nanika
- [ ]P0scarcity gate0.1–0.8enterthe request exists
- [ ]P1crystallize1.1–1.12enterrows 0.1–0.8 evidenced
- [ ]P1Aanchor, verify, then freeze1A.1–1A.8enterrows 1.1–1.12 evidenced
- [ ]P2tournament2.1–2.7enterrows 1A.1–1A.8 evidenced
- [ ]P3converge3.1–3.10enterrows 2.1–2.7 evidenced
- [ ]P3Rreception ↰ the one phase with a return edge3R.1–3R.3enterrow 3.9 carries a verdict
- [ ]P4exit gate4.1–4.2enterrow 3.6 names the artifact, and 3R.3 did not fire
- [ ]P5deliver5.1–5.7enterrows 4.1–4.2 evidenced
The arithmetic is fine. Nothing happened.
No run has produced this report. The four evaluations that would fill the delta ship with status: NEVER RUN. This is the first thing to weigh, and it is not a disclaimer — it is the state of the thing.
A mechanism that leaves no artifact did not happen.
nanika is a self-contained agent skill: a scarcity-gated, one-shot quality harness for any deliverable — code, document, design, or plan. It does not know how to write your thing. It knows how to make sure the something that got written is the best one you were going to get.
Named for the thing in the story that grants any wish, absolutely, and then counts what it cost.
why it is built this way
Agents say they ran the gate. Then they don't.
One study measured instruction compliance at 0% while verbal compliance stayed above 90% — a hundred-point gap that human reviewers caught 0 times out of 15.
So nothing here is taken on report. Every row writes a file, and the run closes on arithmetic that has to resolve: rows ticked against artifacts present, salvage raised against salvage disposed, personas returned against personas valid.
An identity that does not balance is not a reporting error. It is a phase that did not run.
ten identities · each with a worked example · all balance under every legal state
what changes
"Do your best" gives you one attempt, graded by whoever wrote it, against a bar they invented.
| one attempt, iterated | A tournament. Independent candidates on genuinely different bets, judged blind, with the losers' best ideas carried into the winner under a disposition that must balance. |
| a bar the system invented | A sourced ceiling. Every top descriptor cites a real artifact, a named property, a verbatim span at a locator. A worker who did not choose it re-quotes the span, checks the control does not share the property, and is asked to beat it. |
| everyone scores well | Blind-pair calibration. Each scorer grades an unlabelled, shuffled known-good and known-ordinary pair before it sees your work. The key is held elsewhere; no scorer reports a verdict about itself. |
| grading its own homework | A closing audit by a worker that did not run it, recomputing every identity from the files on disk. |
| "nothing was left on the table" | A list of what was left. Angles not run, salvage rejected, trade-offs taken, escalations the budget did not buy. |
how it can end
Seven verdicts. None of them has been reached.
ACCEPT · reception-demoted · diminishing-returns · cap-reached · budget-reached · single-agent-best-effort · BLOCK
A plateau with attacks still open reports as a plateau with attacks still open. A run that could not spawn a worker says so and cannot reach the top verdict at all. The exit vocabulary has no word for "fine".
the price
It counts what it cost.
Computed per run, never quoted from a brochure:
agents = 1 preflight + S sweep + 1 checker + E extra engines
+ C angles + J judges + cycles × (D evaluators + K skeptics)
+ P personas + 4 pairwise + 1 auditA mid-size run lands near forty agents. It stops and asks twice — once before launching, once after the ceiling freezes and the number is recomputed against what the ceiling turned out to demand.
Before any of that, a gate counts how often you have spent one of these, shows how the last ones turned out, and tells you when what you are asking for is ordinary work that does not need it.
install
Three hosts, one command.
make link # symlink into every installed host's skills dir
make status # what is installed, per host
make check # verify the skill is self-contained
make test # assert that make check still catches thingsA host you do not have is skipped and named — nothing creates a skills directory for a tool that is not on the machine. link never removes anything; it refuses a path that already exists and tells you which.
It needs a filesystem it may write to and the ability to spawn an independent worker. Without that it degrades to single-agent(declared), thirteen rows go not-run, and the top verdict becomes unreachable — stated, not hidden.
12 reference files · 12 cited · 12 headed · check ok
is it worth it
Not demonstrated. The evaluation that would settle it is written and unrun.
E3 is a deletion control: strip a mechanism, see whether quality drops. Until it runs, every argument here is a mechanism that sounds plausible, and plausible mechanisms are cheap.
Price is not evidence. Expensive implies expensive.
The ways this can be faked from the inside are named in the skill itself, each paired with the trace that would catch it. Several of those traces read none. That section exists so the claim can be checked instead of trusted.
If the full harness is more than your task needs, four parts change the work rather than the record, and they cost a handful of agents:
- ask what outcome would make you regret the request — then push past the first answer
- ask what changes if it lands
- generate competing bets instead of one attempt, and keep the losers
- never let whoever wrote something be the one who grades it
what is inside