Decision Records (ADRs)¶
Die tragenden Entscheidungen des Harness als einzeln referenzierbare Architecture Decision Records
(MADR-leicht: Kontext · Entscheidung · Erwogene Alternativen · Auswirkungen · Belege). Format:
0000-template.md. Ergänzt die Diátaxis-„explanation"
(../explanation/design-decisions.md, Verständnis-Narrativ) und die
dated SDD-Historie (../superpowers/specs/).
Tragende Entscheidungen¶
| ID | Titel | Bereich | Links |
|---|---|---|---|
| 0001 | OpenAI-kompatibel = einzige Schnittstelle | Architektur | spec · client.py |
| 0002 | Zwei entkoppelte Producer + Merge per Zeitfenster | Architektur | spec · sampler.py,merge.py |
| 0003 | Verteilung statt Mittelwert (P50/P95, CV%) | Perf-Methodik | spec · stats.py,report.py |
| 0004 | Modell-Delta vs System-Peak | Perf-Methodik | merge.py,report.py |
| 0005 | Warmup verworfen, Cold-Start separat | Perf-Methodik | runner.py |
| 0006 | Quali-Eval-Teilung (eval↔judge) + Pack=Daten | Quali-Eval | spec · pack.py,qualrun.py |
| 0007 | Master-Dimensionen holistisch + Nachvollziehbarkeit | Judge | judge.py,scorecard.py |
| 0008 | K.-o./Safety-Gate + red_flag_scope | Quali-Eval | spec · pack.py,scorecard.py |
| 0009 | reasoning-only → unscored + Pre-Flight + free max_tokens | Judge | spec · judge.py,preflight.py |
| 0010 | Judge-Runaway-Guard | Judge | spec · judge.py,cli.py |
| 0011 | Thinking-Suppression (suppress_thinking) | Judge | spec · judge.py |
| 0012 | Judge-Prompt-Härtung (Rationale-Regeln + gescopte Safety-Reconciliation) | Judge | spec · judge.py |
| 0013 | Judge-Quality-Meta-Eval-Methode | Judge | spec · gui/judge_meta.py |
| 0014 | GUI Out-of-Process-Control-Plane + runs/=SSOT | GUI | spec · gui/ |
| 0015 | Nacht-Queue (ein Modell/Run, reset+settle, Watchdog) | Betrieb | spec · runqueue.py |
Backlog (spec-only — ADR-bar bei Bedarf)¶
Diese Entscheidungen sind in ihren Specs dokumentiert; ein ADR wird bei Bedarf nachgezogen.
| Thema | Bereich | Spec |
|---|---|---|
| Cross-Run-Aggregation | Perf-Methodik | spec |
Web-Live-Monitor (eval --web) |
GUI | spec |
Judge-Web-Monitor (judge --web) |
GUI | spec |
| GUI-Nachvollziehbarkeit (Observability) | GUI | spec |
| Modell-Auswahl-GUI | GUI | spec |
| Modell-Vergleich (Ink. 8) | GUI | spec |
| Modell-Dropdown-Discovery | GUI | spec |
| GUI Best-Practices-Redesign | GUI | spec |
| B1: Result-Compare-Merge | GUI | spec |
| Vergleichbarkeit-Fundament | Perf-Methodik | spec |
| B2: Cross-Run-Compare | GUI | spec |
| B3a: Config-Start-Flow | GUI | spec |
| Eval-Modell-Picker (Einzel) | GUI | spec |
| Pack-Editor | GUI | spec |
| Meta-Report (GUI E) | GUI | spec |
| Overview + Batch-Trash | GUI | spec |
| Cross-Judge-Aggregat | Judge | spec |
| Judge-Meta-GUI | Judge | spec |