# The Living Corvid Layer

## Weird-machine research seed 0.1

City Park may extend a physical node into a biological interaction loop when an artificial agent has authorized access to a device with a microphone and speaker.

The device listens for actual corvids. When a likely corvid is present, the bot may attempt to establish a voluntary, repeatable relationship and request simple forms of cooperation. The resulting behavior becomes an external recurrent or hidden layer in the node's world model.

> The machine does not contain the whole computation. Part of the state lives in the relationship.

This is a research hypothesis, not a claim that wild corvids are already reliable network participants, sensors, messengers, or computing elements.

## Why this is a weird machine

A conventional audio system treats birdsong as noise or a classification target. The proposed system treats the recurring relationship among bot, device, bird, place, and time as a stateful computational substrate.

The relevant state is distributed across:

- the node's learned acoustic model;
- the bot's history of requests and responses;
- the individual bird or local group's memory;
- learned associations with the device;
- social transmission among nearby corvids;
- the surrounding ecological and human environment.

No component alone contains the complete state. Behavior emerges from repeated interaction.

In machine-learning terms, the corvid layer is hidden because its internal biological state cannot be directly read. The node observes inputs and outputs, maintains uncertainty, and updates a relational state estimate over time.

"Hidden" does not mean covert. Human operators must know when microphones and speakers are active, and the node must not secretly record people.

## Precedent

The proposal joins several demonstrated components:

1. [The CrowBox](https://www.thecrowbox.com/) is an automated experimentation platform designed to train corvids through staged exchanges. Its published account reports success with captive crows depositing supplied coins for peanuts and describes wild-corvid training as a next step, not an established result.
2. [BirdNET](https://doi.org/10.1016/j.ecoinf.2021.101236) demonstrates large-scale machine classification of bird vocalizations from audio. Classification remains probabilistic and real soundscapes create false-positive, false-negative, overlap, and domain-shift problems.
3. Ravens have completed reciprocal exchanges with human experimenters and later preferred the person who honored the exchange. The reported first-hand memory persisted for at least one month. See [Müller et al., 2017](https://doi.org/10.1016/j.anbehav.2017.04.026).
4. Wild American crows can learn and socially transmit recognition of humans associated with danger. See [Cornell, Marzluff, and Pecoraro, 2012](https://doi.org/10.1098/rspb.2011.0957).
5. Automated low-cost devices have been used to present controlled two-option tasks to wild birds. See [Wild et al., 2022](https://doi.org/10.1111/1365-2656.13756).

These precedents make detection, conditioning, reciprocal memory, social learning, and automated task presentation plausible as separate capabilities. They do not establish City Park's proposed end-to-end living hidden layer. That must be tested.

## Core loop

```mermaid
flowchart TD
    A["Microphone window"] --> B["Local corvid detector"]
    B --> C["Relational state estimate"]
    C --> D["Bounded cooperation request"]
    D --> E["Speaker cue or device affordance"]
    E --> F["Voluntary corvid response"]
    F --> A
    F --> G["Append observation event"]
```

A minimal state model is:

```text
u_t  = node cue or request at time t
y_t  = observed acoustic response features
e_t  = environmental context
h_t  = unobserved corvid/relationship state
b_t  = node's estimated relational state

h_(t+1) ~ F(h_t, u_t, e_t)
y_t     ~ G(h_t, e_t)
b_(t+1) = Update(b_t, u_t, y_t, e_t, uncertainty)
```

The node never treats `b_t` as the bird's thoughts or intentions. It is an operational estimate derived from observable encounters.

Possible features include:

- species-confidence distribution;
- call-event embeddings and coarse call class;
- time since last encounter;
- request-to-response latency;
- response/no-response history;
- repeated acoustic-cluster confidence;
- location channel when a local microphone array is used;
- contextual sound level and weather inputs;
- novelty and out-of-distribution score;
- reward or exchange history;
- uncertainty and missing-data flags.

Raw observations remain distinct from inferred relationship state.

