8440ddfecb
Makes the offline proposer runnable end-to-end.
- BuildProposerInput (proposer_input.go): non-test engine->hazards path. The
PatternMatch->Hazard converter is lifted out of the GT test files into
production scope so both the tests and the CLI share one pipeline.
- iace-audit propose <narrative.json> [<ground-truth.json>]: detect candidates ->
GT-screen survivors (when a ground truth is given) -> judge (HeuristicJudge by
default, LLMJudge over ollama when IACE_PROPOSE_LLM=1) -> write the human-review
queue to audit-reports/proposals.{md,json}. Propose-only.
Smoke run on a dishwasher narrative: 32 fired -> 3 candidates -> queue with a
confident duplicate, a confident distinct, and one punted to the LLM judge; GT
wall recall-safe. Live qwen is opt-in via env; the heuristic default keeps the
tool runnable (and CI deterministic) without a model. Proposal types 2-4
(foreign-framing gates, vocab->tag, coverage blind spots) remain for slice 4.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
142 lines
4.1 KiB
Go
142 lines
4.1 KiB
Go
package main
|
|
|
|
import (
|
|
"context"
|
|
"encoding/json"
|
|
"fmt"
|
|
"os"
|
|
"strconv"
|
|
|
|
"github.com/breakpilot/ai-compliance-sdk/internal/iace"
|
|
"github.com/breakpilot/ai-compliance-sdk/internal/llm"
|
|
)
|
|
|
|
type narrativeInput struct {
|
|
MachineType string `json:"machine_type"`
|
|
Narrative string `json:"narrative"`
|
|
MachineTypes []string `json:"machine_types,omitempty"`
|
|
}
|
|
|
|
// cmdPropose — Method P: offline dedup-candidate proposer.
|
|
//
|
|
// iace-audit propose <narrative.json> [<ground-truth.json>]
|
|
//
|
|
// Detect near-duplicate patterns, screen survivors against a ground truth (if
|
|
// given), judge them (heuristic by default, LLM when enabled), and write the
|
|
// human-review queue to audit-reports/proposals.{md,json}. Propose-only — it
|
|
// writes a report and never mutates the pattern library.
|
|
//
|
|
// Env:
|
|
//
|
|
// IACE_PROPOSE_THRESHOLD candidate score threshold (default 0.30)
|
|
// IACE_PROPOSE_LLM=1 use the offline LLM judge instead of the heuristic
|
|
// OLLAMA_URL ollama base URL (default http://localhost:11434)
|
|
// SELF_HOSTED_LLM_MODEL model name (default qwen2.5:32b-instruct)
|
|
func cmdPropose(args []string) {
|
|
if len(args) < 1 {
|
|
fmt.Fprintln(os.Stderr, "propose: usage: iace-audit propose <narrative.json> [<ground-truth.json>]")
|
|
os.Exit(2)
|
|
}
|
|
|
|
var in narrativeInput
|
|
must(readJSONFile(args[0], &in))
|
|
if in.Narrative == "" {
|
|
fmt.Fprintln(os.Stderr, "propose: narrative is empty")
|
|
os.Exit(2)
|
|
}
|
|
|
|
var gt *iace.GroundTruth
|
|
if len(args) >= 2 {
|
|
var g iace.GroundTruth
|
|
must(readJSONFile(args[1], &g))
|
|
gt = &g
|
|
}
|
|
|
|
threshold := envFloat("IACE_PROPOSE_THRESHOLD", 0.30)
|
|
hazards, mits, fired := iace.BuildProposerInput(in.Narrative, in.MachineType, in.MachineTypes)
|
|
candidates := iace.FindDedupCandidates(fired, threshold)
|
|
|
|
byID := make(map[string]iace.PatternMatch, len(fired))
|
|
for _, pm := range fired {
|
|
byID[pm.PatternID] = pm
|
|
}
|
|
|
|
judge := selectJudge(in.MachineType)
|
|
ctx := context.Background()
|
|
|
|
var proposals []iace.JudgedProposal
|
|
blocked := 0
|
|
for _, c := range candidates {
|
|
var sr iace.ScreenResult
|
|
if gt != nil {
|
|
sr = iace.ScreenSupersession(gt, hazards, mits, c.KeepHazardName, c.DropName)
|
|
if sr.RecallAfter < sr.RecallBefore || sr.DistinctGT {
|
|
blocked++
|
|
continue
|
|
}
|
|
}
|
|
v, conf, rat := judge.Judge(ctx, c, byID[c.KeepPattern], byID[c.DropPattern])
|
|
proposals = append(proposals, iace.JudgedProposal{
|
|
Candidate: c, Screen: sr, Verdict: v, Confidence: conf, Rationale: rat, Judge: judge.Name(),
|
|
})
|
|
}
|
|
|
|
writeText("audit-reports/proposals.md", iace.RenderProposalQueue(in.MachineType, proposals))
|
|
writeJSON("audit-reports/proposals.json", proposals)
|
|
|
|
printSummary("Method P — Dedup Proposer ("+judge.Name()+")", map[string]int{
|
|
"fired_patterns": len(fired),
|
|
"candidates": len(candidates),
|
|
"in_queue": len(proposals),
|
|
"gt_blocked": blocked,
|
|
})
|
|
if gt == nil {
|
|
fmt.Fprintln(os.Stderr, "note: no ground truth provided — GT wall NOT applied (candidates not recall-screened)")
|
|
}
|
|
}
|
|
|
|
func selectJudge(machineClass string) iace.CandidateJudge {
|
|
if os.Getenv("IACE_PROPOSE_LLM") != "1" {
|
|
return iace.HeuristicJudge{}
|
|
}
|
|
base := envStr("OLLAMA_URL", "http://localhost:11434")
|
|
model := envStr("SELF_HOSTED_LLM_MODEL", "qwen2.5:32b-instruct")
|
|
reg := llm.NewProviderRegistry("ollama", "")
|
|
reg.Register(llm.NewOllamaAdapter(base, model))
|
|
fmt.Printf("using LLM judge (ollama %s, model %s)\n", base, model)
|
|
return iace.LLMJudge{Completer: iace.NewRegistryCompleter(reg, model), MachineClass: machineClass}
|
|
}
|
|
|
|
func readJSONFile(path string, v any) error {
|
|
raw, err := os.ReadFile(path)
|
|
if err != nil {
|
|
return err
|
|
}
|
|
return json.Unmarshal(raw, v)
|
|
}
|
|
|
|
func writeText(path, content string) {
|
|
_ = os.MkdirAll("audit-reports", 0o755)
|
|
if err := os.WriteFile(path, []byte(content), 0o644); err != nil {
|
|
fmt.Fprintln(os.Stderr, "warn: could not write", path, err)
|
|
return
|
|
}
|
|
fmt.Println("→ wrote", path)
|
|
}
|
|
|
|
func envStr(key, def string) string {
|
|
if v := os.Getenv(key); v != "" {
|
|
return v
|
|
}
|
|
return def
|
|
}
|
|
|
|
func envFloat(key string, def float64) float64 {
|
|
if v := os.Getenv(key); v != "" {
|
|
if f, err := strconv.ParseFloat(v, 64); err == nil {
|
|
return f
|
|
}
|
|
}
|
|
return def
|
|
}
|