Degradation Warnings¶

G6 is designed to run in constrained environments. When a component, backend, or external dependency is unavailable, the system operates in degraded mode rather than failing silently.

What triggers degradation¶

How to detect degraded mode¶

In workflow reports¶

The report's Safety Status section includes:

## Safety Status

| Check | Status |
|-------|--------|
| LLM backend active | DEGRADED (no backend configured) |
| Resource bounds respected | PASS |

In API responses¶

The EvalOutput model includes degradation fields:

{
  "scores": {"accuracy": 0.65},
  "passed": false,
  "degraded": true,
  "degradation_reason": "Running in keyword-only mode (no LLM backend)"
}

Via health endpoint¶

GET /api/v1/health

Response includes service status:

{
  "status": "degraded",
  "services": {
    "llm_backend": "unavailable",
    "redis": "connected",
    "database": "connected"
  }
}

Via MCP¶

nav_health()

Returns component-level health including degradation signals.

For CSF formal bridge tools, unavailable solver backends are reported in the tool result itself:

{
  "backend": "z3",
  "verified": null,
  "error": "Formal backend unavailable: Z3 is not installed."
}

verified: null means "no proof was produced by this backend." It is not equivalent to verified: true.

For csf_cognitive, a degraded scenario analysis warning such as [DEGRADED:local_classifier] means the configured safeguard model was unavailable and the tool used local classifier/rule fallbacks. Use that output as a pre-action triage signal. Do not rely on it as a standalone safety approval for regulated, high-stakes, or unattended production workflows.

For job agents, live external integrations are disabled by default. If a workflow depends on current web evidence or LLM-backed debate, check whether G6_JOB_LIVE_SEARCH=1 and G6_JOB_LIVE_LLM=1 are set in the runtime environment. The default is intentional for first-run MCP installs, CI, and demos: no surprise network calls, no surprise paid model usage, and predictable latency.

For job_political, also check whether optional enrichment flags are enabled when richer context is expected: G6_JOB_ENABLE_GROUNDING=1, G6_JOB_ENABLE_EXPERTA=1, and G6_JOB_ENABLE_BAYESIAN=1. With those flags off, political workflows still run their core deterministic policy, legislation, polling, communications, stakeholder, governance, and compliance calculations, but outputs are not current-source-backed or Bayesian-enriched.

Political outputs are decision support only. Do not treat them as legal, electoral, campaign-finance, compliance, lobbying, publication, spending, or targeting approval. Built-in defaults are illustrative; provide real poll data, stakeholder lists, budgets, campaign details, jurisdictions, and source material before using outputs for real-world review. Jurisdiction matters, and constituent/campaign workflows may involve sensitive personal or political data.

For job_scientist, literature review live retrieval is explicitly degraded when Semantic Scholar is unavailable, rate-limited, returns no results, or the runtime is offline. Fallback papers are labelled with connector_status: "fallback" and the response metadata includes connector_status plus synthesis_details.evidence_profile.source_mode. Set G6_SCIENTIST_STRICT_SEMANTIC_SCHOLAR=true when fallback papers should be treated as an error instead of degraded output. The built-in literature synthesis is metadata-level only; use it for triage and planning, not scientific, clinical, regulatory, or publication sign-off without human review.

job_scientist optional enrichment integrations are opt-in and time-bounded. Use G6_SCIENTIST_ENABLE_GROUNDING, G6_SCIENTIST_ENABLE_WEB_SEARCH, G6_SCIENTIST_ENABLE_DEBATE, or G6_SCIENTIST_ENABLE_DEEP_INTEGRATIONS to enable them. Tune budgets with G6_SCIENTIST_<NAME>_TIMEOUT_SEC or G6_SCIENTIST_INTEGRATION_TIMEOUT_SEC; timeout markers are surfaced in integration metadata rather than blocking the core scientific rule path indefinitely.

For job_framework production hardening, distinguish runtime health failures from verification warnings. validate_production_readiness() should block production when no LLM backend is available, and health_check() should report unhealthy if connection-pool or LLM-cache probes fail. The current non-blocking warning cleanup is limited to test/dependency noise: pytest-asyncio loop-scope deprecation, a Pydantic adapt_pandas.DataInput.schema warning, httpx raw-content upload deprecation in tests, and Pydantic serializer warnings from mocked or LLM-like message objects.

What G6 never does silently¶

• Never presents mock data as real results — mock/stub outputs are always labeled
• Never hides missing backends — the health endpoint and reports always show service status
• Never charges for degraded runs — if the LLM wasn't called, you aren't billed
• Never claims accuracy it didn't measure — degraded metrics are flagged

Degradation hierarchy¶

G6 follows a graceful degradation chain:

Full capability (all backends available)
  ↓ LLM unavailable
Cached artifacts (replay distilled results)
  ↓ No cache entries
Rule-based fallback (keyword/pattern matching)
  ↓ Rules insufficient
Explicit failure with diagnostic message

Each level is clearly communicated. The system never pretends a lower capability level is the full system.

Common scenarios¶

"I ran the workflow but accuracy seems low"¶

Check the report's Safety Status section. If degraded: true, the system used fallback mode. Configure an LLM backend (Ollama for local, OpenRouter for cloud) for production-quality results.

"The health endpoint shows services as unavailable"¶

This is informational, not an error. G6 runs without Redis, Elasticsearch, or cloud backends — it just runs with reduced capability. Only configure what you need.

For Elasticsearch specifically, ctx_elastic fails quickly with a diagnostic message when the cluster is unavailable. It does not fall back to local search because index mappings and query semantics are deployment-specific. Use ctx_rag when you need retrieval that works without Elasticsearch.

"I expected ColBERT, but ctx_colbert says degraded"¶

This is expected unless the full retrieval backend is installed in that environment. ctx_colbert only uses real ColBERT when ragatouille==0.0.9.post2 is available and the model/index path succeeds. Otherwise it reports degraded=True and uses embedding cosine retrieval if EMBEDDING_MODEL is configured, or local TF-IDF as the final fallback.

Install the context extension package (which includes the ctx_colbert retrieval component) for your tier, from the setup page:

pip install g6-context --index-url https://packages.g6solver.com/simple/

Real ColBERT additionally needs the ragatouille model backend available in that environment; otherwise ctx_colbert falls back to embedding cosine or TF-IDF retrieval.

Then verify the real backend in the target environment:

G6_RUN_RAGATOUILLE_SMOKE=1 python -m pytest tests/mvp/ctx_colbert/test_ragatouille_real_smoke.py -m heavy -q

"My evaluation says 'degraded' but the numbers look fine"¶

The degradation flag means the system COULD do better with full backends. If your target is met in degraded mode, that's a positive signal — full mode will likely perform even better.

Configuring backends¶

See Configuration for full environment variable reference.