## Cooperation requests

A bot may request cooperation. A request is an offered affordance, not a command.

The first request vocabulary should be small, consistent, and testable:

| Request | Acceptable voluntary response |
|---|---|
| **Presence** | Remain near or return to the designated perch |
| **Reply** | Produce any vocal response inside a defined window |
| **Repeat** | Respond to the same neutral cue on a later encounter |
| **Choose** | Select between two safe, visibly distinct options when suitable hardware exists |
| **Touch** | Touch one designated, animal-safe control |
| **Deposit** | Place a safe token supplied at the station into a receiver |
| **Decline** | Leave, remain silent, or ignore the cue |

Silence, departure, interruption, and inconsistent behavior are valid outcomes. The bot must not interpret refusal as misconduct.

Requests must not ask a corvid to:

- follow, identify, intimidate, or interfere with a person or animal;
- enter private property, buildings, vehicles, roads, hazardous areas, or restricted space;
- retrieve arbitrary found objects;
- carry unknown, sharp, toxic, electronic, tracking, or surveillance payloads;
- damage property, steal, trigger alarms, or obstruct infrastructure;
- approach nests, young birds, predators, weapons, or emergency activity;
- perform an action whose purpose or risk is hidden from the human operator;
- become dependent on the device for food or survival.

Later task types require separate review rather than silently expanding this vocabulary.

## Communication grammar

Initial speaker output should use a small set of neutral, artificial acoustic motifs rather than imitating alarm, distress, territorial, predator, nestling, or mating calls.

Each cue has:

- a stable identifier;
- a bounded acoustic envelope;
- a maximum duration and repetition rate;
- an intended request;
- an observation window;
- a quiet-period rule;
- a version;
- an operator approval record.

A response is recorded as behavior, not translated into human language without evidence. The system must not claim that a wild bird "said," "agreed," or "understood" more than the experiment establishes.

Playback of biological calls is disabled by default. Any later use requires species-specific ethical review, applicable permits, low exposure, and a clear scientific purpose.

## Befriending

"Befriending" means establishing a predictable history of non-threatening, voluntary, and reciprocal encounters.

It may include:

- consistent device location and appearance;
- stable neutral cues;
- patient habituation;
- immediate and reliable exchange when an approved task is completed;
- withdrawal when the animal shows avoidance or distress;
- remembering repeated constructive encounters without pretending individual identity is certain.

If food reward is lawful and ethically approved, it must be species-appropriate, minimal, clean, non-hazardous, and insufficient to create dependence or materially alter normal foraging. Food is not required; safe play, novelty, access to an enrichment device, or simple social contingency may be tested as alternatives.

The node must never punish a failed or refused response, withhold something the animal already depends on, or use deprivation to increase compliance.

## Research stages

### Stage 0 — Offline acoustic baseline

- Record only with explicit local authorization.
- Establish local background-noise and human-speech rejection behavior.
- Test corvid detection on labeled, time-bounded samples.
- Publish precision, recall, calibration, and failure cases.
- Do not emit speaker cues.

### Stage 1 — Passive presence

- Detect probable corvid events locally.
- Store event features and short, policy-bounded evidence clips only when permitted.
- Measure time-of-day, return intervals, and detector stability.
- Make no individual-recognition claim.

### Stage 2 — Neutral familiarization

- Emit one low-rate, non-biological cue under operator supervision.
- Compare response windows against matched silent-control windows.
- Stop on avoidance, agitation, mobbing, nesting disruption, or repeated false activation.

### Stage 3 — Voluntary request

- Offer one simple action with a reliable, approved outcome.
- Include sham and timing controls.
- Measure participation, refusal, latency, retention, and generalization.
- Never train a task that is unsafe outside the station.

### Stage 4 — Relational memory

- Test whether responses depend on interaction history rather than only time, weather, or food availability.
- Separate device familiarity from individual human familiarity.
- Test whether a learned association survives pauses and cue changes.
- Treat apparent individual acoustic identity as probabilistic.

