Nishi Family › Compare › Pitwall

Nishi Pitwall -- Crew Roadmap Board

Every Nishi crew role, its measured position vs named incumbents, and the ranked next rungs. The racing crew's board -- drive the roadmap from honest delta-to-incumbent.

The number-one cross-cutting rung. An LLM lane built UNDER the coach, verify, and refusal gates the roster already has is the top gap on FIVE frontier radars — Engineer (momentum 145), Doctor (128), Janitor (111), Team (95), and Racing (LLM-test). The governance the field's agents lack, wrapped around the authorship we lack. Rows marked LLM-lane arc feed it.
Scores are the liar-killed /compare coverage for each role (presence, not depth). Green ≥700, amber 500-699, red below 500. Rungs are ranked by leverage; the lowest-scoring crews are the most urgent. Generated by nx_pitwall_board from knowledge/compare/roadmap.data — a new crew is one data line.
Supervisor347/1000
Service monitoring vs Prometheus and Netdata
  1. Sovereign metrics store feeding the /health surface
  2. Alert routing through the vizsla reminder engine
  3. Per-service resource metering
Genealogist444/1000
Code genealogy vs Sourcegraph and CodeQL and jscpd
  1. An AST-precise typed call-graph (reference search over-reports)
  2. Coupling and cohesion architecture metrics
  3. SBOM attestation (also serves the janitor and warden)
Doctor470/1000LLM-lane arc
Debugging vs rr and GDB and Delve and C-Vise
  1. DWARF line tables in the backends (unlocks the whole field's tooling)
  2. A minimal sovereign breakpoint runner
  3. Input reduction at C-Vise scale
Team473/1000LLM-lane arc
Software agents vs SWE-agent and OpenHands and Aider
  1. Machine-executable targets plus heartbeat-on-advance for real workstreams
  2. Team bench harness (SWE-bench-class scored eval)
  3. LLM-authoring lane UNDER the coach and verify gates
Modelwright476/1000LLM-lane arc
Own LLM vs PyTorch and HuggingFace and llama.cpp and Megatron
  1. Alignment: SFT then RLHF/DPO (radar momentum 120)
  2. A standardized eval harness (shared with the referee)
  3. Train a real model at real scale on our own substrate
Racing500/1000LLM-lane arc
Auto-test vs Playwright and Cypress and Selenium
  1. A sovereign WebDriver or CDP driver so the harness drives real pages
  2. Wire competitive-intel to auto-open a track per census
  3. Flaky-detection plus accessibility audit
Engineer500/1000LLM-lane arc
Language vs Rust and Zig and Go and Elm
  1. Call arity plus param-type checking in nx_parse
  2. Statement-boundary becomes a parse error
  3. An x86-64 register allocator (the 3.65x-vs-C wall)
Warden500/1000
Policy vs OPA and Kyverno and Checkov
  1. A policies-as-data spec the conductor evaluates
  2. Violation reports on the compare plane
  3. Admission-style gates on the deploy path
Janitor583/1000
Code-health vs SonarQube and Semgrep and Renovate
  1. Conductor-wire the detect-clean-ratchet loop autonomous
  2. Re-lock ratchet floors and resolve the 19 shadow collisions
  3. An SBOM organ (frontier momentum 1864)
Referee600/1000
Evaluation vs OpenSSF Scorecard and openai-evals and HELM
  1. Score qabench -- scoring is the named wall (bar 251)
  2. Wire the orphaned honesty grader to a consumer
  3. Contamination detection before TRAIN-our-own ships
Researcher666/1000LLM-lane arc
Deep research vs Perplexity and Tongyi and deep-searcher
  1. Answer-scoring (qabench is the wall)
  2. LLM synthesis and refutation
  3. Neural rerank plus multimodal
Publisher720/1000
Deploy and hosting vs Vercel and Fly and Render and Argo
  1. Canary plus preview environments
  2. GitOps plus autoscale-to-zero
  3. Tenant-scoped ACL on the /relate daemon
PM1000/1000LLM-lane arc
Project mgmt vs MS-Project and Asana and OpenProject
  1. LLM planning copilot
  2. Depth behind the present axes
  3. Arena-reset multi-tenant isolation
Product763/1000LLM-lane arc
Surveys and feedback vs SurveyMonkey and Google Forms and Typeform and Qualtrics
  1. Attention checks and careless-response guards (evidence-ranked from the banked 24-26 corpus)
  2. LLM open-text coding under the coach and verify gates
  3. Adaptive branching questionnaires
Every role above is a censused crew member. The full 35-role crew is enumerated in knowledge/registry/crew_roster.tsv and guarded by nx_crew_roster_gate (roster and RACI must name the same set); roles not yet carded here are RACI'd with a real organ and awaiting their /compare census. Each new census is one line added to roadmap.data.

Generated by nx_pitwall_board (sovereign NishiLang organ) from knowledge/compare/roadmap.data — 14 crew roles. Zero JS, zero trackers.