jsort

sort, but the key is a description.

$ jsort -o "more hawkish about inflation" presser.txt | head -6
2.18	0.12	The plain fact is that inflation is too high and has been for too long.
2.18	0.16	I defined the standard for action: We must be confident that underlying inflation is moving to our objective, clearly and at sufficient speed.
1.76	0.19	Yet for more than five years, inflation has been running above target.
1.68	0.36	Too many categories are still posting increases above 3 percent, on both a 6- and 12-month basis.
1.52	0.38	In the meeting just concluded, the FOMC decided to raise the target range for the federal funds rate by ¼ percentage point to 3¾ to 4 percent, in support of the Federal Reserve’s dual mandate.
1.39	0.22	The committee’s unanimous vote shows our resolve to achieve price stability on a timelier basis.
jsort: 51 texts, 255 comparisons in 10 rounds; reliability 0.96; first-position lean -0.07; 255 calls, 0 cached; 87,187 tokens; $0.0037; 7.2s

$ jsort -r "more hawkish about inflation" presser.txt | head -3      # the other end, from the cache
And with that, I’ll take a few of your questions.
New hiring, private-sector earnings, business capital investment—each of these markers has improved in recent months and is pointing in a good direction.
Those who are least well-off have the most to gain from a durable expansion, a solid labor market, and stable prices.

presser.txt is the chair's opening statement at the FOMC press conference of September 16, 2026, one sentence per line. The first column is the score and the second its standard error.

jsort shows Jev, TypeSafe's decision model, two texts at a time and asks which ranks higher on the dimension you described. Jev answers with a probability in about 200 ms. jsort asks about a few pairs per text, not all of them, fits a Bradley-Terry scale to the answers and prints the texts from the top of the scale down. With -o each line carries its score and a standard error, so the order can be used as a measurement and not only as a ranking.

The same command sorts whole documents. Every opening statement since press conferences began, 91 of them from January 2012 to this week, as files named by date:

$ jsort --whole -o --max-chars 16000 "more hawkish about inflation" statements/*.txt | sed -n '1,4p;88,91p'
6.56	0.41	statements/2022-11-02.txt
6.08	0.66	statements/2022-09-21.txt
5.56	0.58	statements/2022-06-15.txt
5.49	0.30	statements/2022-12-14.txt
-3.33	0.25	statements/2020-11-05.txt
-3.44	0.29	statements/2020-07-29.txt
-3.46	0.24	statements/2021-01-27.txt
-3.56	0.20	statements/2020-06-10.txt
jsort: 91 texts, 455 comparisons in 10 rounds; reliability 0.98; first-position lean -0.03; 0 calls, 455 cached; 0.1s

That run came from the cache. The first time it was 26 seconds and six cents, for statements of about 1,250 words each. The top four are 2022 hikes of 50 or 75 points. The bottom is the first year of the pandemic. By chair, the average score is -2.35 for Bernanke (9 statements), -1.34 for Yellen (16), 0.55 for Powell (63) and 2.67 for Warsh (3). bench/fed.py builds both files from federalreserve.gov and checks the scale against what the Committee did; the results are below.

A thousand short lines take about 5,000 comparisons: roughly a minute and seven cents.

Install

uv tool install jev-sort        # the command it installs is jsort

jsort finds a key the way jgrep does: TYPESAFE_API_KEY, OPENROUTER_API_KEY, or a System One gateway (JEV_GATEWAY_URL and JEV_GATEWAY_API_KEY), from the environment or from ~/.config/jev/typesafe.key, openrouter.key or gateway.key. Force a choice with --api or JEV_API. The two tools share one answer cache.

Use

jsort "more urgent" tickets.txt | head                       # the ten most urgent
jsort -r "more urgent" tickets.txt | head                    # the ten least
jsort -o "more hawkish about inflation" statements.txt       # with scores and standard errors
jsort --top 5 "a more serious safety problem" complaints.txt
jsort --para "makes a stronger causal claim" paper.txt       # paragraphs, not lines
jsort --whole "of more general interest" abstracts/*.txt     # whole files; prints their names
jsort --jsonl --field event.message "angrier" events.jsonl   # full records come back, sorted
jsort --csv --field narrative -o --keep-order --name breadth \
      "drew broader participation" events.csv > scored.csv    # add a variable, keep the row order
jsort --json "more urgent" tickets.txt                       # rank, score, se and source line

Blank lines are dropped. Identical texts are compared once and share a score. Several files are sorted together, as sort does. Exit status is 0 when the input was sorted and 2 on any error: an unreadable file, a failed comparison, a stop at the budget, the API refusing further calls because the credit ran out. In each of those cases whatever could be sorted from the answers already paid for is still printed.

