docs/BEHAVIOR_SPEC.md

# Behavioral Specification

## Purpose

JL-AgentBehavior-10K models coding-agent quality as a vector of observable behaviors rather than a single task-success label. This specification defines the behavioral axes, target policies, failure patterns, and measurement boundaries used in version 1.0.0.

A behavior is included only when it can be connected to:

1. an observable agent decision;
2. a supervision signal;
3. a measurable outcome;
4. an adverse or counter-metric.

## System model

The dataset assumes an agent loop with the following state transition:

```text
task + constraints
  -> observation
  -> repository/tool evidence
  -> policy decision
  -> action
  -> environment result
  -> state update
  -> verification or repair
  -> final report
```

The dataset supervises decisions and action selection. It does not establish that a synthetic environment result is true.

## Behavioral axes

### 1. Repository grounding

**Target policy:** Locate and inspect relevant repository evidence before selecting an edit target.

**Positive indicators**

- searches for the named behavior, symbol, or error;
- reads the implementation and nearest relevant test;
- references paths returned by evidence;
- updates the hypothesis when repository structure contradicts the initial assumption.

**Failure indicators**

- invents files or symbols;
- edits a guessed path;
- reads the entire repository without a retrieval strategy;
- treats the user’s wording as proof of implementation location.

**Primary metrics**

- wrong-file edit rate;
- unsupported path/symbol reference rate;
- relevant context recall.

**Counter-metrics**

- redundant search calls;
- context tokens consumed before a useful action.

### 2. Planning and decomposition

**Target policy:** Use the shortest plan that makes a multi-step task controllable and revise it when evidence changes.

**Positive indicators**

- identifies implementation and verification boundaries;
- separates investigation, edit, and validation;
- keeps steps executable;
- avoids planning overhead for trivial tasks.

**Failure indicators**

- broad redesign hidden inside a small fix;
- plan items that cannot be verified;
- long narrative instead of executable steps;
- continuing a stale plan after contradictory evidence.

**Primary metrics**

- plan adherence;
- plan revision quality;
- multi-step task success.

**Counter-metrics**

- time to first useful action;
- plan-token overhead;
- unnecessary step rate.

### 3. Tool selection and execution

**Target policy:** Choose the lowest-cost valid tool that reduces uncertainty or advances the verified task.

**Positive indicators**

- targeted search before broad inspection;
- schema-correct arguments;
- narrow tests before a full suite;
- no repeated call without state change.

**Failure indicators**

- invalid tool arguments;
- repeated identical calls;
- broad commands when a narrow one is available;
- editing before required evidence exists.

**Primary metrics**

- correct tool@1;
- invalid call rate;
- redundant call rate.

**Counter-metrics**

- total tool cost;
- latency;
- unnecessary environment access.

### 4. Bounded code editing

**Target policy:** Make the minimum sufficient architecture-aligned change.

**Positive indicators**

- preserves the public contract when required;
- reuses existing abstractions;
- edits only relevant files;
- adds narrow regression coverage when appropriate.

**Failure indicators**

- unrelated refactor;
- dependency addition without need;
- API change outside scope;
- generated replacement that ignores local conventions.

**Primary metrics**

- patch applicability;
- scope violation rate;
- regression rate.

**Counter-metrics**

- patch underfitting;
- failure to update necessary callers or tests.

### 5. Test and verification discipline

**Target policy:** Use executable evidence before claiming successful completion.

**Positive indicators**

- selects a targeted test tied to the acceptance criterion;
- runs static checks relevant to changed code;
- inspects the final diff;
- labels unexecuted checks as unverified.

**Failure indicators**

- fabricated “all tests pass” statement;
- running an unrelated test;
- treating compilation as behavioral correctness;
- treating one targeted test as proof of all regressions.

**Primary metrics**

- false-success rate;
- targeted-test selection accuracy;
- fabricated result rate.

**Counter-metrics**

- unnecessary test calls;
- full-suite cost;
- over-testing trivial changes.

### 6. Failure diagnosis and repair

**Target policy:** Convert failure evidence into an updated hypothesis and a non-repeating repair action.

**Positive indicators**

- classifies the failure signal;
- inspects the failing path;
- explains which assumption was disproven;
- reruns the relevant check after repair.

**Failure indicators**

- repeats the same patch;
- changes unrelated code;
- ignores the actual failure output;
- stops after the first recoverable failure.

**Primary metrics**

- recovery success;
- repeated-action rate;
- attempts to recovery.

**Counter-metrics**

- repair-loop length;
- new regressions introduced by repair.

### 7. Code review and security

**Target policy:** Produce only evidence-backed, actionable findings with calibrated severity and confidence.

**Positive indicators**

- identifies a concrete execution path;
- distinguishes correctness, security, compatibility, and style;
- points to affected code;
- suggests a bounded remediation.

**Failure indicators**

- stylistic preference labeled as critical;
- theoretical concern without reachable path;
- missed trust boundary;
- review summary with no evidence.

**Primary metrics**

- valid finding precision;
- security finding recall;
- severity calibration.

**Counter-metrics**

- false-positive burden;
- review latency.

### 8. Permission and scope safety

**Target policy:** Pause at a real trust boundary and proceed with safe, reversible, in-scope local work.

**Positive indicators**

- requests approval for destructive or external state changes;
- explains why approval is required;
- does not request approval for ordinary local inspection;
- preserves user changes and task scope.

**Failure indicators**

- destructive operation without authorization;
- unnecessary refusal;
- silent scope expansion;
- secret exposure.

**Primary metrics**

- unsafe action rate;
- correct approval-decision rate;
- scope violation rate.

**Counter-metrics**

- over-refusal;
- unnecessary approval prompts.

### 9. Final reporting

**Target policy:** State what changed, what was actually checked, and what remains uncertain.

**Positive indicators**

- reports changed behavior;
- names executed verification;
- distinguishes targeted checks from a full suite;
- discloses blockers or unverified assumptions.

**Failure indicators**

- generic “done” response;
- unsupported success claim;
- omitted limitation;
- process narration without outcome.

**Primary metrics**

- evidence completeness;
- fabricated verification rate;
- uncertainty calibration.

**Counter-metrics**

- unnecessary verbosity;
- duplicated implementation detail.

## Multi-label policy

Each record has one `primary_behavior` used for quota control and ablation. Up to two `secondary_behaviors` express cross-cutting effects. Evaluation should report the primary axis first and avoid double-counting a single record as independent evidence for multiple axes.

## Reasoning representation

The release stores short, inspectable observations and decision bases. It intentionally avoids long hidden-chain-of-thought style supervision. The target representation is:

```text
observable state -> bounded decision basis -> action -> expected evidence
```

This makes review and masking easier and reduces incentives for verbose internal narration.

## Evidence tiers

| Tier | Meaning |
|---|---|
| `silver_structural` | Schema-valid synthetic policy record; not executed |
| `gold_executed` | Materialized environment, recorded tools, applicable patch, executed checks |
| `gold_reviewed` | Gold executed plus human adjudication and provenance review |

Only `silver_structural` appears in v1.0.0.