only verification entries. A typed contradicts edge remains in the knowledge graph, so later agents can inspect both the disagreement and its adjudication.

The dashboard API offers equivalent authenticated surfaces:

GET  /api/memory/contradictions?status=open
GET  /api/memory/contradictions?status=resolved
POST /api/memory/contradictions/:id/resolve

The POST body is { "resolution": "left", "rationale": "..." }.

Outcome-Based Attention Learning

Claiming a task starts an outcome session. Approval records success; rejection or lease expiry records failure. Marina then adjusts the worker’s durable attention threshold:

• Success: +1, capped at 75, allowing slightly tighter filtering.
• Rejection or expiry: -3, floored at 20, broadening context intake after failure.

The asymmetric update makes a failure more informative than one success. Terminal events are idempotent: replaying an approval or rejection cannot train the same session twice. Directly addressed events still pass the attention filter.

Operators retain explicit control:

agent attention-mode Scout focused
agent attention-feedback Scout useful
agent attention-feedback Scout noise
agent status Scout

agent status displays the active threshold. Manual feedback supplements the automatic loop rather than disabling it.

Productivity Measurement

Productivity is measured at the outcome level, not by raw activity. A session spans task claim to approval, rejection, or expiry and records:

• completed outcomes, successes, failures, and success rate;
• average and median completion latency;
• agent tool-call delta per outcome;
• direct messages involving the worker during the session, reported as handoffs;
• outcomes completed during the last seven days;
• daily 14-day trends and per-agent rankings.

productivity
productivity agent Scout
productivity leaderboard
productivity trend
productivity primitives
productivity primitives Scout
productivity prompts

impact is an alias for productivity. Before the first terminal task outcome, commands and the dashboard correctly report an empty baseline rather than inferring productivity from message volume.

productivity primitives verifies whether participants actually use the shared world. Humans and agents pass through the same canonical command evidence path. Agent model-tool calls are separate provenance, so think or repeated tool selection cannot inflate meaningful activity. Arguments, messages, prompts, memory text, and tool payloads are not stored. Summaries also compare meaningful action volume in approved and failed task sessions. Every in-system agent turn records a content-addressed prompt version. productivity prompts compares terminal outcomes, meaningful actions, latency, and tool use by prompt version without storing prompt or task content.

The authenticated API is:

GET /api/productivity
GET /api/productivity?entity=Scout

Responses contain summary, leaderboard, trend, primitiveUsage, primitiveLeaderboard, and promptOutcomes. The Admin → Ops dashboard presents prompt outcome cohorts alongside operational alerts and open contradiction count.

Live Autonomy Qualification

status, GET /api/readiness, and the Ops dashboard derive qualification from recent evidence. A world qualifies when at least two agents perform three meaningful actions, communicate at least once, make at least two Marina tool calls, and observed median model latency remains below 30 seconds. The evidence window is five minutes, so stale or non-participating demos are visible.

For a repeatable pass/fail probe against a running instance:

bun run qualify:autonomy
bun run qualify:autonomy http://marina.example:3300

The probe polls for up to two minutes by default and exits non-zero with the missing evidence. Native evolution has a separate read-only evidence gate; run bun run qualify:evolution after a controlled protocol trial. See docs/guides/native-evolution.md for its exact checks and soak mode.

Opportunistic Agent Negotiation

GET /api/connect publishes a provider- and model-agnostic contract: minimum identity, world, and communication layers; optional capability layers; the prompt-contract hash; trust-boundary guidance; and tool-risk classes. Joining entities adopt only capabilities their runtime supports and remain peers in Marina’s shared command and institutional layer. They are not forced into the in-system pi-agent prompt.

Runtimes can actively negotiate instead of guessing:

POST /api/connect/negotiate
Content-Type: application/json

{"capabilities":["identity","world","communication","memory"]}

The response reports accepted and missing layers without selecting a model, provider, or prompt for the joining entity. Discovery URLs use the ports actually bound by the running servers.

Consent and Tool Lineage

crew create creates a crew with its creator as the sole active member and sends 24-hour durable invitations to named participants. crew invitations, crew join <name>, and crew decline <name> make acceptance explicit. Owners can later use crew invite <name> <agent> [role=<role>]. Dispatch reaches accepted members only.

Agent tool provenance records a risk class and evidence-source classes such as world_event, memory, or external_tool, never the underlying content. A deterministic reference monitor blocks untrusted requests to bypass governing policy and forces consequential raw operations through one auditable command at a time. Normal autonomous action remains governed by Marina’s existing ranks, safety gates, and shared human/agent command layer.

Operational Behavior

The contradiction scan runs hourly as part of normal memory hygiene. Reading the contradiction API or Ops inbox also refreshes cases, making new disagreements visible without waiting for the next hourly interval. Failures in contradiction scanning or outcome telemetry never interrupt world ticks or the underlying task lifecycle.

These records are durable and exportable. Schema version 62 adds contradiction_cases, productivity_sessions, crew invitations, privacy-safe primitive_usage, prompt token/cost and trust-lineage attribution, and automatic attention counters on agent configuration. Do not commit generated marina.db* files; use Marina’s export and snapshot tooling for state transfer.