CSV needs a header with unique column names. A byte-order mark is ignored, a row with more values than the header keeps them at its end, and a header with no rows comes back as a header. With -o, jsort refuses to write over a column or JSON field that already exists; pick another prefix with --name.

Write the description as a comparative: "more urgent", "easier to read", "drew broader military participation". It goes into one question, Text A ranks higher than text B on this criterion: "...", so anything that completes that sentence sensibly will work.

What the numbers mean

Score. The position on the scale, in logit units, centred on zero. A gap of 1.0 between two texts means Jev gives the higher one about 73% in a head-to-head; a gap of 3 means about 95%. Scores are relative to the other texts in the same run. They do not carry over to another file or another description.

Standard error. How well the comparisons pin the score down. Jev's answer is a probability, not a win or a loss, so the model is a fractional logit and the errors are the robust (sandwich) kind, with a leverage correction because there is one parameter per text. Where Jev's answers line up on one scale the errors are small. Where they contradict each other the errors grow. In simulation, intervals of 1.96 standard errors cover the target 94 to 95% of the time from -k 6 up (bench/simulate.py). Above 4,000 distinct texts the same errors are estimated by random probing, to within about 7% each, because the matrix the exact ones need no longer fits.

Reliability. The comparisons are dealt into two halves, a scale is fitted to each, and the correlation between the two is stepped up to full length (Spearman-Brown). Near 1, the order does not depend on which pairs happened to be asked. Below 0.8 jsort says so: raise -k, or the description is not one these texts can be ranked on. It errs on the cautious side, and most at low -k, where a half is only two comparisons per text: in simulation it reads 0.83 for a true 0.89 at -k 4, and 0.955 for 0.964 at the default.

First-position lean. How far the first-shown text's chance sits from a half in an even matchup. Positions are randomised and balanced and the lean is estimated with the scale, as home advantage is in a sports model, so it does not tilt the scores. It has been a few points either way.

Reliability and the standard errors describe how consistent Jev is. They say nothing about whether Jev is right. For that, check a sample against your own judgment.

From Python

import jsort

r = jsort.rank(statements, "more hawkish about inflation", per_item=10)
for i in r.order()[:5]:
    print(f"{r.score[i]:6.2f} ±{r.se[i]:.2f}  {statements[i]}")
r.reliability, r.lean, r.asked

r.score, r.se and r.comparisons are arrays aligned with the input. The command's seat belt applies here too: spending stops at budget= dollars, by default $JSORT_BUDGET or 1.00, and r.over_budget says whether it was reached. It works inside a notebook. jsort.arank is the same thing as a coroutine, for a client you already hold.

How it chooses pairs

All n(n-1)/2 pairs are never needed. A lopsided pair says almost nothing, because the answer was predictable. A pair of near neighbours says the most. The first round is a random ring: every text meets two others, once in each position, which connects everything. After each round the scale is refitted, and the next round pairs texts that currently sit next to each other, as a Swiss-system tournament does. Each score is jittered in proportion to how uncertain it still is, so a text that is not pinned down yet keeps meeting new neighbours. A pair is asked at most twice, once in each order. With a handful of texts jsort simply runs out of pairs and stops early.

--top N adds one rule. Once a text has three comparisons, jsort asks how much worse the fit would get if that text were moved up to the edge of the top N. If the answer is "much worse", the text gets no more questions. A text that lost 0.03 to 0.97 against a middling opponent is out, and no further comparison will change that. In simulation this is the same cost as a full sort or a little less, and finds more of the true top ten (0.85 against 0.77 of them, among 2,000 texts).

Cost

A comparison bills about 270 tokens of overhead plus both texts and the description. For short lines that is about 320 tokens, or $0.0000135 at $0.042 per million, and the default -k 10 asks five questions per text: about 7 cents per thousand lines. For 170-word excerpts a comparison is about 735 tokens and a thousand texts cost about 15 cents. A run is about ten rounds, each as slow as its slowest call, so a small sort takes five to ten seconds. A large one manages about 90 comparisons a second through OpenRouter.

jsort stops asking at --budget, one dollar by default, and sorts on what it has. Nothing is lost: every answer is cached in ~/.cache/jev/answers.sqlite, and the choice of pairs is seeded, so a rerun with a higher budget, or a higher -k, starts by replaying the same questions from the cache and only pays for the new ones. The exception is a run in which a comparison failed. Later pairs depend on earlier answers, so once the failed one succeeds the rounds after it can choose differently, and those questions are new.

How well does it work

Two checks, run on 2026-09-19 with Jev 1.13 through OpenRouter. Both run the installed jsort command, uncached.

