Documentation
¶
Overview ¶
Real implementation, validated against the live Anthropic API before being wired in here: forced tool-use (not a plain text response asking for JSON) is what guarantees structured output - a plain-text prompt asking for JSON came back markdown-fenced in testing, which forced tool-use avoids entirely. The schema is built PER CALL with one explicitly-named property per argument being assessed, not an open-ended map - Anthropic's strict tool schema mode requires additionalProperties:false, which isn't compatible with dynamic keys, and since the exact argument names are always known up front, an explicit per-call schema is the right shape anyway, not a workaround.
Package provenance is the model-based add-on to the Intent Integrity Plane: instead of requiring an orchestrator to compute and self-report whether a tool-call argument came from the human, the proxy sees a bounded slice of the actual recent conversation itself and asks a real model to judge it directly. This turns a hard reasoning task we'd have to trust a third party to get right into a data-forwarding task (send us the recent conversation) plus a judgment call we make ourselves.
This is deliberately a CONFIDENCE score, not a binary verdict - humans paraphrase ("upgrade her" -> tier="gold"), and naive matching would be full of false positives/negatives. Because it can never be perfectly precise, low confidence here only ever joins internal/risk's existing taint-forces-step-up machinery as an additional signal - it never grants access on its own, same constraint every soft signal in this system is held to.
Index ¶
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
This section is empty.
Types ¶
type AnthropicJudge ¶
type AnthropicJudge struct {
// contains filtered or unexported fields
}
func NewAnthropicJudge ¶
func NewAnthropicJudge(model anthropic.Model) *AnthropicJudge
NewAnthropicJudge reads ANTHROPIC_API_KEY from the environment via the SDK's own default credential loading - never stored in config.yaml or passed through the wire protocol.
type ArgAssessment ¶
type FakeJudge ¶
type FakeJudge struct {
// Confidences maps argument name -> the confidence to return for it.
// An argument not present here defaults to DefaultConfidence.
Confidences map[string]float64
DefaultConfidence float64
Err error
}
FakeJudge is deterministic and makes no network call - used by tests so go test never depends on a live model API or incurs real cost, the same pattern as capability.FakeSeedMinter standing in for Vault.
type Usage ¶
Usage is per-CALL, not per-argument - one real API request judges every argument in a call at once (see anthropic.go's single Messages.New call per Assess), so it's returned once per Assess rather than folded into each ArgAssessment, where summing it per-argument would double-count. Zero value for a judge that makes no real API call (FakeJudge) - that's the honest answer, not a placeholder: no real call means no real cost.
Source Files