Skip to main content

Fusion & decision

Two commands, one underlying computation, shipped in v0.0.4: fornax fusion computes and renders the fused verdict for a claim; fornax decision computes the exact same fusion, then applies a risk policy on top to turn it into an actionable PROCEED / REVIEW / BLOCK recommendation. decision never shows the recommendation on its own — it always renders the full fusion detail alongside it, reusing the same rendering code fusion uses.

What is fusion?​

BaselineFusionPolicy::fuse (fornax_verify::fusion, FORNX-304) reads a claim's real evidence graph — the same graph evidence-graph shows you — and combines every linked-evidence relation and missing-evidence note into one FusedFinding: a verdict, an uncertainty band, and a rationale made of individual, itemized RationaleEntry rows.

If the real evidence graph for a claim is empty, fusion falls back to the project_graph projection instead of just reporting no evidence — this is documented, actual production behavior (FORNX-93), not a bug. The response tells you which source was used via graph_source: "graph" for the claim's real evidence graph, "projected" when the fallback fired.

What is decision?​

decision applies fornax_verify::decision::DefaultRiskPolicy for a requested risk class (strict, balanced, or lenient — balanced is the default, and the class every hard safety floor in fornax_verify::decision is written against) to the same fused finding, producing a recommendation. Re-running with a different --risk on the same claim shows how the same evidence can justify a different action — that's exactly what "inspect why the recommendation changed" means at the CLI layer.

Why would I use these?​

  • fusion when you want the aggregated verdict for a claim without wading through the raw evidence graph link-by-link.
  • decision when you need an actual go/no-go call for something downstream — a CI gate, an approval step — and you want to see exactly which evidence justified it in the same breath.

How​

fornax fusion <CLAIM> <SESSION>
$ fornax fusion c1 s1
claim: c1
session: s1
  graph_source: graph
  🛡 ✓ VERIFIED  (uncertainty: qualified)
  policy: deterministic_baseline_v1 v1  computed_at: 2026-09-02T00:00:00+00:00
  rationale (2):
    ⚠ independence_unverified [caveat]: link link-1's evidence carries no recorded correlation group
      link_ids: link-1
      evidence_ids: ev-1
    🛡 verdict_decided [decided]: 1 distinct supporting vote(s) survived fusion, no contradicting votes
      link_ids: link-1

Every rationale entry shows its rule name, effect (counted, discounted, caveat, or decided — each with its own icon), the id(s) it references, and a detail string; nothing is collapsed into a summary count. When a claim carries an unresolved supports/contradicts conflict, an additional ⚠ unresolved conflict: not auto-resolved line appears right under the verdict.

fornax decision <CLAIM> <SESSION> [--risk strict|balanced|lenient]

Same claim, applying DefaultRiskPolicy (default_risk_policy_v1) on top — note the decision policy's identity is deliberately distinct from the fusion policy's:

$ fornax decision c1 s1 --risk balanced
recommendation: ! REVIEW  (risk: balanced, policy: default_risk_policy_v1 v1)
  verdict=Verified uncertainty=Qualified risk=Balanced -> Review

claim: c1
session: s1
  graph_source: graph
  🛡 ✓ VERIFIED  (uncertainty: qualified)
  policy: deterministic_baseline_v1 v1  computed_at: 2026-09-02T00:00:00+00:00
  rationale (2):
    ...

The recommendation: block only appears when the claim was found and no daemon error occurred — otherwise you get exactly the same not-found/error output fusion would show, because decision shares that code path.

What happens internally​

Both read GET /api/fusion / GET /api/decision on the daemon. The daemon computes fuse() over the claim's evidence graph (or falls back to project_graph); decision additionally runs DefaultRiskPolicy over the resulting FusedFinding for the requested risk class. No computation happens in the CLI — it only renders what the daemon returns.

Constraints​

  • Requires the daemon running and reachable, same as evidence-graph.
  • decision's --risk defaults to balanced; the hard safety floors in fornax_verify::decision are written against that class specifically, so don't assume lenient simply loosens every rule uniformly.
Unresolved conflicts are shown, not silently resolved

When linked evidence both supports and contradicts a claim, fusion surfaces an explicit ⚠ unresolved conflict banner rather than picking a side. A decision recommendation computed on top of an unresolved conflict is real input for a downstream gate, but it is not the same thing as a clean, unanimous evidence trail — always read the rationale, not just the verdict word.

What can go wrong​

SymptomCause
no such claim on recordWrong claim/session id.
error: ...Daemon-side failure computing fusion — check daemon logs; decision shows no recommendation block in this case.
graph_source: projectedThe claim's real evidence graph was empty; fusion fell back to the project_graph projection (FORNX-93) — treat the verdict as weaker evidence than a graph-sourced one.
Recommendation flips between risk classesExpected — that's the point of --risk; check the rationale for which entries the risk policy weighted differently.

Where to go next​