evidence / Engineering depth
AI Engineering at the Current Frontier -- the five skills 2026 teams hire for, each mapped to committed code
The AI-engineering job market has converged on a recognizable skill set: eval harnesses, fail-closed LLM answer systems, MCP tooling, guardrail engineering, and cost-aware model routing. This page claims the repo demonstrates each one -- not as a resume keyword but as a committed artifact a reviewer can open and run. This is an engineering-parity claim, not a performance or edge claim: it maps machinery, not results. It is the skills companion to agent-fleet-direction.md (how the system was built); read that for the orchestration mechanics this page does not repeat. Truth-source for any figure: JOB_EVIDENCE_PACKET.md.
strongest single receipt
the claim
The current-era AI-engineering interview does not ask "can you call an LLM API." It asks whether you can make an LLM system trustworthy in production: measure it against held-out truth, stop it from hallucinating, expose it as tooling other agents can call, guard it in code rather than in prompt text, and pay for it sensibly. Those five competencies are the job. This repo carries a working, committed artifact for each -- and each was built for a forecasting system, where getting any of them wrong fails silently.
receipts
| claim | committed artifact |
|---|---|
| Eval harnesses | [EVAL_PLATFORM_MAPPING.md](../EVAL_PLATFORM_MAPPING.md); scripts/platformkit/calibration_diagram.py; src/prediction/walk_forward_backtester.py; scripts/validate_calibration_multicorpus.py; tests/test_ingame_leak_free.py |
| Fail-closed answer system | [AI_CONSUMER_CONTRACT.md](../AI_CONSUMER_CONTRACT.md); scripts/platformkit/answers/resolver_registry.py; scripts/platformkit/answers/contract_client.py |
| MCP tooling | scripts/platformkit/mcp_server/server.py; [USE_WITH_CLAUDE.md](../USE_WITH_CLAUDE.md); [mcp-live-demo.md](mcp-live-demo.md) |
| Guardrail engineering | [HONESTY_SYSTEM.md](../HONESTY_SYSTEM.md); scripts/platformkit/hygiene_lint.py; predict_service/honesty_mw.py; scripts/hooks/pretooluse_guard.py |
| Cost-aware routing | [BUILT_WITH_CLAUDE.md](../BUILT_WITH_CLAUDE.md) |
why this matters
Every one of these five is a line on a 2026 AI-engineering job description, and every one is usually asserted rather than shown. Here each is a file you can open and a command you can run. An LLM system is only as good as the harness that measures it, the contract that stops it lying, and the guardrails that hold when no human is watching. I built those first, and the numbers second.
reproduce
# Fail-closed answer envelope -- returns a typed status + source_artifact, or no_data
python -m scripts.platformkit.answers.contract_client "does b2b_rest_penalty hold up" --sport nba
# Eval gate -- walk-forward with a per-fold leak assertion; exits non-zero on overfit
python scripts/run_walk_forward.py --gate
# Guardrail static lint -- exits non-zero on a banned number/phrase outside retraction context
python scripts/platformkit/hygiene_lint.py
# MCP server -- one-command install + connect (prints the exact Claude config)
python scripts/platformkit/publish_pack/install_pack.pyThis site reports calibration and sharpness only, never a dollar edge, ROI, or bankroll result. An honest null is a success.