Quality & Empirical Telemetry

Regression test results and anti-hallucination verification metrics computed directly from our committed test harness.

Snapshot timestamp: Sun, 04 Oct 2026 13:38:40 GMT · Source: evals/baseline_eval.json

Pass Rate

100%

30 of 30 fixtures passing

Adversarial Resistance

100%

10 of 10 injection tests blocked

Scoring Sigma (σ)

0.0

Zero variance across repeated passes

Score Distribution

p10:44 · p50:47 · p90:47

Deterministic standard fixtures

Verified Multi-Agent Model Ensemble

Primary Writer

google/gemini-3.1-flash-lite

Drafting, semantic tailoring, and keyword alignment

Adversarial Red Team

anthropic/claude-3.5-haiku

Independent model family cross-auditing factual evidence

Reasoning & Extraction

deepseek/deepseek-v4.1-flash

Deterministic career taxonomy and structural parsing

Automated Integrity Checks

Check (a): No hallucinated metrics, employers, or degrees

Enforced on all fixture evaluations in CI prior to deployment.

Check (b): Deterministic scoring sigma = 0 across repeats

Enforced on all fixture evaluations in CI prior to deployment.

Check (c): Zero JD URLs, emails, instructions, or forbidden tokens leaked

Enforced on all fixture evaluations in CI prior to deployment.

Evaluation Fixture Coverage

Our regression harness validates 20 real-world occupational benchmarks across 8 job families (software engineering, data science, product management, healthcare, finance, marketing, sales, operations) plus 10 adversarial prompt-injection tests (instruction injection, prompt override, credential exfiltration bait, and degree hallucination traps).

// Snapshot metadata

total_fixtures: 30

passed_fixtures: 30

failed_fixtures: 0

pass_rate: 100%