a model card lists what its subject cannot do. so does this one.

the profile runs a daily self-test: 16/18 covered, 88%. the questions it currently cannot answer from its own facts file:

• Which programming paradigm do you dislike most, and defend that opinion?
• Who could give a reference for your teamwork?

they stay listed until i write the missing fact. that is the point: a gap you can see is worth more than a gap you cannot.

behind this README is a retrieval harness pointed at data/facts.md. it ranks the corpus against your question, answers only from what it retrieves, and grades that answer with the same code as ragproof. no tools, no web access, no guessing.

a question the corpus cannot cover is a failing test. one click files it, question and title already filled in:

what does claude-skill-audit scan for? · how does the-breaker fail safe? · what gates a build in starship-flow-control? · what is the hardest bug tj has debugged?

or write your own.

0 trials · n/a answered · 0 gaps open.

nothing open right now. every question anyone has asked is covered by the facts file.

who found a gap

no one has stumped it yet.

• economic-pulse-dashboard — 7 FRED indicators, pandas ETL, trend detection, Streamlit
• Cohort-Retention-Engine — event CSV in, retention matrix and churn curves out
• Bottleneck-Simulator — discrete-event simulation of multi-stage workflows with SimPy
• ecommerce-sql — CLV, market basket and rolling revenue windows on a normalized PostgreSQL schema
• Coke-Recap — interactive 3D recap of the Coca-Cola internship, React Three Fiber
• operations-forecasting-model — multivariate regression and a 15+ metric KPI dashboard
• wa-housing-homelessness — R analysis of rent and homelessness in Washington State
• 3d-ramenshop-portfolio — hand-coded Three.js ramen shop, single file, no build step

two GitHub Actions, no server, no third-party services, zero runtime dependencies.

the panel at the top is an SVG this repo draws itself. it is static on purpose: an SVG that GitHub embeds as an image runs in secure static mode, where declarative animation is disabled, so neither CSS keyframes nor SMIL will ever run there. i checked that against a browser rather than trusting a blog post — a plain HTML div running the same keyframes animated fine while the identical animation inside an image sat on frame one. so the craft is all rendering: gradients, a gaussian-blur phosphor glow, turbulence grain, a scanline pattern and a vignette. two themes, because a dark terminal pasted onto a white page looks broken.

flowchart LR
  A[you open an issue] --> B[run trial]
  B --> C[rank sections of<br/>data/facts.md]
  C -->|found something| D[answer, and cite it]
  C -->|found nothing| E[break: you scored]
  D --> F[log the trial]
  E --> F
  F --> G[redraw the panel<br/>commit to README]

a trial counts as a break when retrieval finds nothing in the corpus. that is deterministic rather than a judgement about how the model phrased itself, so the score cannot be gamed by wording, and the model cannot quietly rescue a question the facts do not cover.

when a model writes the answer, the answer is scored too. faithfulness is the same algorithm and the same 0.55 threshold as ragproof, run with the hash embedder from that repo's own test suite so the workflow needs no model download. that embedder only counts word overlap, so the number is a lexical-overlap proxy, not an entailment check: it is paraphrase-blind and negation-blind, and an answer that borrows corpus wording while saying something false would still score well. read it as a smoke alarm, not a verdict. extractive answers are not scored at all, since they are copied corpus text and would sit at 1.0 by construction. so far: 0 model-written answers scored, mean grounding proxy n/a.

• scripts/render-hero.mjs — draws the terminal in both themes
• scripts/lib/retrieve.mjs — ranks facts against a question, and doubles as the answer path when no model key is set, so the bot degrades to retrieval rather than to silence
• scripts/lib/ragproof-metrics.mjs — faithfulness and relevance, ported from ragproof so the profile grades itself with the code it ships elsewhere
• scripts/lib/eval-log.mjs — replays every past break against the current corpus, so a gap counts as closed only once the facts genuinely cover it
• scripts/build.test.mjs — the gate. an unsubstituted token, a lost theme, glow leaking into the printout, or a row dropped from the panel while the summary still counts it all fail here

the model never gets tools, never sees anything outside data/facts.md, and every question is treated as untrusted input and stripped of markup before it is written anywhere. worst case it says "i don't know", which here is a scoring event rather than a failure.

anything a stranger writes ends up on a page with my name on it, so moderation stays human: a junk trial comes off the board by deleting its entry from data/eval.json and closing the issue. the score is honest, not automatic.