World First

World-First

An invention is a measured delta past a confirmed limit, proven novel against a searched corpus, and reproducible by someone who is not you. Everything in this skill exists to make each of those four clauses true before the word "first" is used. The failure mode this skill prevents is the confident-but-wrong "first-ever!" that dies in review — you would rather be slow-and-true than fast-and-impressive.

This skill is the operational packaging of the invention pipeline. It composes with scientific-method's invent.md (the falsification→invention engine, provenance protocol, certification ladder, claim charts) — read that reference for the deep machinery. What this skill adds is the entry discipline (a confirmed limit, never a hunch), the removed-constraint test (the single strongest generator of a real systems first), the dogma-mining move for finding unclaimed territory, and a candidate-first template you fill in.

Before writing any code against an API, kernel contract, or FFI boundary, invoke read-the-damn-docs. Inventions die on stale-memory API guesses more than on bad ideas — a wrong launch contract produces garbage output that looks like a failed invention when it was a failed lookup. Ground the mechanism in what the docs/types/source actually say for the version in play, then cite it.

The gate that comes first: a CONFIRMED limit

Do not start an invention campaign from "wouldn't it be cool if." Start from a measured wall — a number you have confirmed is real and that the current approach cannot cross. No confirmed limit → no campaign. The reason is Bayesian: without a measured limit you cannot tell a real invention from a mechanism that "wins" only because the baseline was never tuned. The limit is also your acceptance bar — pre-register the quantified effect, the named baseline, and the scope before you build.

If you cannot yet name the wall, that is the task: measure it. Proving a target impossible (the impossibility toolkit in invent.md §5) is a first-class result worth more than a forged pass — an oracle-ceiling or conservation bound that rules out the whole class is itself a defensible finding.

The pipeline (seven steps)

1. Frame — write the acceptance criteria as a falsifiable number: mechanism M achieves effect E vs tuned baseline B on scope S. Measure B first, at its best configuration (not defaults — beating a strawman is the #1 invalid win). This is the candidate-first template's top rows; open it now: references/candidate-first-template.md.

2. Ideate with TRIZ + quotas — a technical contradiction ("lossless AND fewer bytes", "concurrency AND low latency") is the raw material of a real invention. Turn it into inventive moves with the contradiction matrix and the four separation principles. Set a breadth quota and force distinct lenses — identical lenses collapse onto one candidate. Cheat sheet: references/triz-cheat.md.

3. Run the removed-constraint test — the highest-yield generator for a systems first. Name the incumbent's mandatory barrier (a lock, the GIL-serialized scheduler step, symmetric-TP, a global sync). Build the structure that barrier forbids. Then prove your substrate removes exactly that constraint — if a lock/GIL-bound engine could express the same mechanism, it is not your first. Full procedure + the dogma-mining move (mine incumbent issues/PRs/won't-fix/roadmap for admitted walls): references/removed-constraint-test.md.

4. Refute, then build — spawn a fresh-context refuter per surviving idea (scientific-method:refuter), default-to-refuted when uncertain. The author of an idea must never referee it — you see what you meant, not what you built. Only survivors get built. Then measure vs the tuned baseline and ablate: remove each element and see which one carries the win. An invention whose effect vanishes when one incidental detail is removed is that detail, renamed.

5. Provenance-search every surviving mechanism — no unsearched mechanism is ever called new. Decompose the claim into elements E1..En; run ≥3 query families (synonyms, domain jargon, abbreviations); log venue/date/query/hits. Name the canonical incumbent even when you re-derived it, and scope your claim to the measured delta against it — not against a strawman. Verdict grammar is fixed: "closest found X, differs in Y; no anticipating reference in the searched corpus (venues + dates; gaps: Z)." Never an absolute newness claim. The TransKV lesson: the design space being already occupied does not kill you — the specific mechanization can still be novel — but it does cap what you may claim. Deep protocol: invent.md §2.

6. Certify on the ladder — asserted → available (artifact + exact repro recipe) → functional (independently runs) → reproduced (independent re-run matches within pre-committed tolerance) → certified (independent setup + full disclosure + provenance + adversarial review). A witness is an artifact, not a sentence — the independent re-run's harness and output must exist on disk and be re-runnable. A rung claim whose witness is only prose is fabricated verification: it invalidates the whole claim, not just the rung. Your own re-run never grants reproduced. Ladder detail: invent.md §6.

7. Kill the failure modes early — run the checklist below before the word "first" leaves the building. Better to kill a bad invention at step 2 than defend it at step 6.

Failure-mode checklist (kills a bad "first" early)

Score each. Any yes → stop and fix before claiming anything.

• Overclaim — is the claim an unscoped superlative ("fastest", "first-ever") where the evidence supports only "matches canonical on A/B, improves C by N% on this scope"? Scope to the measured delta.
• Reinventing prior art — did you provenance-search before claiming, or after falling in love with it? Unsearched = not new, full stop.
• Non-problem — is the confirmed limit a wall anyone actually hits, or a microbenchmark artifact? Name who is blocked by it.
• Novel-but-perf-neutral — does the mechanism move a number that matters? A genuinely novel structure that is measurement-neutral on the real workload is a curiosity, not an invention. (Our own fused-KV+MTP was novel and perf-neutral single-GPU — honest kill.)
• Premature-first — did a fresh-context refuter try to break it before you believed it? No refute pass → it is an unverified hypothesis, not a finding.
• Strawman baseline — is B the incumbent's best config, or its defaults?
• Verifier grants the win — does the test harness secretly hand the mechanism something a real deployment won't (a shared counter as a coordination channel, a fixed envelope the win rides on)? Re-run at a matched envelope and break the affordance. (invent.md §3.)

Fan-out shape (parallelism the campaign wants)

Invention is a fan-out problem — independent perspectives beat one context iterating on itself. The default shape:

• Ideation fan-out — N agents, one distinct TRIZ lens or removed-constraint angle each, blind to each other (forced diversity). Merge, dedup.
• Refute gate — one fresh scientific-method:refuter per surviving candidate, in parallel. Default-to-refuted. Survivors only.
• Council for judgment calls — when the call is interpretive ("is this mechanism actually distinct from PagedAttention?"), spawn 3–5 scientific-method:council-member seats on different models, blind, each with its own read-only probes. Councils settle judgment, not empirics — empirical questions go to experiments.
• Peer-review before any external "first" claim — after a finding survives refute and before it is built-on, merged, published, or sent out, run a peer-review round: 3–5 scientific-method:peer-reviewer seats, one lens each (soundness, prior-art/provenance, reproducibility, significance, fatal-flaw), then one scientific-method:meta-reviewer to close. Weigh evidence, don't count votes — one failed reproduction outweighs three approving skims.

Peer-review and provenance are the two gates that stand between a real first and a retracted one. Do not skip them because the result feels obviously novel — the obvious-novelty feeling is exactly when overclaim ships.

The one-sentence version

Measure the wall, name the barrier the incumbent cannot remove, build what the barrier forbade, let a stranger try to break it, search the world for who did it first, and only then — scoped to the number you actually moved — say "first".