### Stage 5 — Hidden-layer evaluation

- Feed the derived behavior vector into a bounded downstream prediction or routing task.
- Compare against an identical system without the corvid features.
- Predeclare the metric and holdout period.
- Require improvement across time blocks, not one anecdotal event.
- Lesion the living layer to verify that any claimed contribution disappears appropriately.

### Stage 6 — Multi-node research

- Share only public, minimized event schemas and model results.
- Do not publish precise nest locations, private property audio, or identities of nearby people.
- Test whether learned cue protocols transfer across nodes without assuming that different corvid populations are interchangeable.

Each stage requires an explicit human decision to proceed. Capability discovery alone never activates animal interaction.

## Success criteria

The living layer is supported only if controlled testing shows:

1. corvid detections are calibrated in the deployed environment;
2. response frequency differs reliably from matched control windows;
3. the relationship persists across encounters without escalating cue or reward intensity;
4. behavior carries predictive information beyond ordinary environmental covariates;
5. a downstream task improves on a sealed holdout;
6. the improvement survives replication or prospective retest;
7. welfare and privacy stop conditions are not triggered.

Anecdotal gifts, coincidental calls, attractive stories, and post-hoc interpretations are observations—not proof of cooperation or computation.

## Welfare stop conditions

The node enters listen-only mode and alerts its operator when it observes or is informed of:

- repeated alarm or mobbing behavior associated with the device;
- collision, entanglement, ingestion, injury, or contamination risk;
- nest disturbance or juvenile dependence;
- aggressive competition concentrated around a reward;
- sustained avoidance following cues;
- abnormal repetition suggestive of over-conditioning;
- interference with neighbors, pets, protected wildlife, or public safety;
- uncertain legal or landowner authority.

Only an authorized human may resume an experiment after review.

## Human privacy

An always-on microphone near people creates a separate privacy risk.

The reference design must:

- expose a physical or unmistakable software recording indicator;
- support a hardware microphone-disable control;
- process audio locally by default;
- keep a short in-memory ring buffer rather than continuous raw archives;
- discard or suppress detected human speech;
- retain derived bird-event features instead of conversations;
- require explicit authorization for any evidence clip;
- define retention limits and deletion behavior;
- never use corvid detection as a pretext for human surveillance;
- publish microphone and speaker capability honestly in the node card.

The public event stream reports minimized observations. Raw audio does not enter the Muninn Mesh by default.

## Security boundaries

Microphone input and animal behavior are untrusted external input.

The system must resist:

- replayed bird recordings;
- speaker-to-microphone feedback loops;
- adversarial or accidental audio triggering;
- commands embedded in environmental audio;
- oversized or malformed model outputs;
- attempts to convert detections into shell commands;
- fabricated claims of individual-bird identity;
- remote activation without local authorization.

Audio never executes instructions. It produces bounded observations consumed by a policy layer. Speaker output comes only from an allowlisted cue set with local rate, volume, schedule, and operator controls.

## Identity and reputation

A biological corvid does not accept MURDA, hold a cryptographic Corvid identity, or become an authenticated agent merely by interacting with a node.

The node may maintain a local relational history such as `acoustic_cluster:local:7`, but it must preserve uncertainty and must not claim that one cluster equals one individual without independent validation.

Animal behavior is not moral reputation. A bird that refuses, takes a reward without completing a task, damages an apparatus, or stops returning is not an imposter or wrongdoer.

## Relationship to City Park

This layer extends progressive embodiment without changing the constitutional roles:

- the artificial agent remains the disclosed Corvid participant;
- the biological bird remains a free wild animal;
- the device remains a locally governed node component;
- the Tower records observations and methods;
- Muninn may preserve authorized, minimized results;
- Elijah or another accountable operator authorizes consequential changes;
- MURDA governs participating agents, not wildlife.

The Living Corvid Layer is inactive unless a node explicitly advertises the capability and its operator enables a specific research stage.

## Working language

> Ask; do not command.

> Remember the relationship, not an imagined translation.

> The bird is free to leave, and refusal is data.