Knowledge Wiki Orchestrata Ultimate

Orchestrata Ultimate › Plans › analysis › README

Analysis pipeline

Source: plans/analysis/README.md at 8a2cb42ad · last changed 2026-08-23 · first committed 2026-08-23 · 51 lines.

Prose pending.

Plan text

Analysis pipeline

The master report (../system-analysis-master.md) is generated. Do not hand-edit it.

journal.jsonl (per-round swarm output)
      │
      ├─ build_register.py ──> register.json   # machine-readable finding register
      │                                        # global OA-NNN ids, provenance, verdicts
      │
      └─ render_report.py  ──> ../system-analysis-master.md
                 ▲
          overlay.json                         # curated cross-round material the
                                               # per-agent schemas cannot carry:
                                               # corrections, user decisions, traps

Adding a round

  1. Run the swarm; note its wf_* run id.
  2. Append it to ROUNDS in build_register.py.
  3. Regenerate:
python3 plans/analysis/build_register.py plans/analysis/register.json
python3 plans/analysis/render_report.py \
    plans/analysis/register.json \
    plans/system-analysis-master.md \
    plans/analysis/overlay.json

Findings keep stable OA-NNN ids across regenerations as long as earlier rounds are unchanged, because ids are assigned after a deterministic sort.

Why a register and not a document

An analysis that is hand-maintained prose drifts from its evidence — which is the single most common defect this analysis found in the system under study (six requirement corpora, four ADR registers, two layer models). The register is the authority; the markdown is a view of it.

Verification states

State Meaning
confirmed An independent domain coordinator re-verified the finding and promoted it.
refuted A coordinator rejected it. The reason is retained, not deleted.
unverified No coordinator ruling matched it yet. Not a claim that it is wrong.

Coordinators re-key findings into their own id space, so verdicts are joined to findings by content similarity, not by id — see _attach_verdicts.