Model Routing Simulator

Static MOCK presets · deterministic policy · no live model calls · recommendations are advisory
DCP READ-ONLY: CONNECTED DEVELOPMENT / UI PREVIEW DCP BATCH 5/6 SOAK IN PROGRESS TELEMETRY: UNKNOWN COMMAND EXECUTION LOCKED
READ-ONLY SIMULATOR · Deterministic policy evaluation over a STATIC development registry. No live model calls, no routing changes, no execution. Recommendations are advisory; authority stays with the owner.

Task Input (preset scenario)

3 · Medium multi-file feature

Task typeMEDIUM_FEATURE
Agent roleDEVELOPER
Risk levelR2
ComplexityMEDIUM
Context sizeMEDIUM
VisionNONE
Repo writeyes
Long tool loopyes
Architecture reasoningno
Code generationyes
Security reviewno
Cost sensitivityMEDIUM
Latency sensitivityLOW
Preset expectation: DeepSeek preferred until GLM medium-task evidence exists.

Recommended Agent

DEVELOPER

Static default policy: coding evidence dominates · multi-file/long-loop evidence required for larger tasks

Primary Model

DeepSeek V4.1 Flash deepseek-v4.1-flash VERIFIED

Fallback Model

GLM-5.3-Flash(x) glm-5.3-flashx VERIFIED_FOR_MEDIUM_MULTI_FILE

GLM-5.3-Flash(x) (glm-5.3-flashx)

Status: VERIFIED_FOR_MEDIUM_MULTI_FILE · last verified: 2026-10-05 · provenance: STATIC_CONFIG

Verified task classes: GIT_AUDIT, README_UPDATE, SMALL_PATCH, MEDIUM_FEATURE · Unverified: LARGE_FEATURE, SECURITY_REVIEW, UI_VISION_REVIEW

  • read-only Git inventory PASS [REAL_OBSERVATION]
  • untracked-file inspection PASS [REAL_OBSERVATION]
  • small repo hygiene coding patch PASS [REAL_OBSERVATION]
  • 531-test verification PASS [REAL_OBSERVATION]
  • medium multi-file task PASS (MODEL_ROUTING_POLICY_SIMULATOR_V1: ~9 files, 286/286 portal tests, 39 new) [REAL_OBSERVATION]
  • large feature / long agent loop work NOT YET VERIFIED; one medium task is not enough evidence [REAL_OBSERVATION]

Evidence — primary model

DeepSeek V4.1 Flash (deepseek-v4.1-flash)

Status: VERIFIED · last verified: 2026-10-05 · provenance: STATIC_CONFIG

Verified task classes: MEDIUM_FEATURE, LARGE_FEATURE, SMALL_PATCH, CODE_REVIEW, GIT_AUDIT, README_UPDATE · Unverified: UI_VISION_REVIEW

  • heavy coding proven [REAL_OBSERVATION]
  • long agent loops proven [REAL_OBSERVATION]
  • multi-file repo work proven [REAL_OBSERVATION]
  • testing/tool loop proven [REAL_OBSERVATION]

SCORE (transparent, advisory)

The weighted score is advisory only. Hard blockers always override the score. UNKNOWN never counts as positive evidence.
DimensionValueStatic weight
Capability Match0.60weight 0.30
Evidence Strength1.00weight 0.25
Task Complexity Fit1.00weight 0.15
Risk Fit1.00weight 0.10
Context Fit1.00weight 0.08
Cost Fit0.72weight 0.07
Latency Fit0.80weight 0.03
Vision Fit0.50weight 0.02
TOTAL0.844

Routing Reasons

WHY THIS MODEL?

  • STATUS — DeepSeek V4.1 Flash status = VERIFIED
  • EVIDENCE — verified task classes: MEDIUM_FEATURE, LARGE_FEATURE, SMALL_PATCH, CODE_REVIEW, GIT_AUDIT, README_UPDATE
  • TASK_MATCH — direct verified evidence for task class MEDIUM_FEATURE
  • MULTI_FILE — proven multi-file/long-loop evidence for repo implementation
  • NOT_PREFERRED_glm-5.3-flashx — GLM-5.3-Flash(x) scored 0.79 vs 0.844 (gap 0.054)
  • EXCLUDED_kimi-k3 — Kimi K3 excluded: unavailable (upstream HTTP 503 on trivial prompt; provider retry failed)
  • FALLBACK — fallback GLM-5.3-Flash(x) (glm-5.3-flashx): VERIFIED_FOR_MEDIUM_MULTI_FILE, next-best eligible score 0.79

Hard Blockers

  • none

WHAT IS STILL UNKNOWN?

  • DeepSeek V4.1 Flash: vision support UNKNOWN
  • GLM-5.3-Flash(x): vision support UNKNOWN
  • Kimi K3: vision support UNKNOWN
  • Kimi K3: cost multiplier UNKNOWN

WHAT WOULD CHANGE THE DECISION?

  • If a cheaper model accumulates verified evidence for this task class, cost preference may flip the primary
  • If Kimi K3 becomes reachable AND accumulates verified evidence in our environment, it re-enters the candidate pool
  • If this task is R3/R4, owner approval is required regardless of model choice

Diversity-of-Failure

DIVERSITY_PREFERRED QA reviewer should prefer GLM-5.3-Flash(x) (glm-5.3-flashx) — different failure modes from DeepSeek V4.1 Flash; evidence quality preserved

MODEL ROUTING != EXECUTION AUTHORITY

Selecting a model never grants sudo, Docker mutation, production write, DB mutation, deploy, network change or approval power. Authority is governed by risk policy, owner approval and DCP controls.

Model Status (static development registry)

Kimi K3 is UNAVAILABLE (HTTP 503 upstream provider failure) and is never displayed as selectable or recommended.
ModelStatusProven task classesKnown failuresLast verification
DeepSeek V4.1 FlashVERIFIEDMEDIUM_FEATURE, LARGE_FEATURE, SMALL_PATCH, CODE_REVIEW, GIT_AUDIT, README_UPDATE—2026-10-05
GLM-5.3-Flash(x)VERIFIED_FOR_MEDIUM_MULTI_FILEGIT_AUDIT, README_UPDATE, SMALL_PATCH, MEDIUM_FEATURE—2026-10-05
Kimi K3UNAVAILABLEupstream HTTP 503 on trivial prompt; provider retry failed