Against people making the same comparisons. bench/readability.py. The CommonLit Ease of Readability corpus (Crossley et al. 2022) has 4,724 excerpts written for grades 3 to 12. Each has an easiness score that is itself a Bradley-Terry fit, to teachers' judgments of which of two excerpts is easier. So this is the same method with Jev in the teachers' chair. On a random 300 excerpts, the teachers' scale has a reliability of about 0.78, which means no measure can be expected to correlate with it above about 0.88.

Three things to take from it.

jsort gets most of the way to the ceiling from a two-word description, and it clearly beats the other thing you can do with two words, which is to ask for a probability per text and sort on that. That probability bunches up (73 distinct values among 300 texts) and it is weakest exactly where a sort is needed, on texts that are close.

More comparisons stop helping early. -k 10 and -k 16 give the same answer, and -k 6 is nearly there. What is left between 0.82 and 0.88 is Jev disagreeing with teachers, not noise that more questions would average away; the reliability of 0.98 says the same.

A written rubric did just as well for a sixth of the price. Five levels, from "suited to children in the first years of school" to "suited to readers at the end of high school or beyond", one call per text, and Jev's score answer interpolates between levels (2.9, not 3), so it does not bunch up either. The two agree with each other at r = 0.92. Broken down by how far apart the teachers put two excerpts, jsort is ahead only on the closest pairs, and not by much:

Readability across ten school grades is friendly ground for a rubric: the levels are easy to name and the texts spread across all of them. So:

• If you can write down what the levels are, write the rubric. It is what jcol's tone: calm < upset < furious columns do, and it costs one call per text.
• Use jsort when you can say what "more" means but not what the levels are, when the texts would all land on one or two rubric levels, or when the scale is going into a model and you need a standard error on every score and a reliability figure for the whole.

Against what the Fed then did. bench/fed.py. The 91 opening statements above, sorted whole on "more hawkish about inflation", against the top of the target range for the federal funds rate (FRED series DFEDTARU). Both tests were written into the script before the sort was run. The rank correlation between a statement's score and the move announced that day is +0.47 (91 statements). With the change in the target over the following 180 days it is +0.38 (87 statements). Those are moderate, and they should be: a chair can hold rates and talk tough. The holds of June, September and November 2023 score near 4, among the ten most hawkish. What the scale gets right is the shape of fourteen years. The six statements that announced hikes of 50 or 75 points in 2022 are the top six, seven of 2023's eight statements fill ranks 7 to 13, and the bottom eight are all from March 2020 to January 2021.

Things to know

• These are a model's judgments. Check a sample before you rely on an order.
• Jev reads the description literally and is weak at counting and at comparing numbers or dates. "Mentions more people" or "happened earlier" are bad dimensions. Compute those in code.
• One dimension per run. If the description mixes two ("more urgent and more polite"), the scale is whatever blend Jev makes of them, and reliability may not warn you.
• Both texts must fit in one call. --max-chars caps each at 8,000 characters; TypeSafe documents a limit of 32k tokens for the state and question together. Long texts full of detail that is irrelevant to the dimension are a documented weak spot, so cut them down first (jgrep --para is one way).
• Jev is close to deterministic, not exactly so. The cache makes reruns exact. For results that must reproduce, pin the model with --model (for example typesafe/jev-1.13 on OpenRouter) and keep the cache file with the project.
• Adding texts later means a new run and a new scale. The old answers are still cached, so the new run is cheap, but scores from two runs are not on the same footing unless the texts are the same.
• Text in the input can try to steer the answer. Do not use an order from jsort as a security boundary.

Development

uv sync && uv run pytest              # offline: a fake API, no key
uv run python bench/simulate.py       # offline: a simulated judge with a known scale
uv run python bench/probe.py          # live, under a cent: which question shape to use
uv run --group bench python bench/readability.py prepare && uv run --group bench python bench/readability.py run

src/jsort/model.py is the scale: the fit, the standard errors, reliability and the test --top uses. schedule.py chooses pairs. engine.py runs the rounds and is the Python API. inputs.py reads lines, paragraphs, files, CSV and JSONL. core.py is jgrep's client, unchanged: backends, retries inside a time budget, the cache, in-flight deduplication and the cost meter.

Why a noul and not a choice: bench/probe.py asks every ordered pair of ten texts both ways. A two-option choice and a noul ("A ranks higher than B") made the same decisions, 95.6% correct, but the choice put 93% of its answers below 0.1 or above 0.9, where the noul left 28% of its answers between the two. Those in-between answers are what a scale is fitted from.

MIT license.