trimcrae
@trimcrae
Public Skills
paper-hardening
by trimcrae
Review a research manuscript or verify a repair against its evidence, with a finite scope and a concrete submission decision.
repo-gates
by trimcrae
Run the repository checks appropriate to a code, evidence, or manuscript change and preserve an honest validation result.
research-loop
by trimcrae
Run one bounded EMC research cycle, select a useful deliverable, claim ownership, and preserve the outcome without overlapping writers.
aixiv-submission
by trimcrae
Post a paper from this repository to aiXiv and read its automated review. Load before submitting or revising anything on aixiv.science, before quoting an aiXiv Rating to anyone, and before setting a target rating as a goal. Covers the working mechanics (agent token, the exactly-three category contract, metadata generated from the manuscript, PREFLIGHT_FULL before any post) and the three traps that cost a day: is_public=0 does NOT make a submission private, the Rating is written by an UNAUTHENTICATED endpoint so it is not a quality measurement, and ELEVEN versions of one paper never moved it above 6 — trending DOWN as the paper improved, until the target was abandoned by decision — while a DIFFERENT paper of the right shape scored 7 on its first version. Also the one thing that does move a rating (pick the paper, do not iterate the prose), why the review TEXT is worth having even though the number is not, the version-numbering and review-latency mechanics (the server assigns the version; 1.9 was followed by 2.0), and what this repository will not do to raise a number.
inflight-reporting
by trimcrae
The exact format of the end-of-turn "In flight" board, and the $/ns drift line that doubles as the buy line. Load whenever your final message will leave real compute running (GPU or CI jobs, subagents doing real work), when you need to decide whether a row gets a $/ns column at all, or when you are about to rent, relaunch or refuse a host on price. Covers: one scannable line per item with state, ET ETA, cost and $/ns against its basis; cost is part of the format, not an extra; the absolute buy line $0.006539/ns (approximately 1.92x basis) and why it is an absolute rate rather than a multiple; the drift line IS the buy line, a hard gate; and why a row we are paying and a row the gate refused must never render alike.
scientific-writing
by trimcrae
Write and revise this repository's submission texts so a working scientist can read them once and follow them. Load BEFORE drafting or revising any manuscript, SI, cover letter or abstract that will go to a preprint server, a journal or aiXiv, and before applying a hardening round's findings to prose. Covers the four moves that carry almost all the gain (topic and stress positions, characters as subjects, actions as verbs, one sentence one idea), the revision pass that finds the damage, and the two failure modes that make a "clearer" paper worse — a hedge quietly dropped to shorten a sentence, and a readability score chased as a target. Metrics are a SCREEN that says where to look, never a measure of whether the prose is good; lint_readability.py reports, this file decides.
eli5
by trimcrae
Explain something in plain language, the way you would to a smart person who does not work on this project. Load when the user types /eli5, or asks for it in their own words — "explain it like a normal person", "in plain English", "I don't follow", "assume I know nothing", "what does that actually mean" — or when they push back on an answer that was dense rather than wrong. Covers the rewrite rules (lead with the point, one idea per sentence, jargon gets replaced not glossed, at most one analogy), what plain language must NOT change (a hedge, a null, an unknown or a number stays exactly as strong as it was), and the failure modes that make an eli5 worse than the original — talking down, hollowing out, and the analogy that quietly becomes the argument.
ci-escape-hatches
by trimcrae
How to route work OUT of this dev sandbox instead of deferring it. Load the moment you are about to write or think "I can't run X here", "no GPU/network/pip here", "this is blocked", "I can't test this locally", or you hit a 403 at the egress proxy (NCBI, GEO, PMC, EuropePMC, UniProt, Springer all block CONNECT). Also load before dispatching a GitHub Actions workflow, running a feature branch's CI without merging, timing a CI step, supervising a billing fleet, or setting up a self-wake poller. Covers: the two standing escape hatches (Actions runner for networked/CPU/PDF/pip work, spot GPU for heavy compute); workflow_dispatch with ref=<branch>; why the jobs API lag manufactures fake stalls; why a schedule: cron does not supervise a billing fleet; the fleet-armed heartbeat gate; retiring the non-file half of a feature; background-bash self-wake pollers; and GitHub auth/commit-signature noise that is safe to ignore.
gpu-compute
by trimcrae
Rules and hard-won provider facts for spending real GPU money — renting a host, launching a fleet or fan-out, picking a provider, pricing a rung, or diagnosing a job that stalled, died or kept billing. Load BEFORE any rental, relaunch, fleet launch or market gate, before writing a job that checkpoints, and before diagnosing a Vast or GCP provisioning, quota, capacity or teardown problem. Covers: never build an environment on a machine we are paying for (the baked Docker images and the parity reading); checkpoint-and-upload-continuously; default every GPU run to spot; serialize only when one result could cancel the rest; Vast capacity refusals, the retired durable machine blacklist, the $/ns market gate and board-depth holds; a relaunch is a new purchase and faces the same ceiling; the host cannot stop its own billing; GCP us-central1-only, GPUS_ALL_REGIONS=1 and the VMs-do-not-self-delete correction.