PLUMB 0.2 · exact transform compiler
Give PLUMB a few input/output pairs. It writes a Python function that reproduces every pair exactly: offline, in under a tenth of a second, with no model and no tokens. If the pairs don’t pin the program down, it stays silent rather than guess.
pairs.jsonPLUMB 0.2.0 · real outputPairs in
2026-11-03 10:15:02 ERROR disk full→"ERROR"2026-11-03 10:15:07 WARN cpu at 91%→"WARN"2026-11-03 10:16:00 INFO backup done→"INFO"Function out Shipped
def f(s):
return ([w for w in (s.split()) if w.isalpha()])[0]How it works
Three to twelve inputs with the outputs you want. Sample log lines and the fields an existing parser already pulled from them are ideal.
PLUMB searches short chains of its own basic steps over text, lists and numbers, and keeps only chains that reproduce every pair exactly.
If its ship rule judges that the pairs pin one program down, that function ships. Otherwise nothing ships, and plumb.first() hands the job to your model or a person.
Real runs
Each card is PLUMB 0.2.0's actual output, run on a laptop on 3 November 2026. The new lines were never shown to PLUMB; they were run through the function it shipped.
Pairs
2026-11-03 10:15:02 ERROR disk full→"ERROR"2026-11-03 10:15:07 WARN cpu at 91%→"WARN"2026-11-03 10:16:00 INFO backup done→"INFO"Function
def f(s):
return ([w for w in (s.split()) if w.isalpha()])[0]New lines
2026-11-03 10:17:44 DEBUG cache warm→"DEBUG"2026-11-03 10:18:02 CRITICAL raid degraded→"CRITICAL"2026-11-03 10:18:09 ERROR 3 retries left→"ERROR"Pairs
GET /index.html 200→"200"POST /login 302→"302"GET /missing.png 404→"404"Function
def f(s):
return (s.split())[-1]New lines
PUT /api/v1/items 201→"201"GET /health 503→"503"Pairs
[1, 2, 3, 4]→6[6, 7]→6[5]→0PLUMB says
# silent: the examples do not pin the answer down (…)“Sum all but the last” and “sum the even numbers” both reproduce every pair. On [2, 4] they disagree: 2 or 6. Its best candidate, droplast › sum, was contested, so nothing shipped.
Pairs
GET /index.html 200→200POST /login 302→302GET /missing.png 404→404PLUMB says
# silent: no chain of its basics reproduces these examplesThe same field, wanted as a number instead of text. No short chain of PLUMB’s basics makes that turn, so it says so. plumb.first() passes jobs like this to your model.
Proof
Tasks, predictions and scoring were fixed and hashed before every run. A wrong answer counts against a right one.
0 wrong. PLUMB alone on machine-written tasks across three sealed sets. Silent on about half.
Head-to-heads won. PLUMB first, with Claude Opus 5.5 only where PLUMB was silent, beat Opus 5.5 alone: 73–63, 79–75, 87–81.
Fewer LLM calls in those head-to-heads, because PLUMB answered first.
Per call, offline. Under a tenth of a second per task on a laptop, the same program every run, no tokens.
Zero wrong in 112 still allows a true error rate of up to about 2.7% at 95% confidence. Every test task was machine-written and checked so that its examples pin it down; real data comes with no such check, and pairs that fit more than one program can slip through (in one test, 2 of 58 answers were wrong that way). The next sealed test rebuilds public log parsers’ field extractors from a few pairs and scores every held-out line.
Where it fits
Seven possible uses, ranked on four questions: are input/output pairs the native input, is a wrong answer costly, does volume rule out a model call per item, and is the space still open? The ranking runs nearly opposite to market size.
Writing, migrating and repairing the extractors that pull fields out of log lines in security and observability pipelines.
A by-example step licensed (OEM) into import, ETL and integration tools.
A model’s logged outputs are ready-made pairs; each compiled function replaces a stream of future calls.
Exact transforms on demand, when the agent already holds pairs.
Transforms written where the data may not leave the building.
A first pass in front of an LLM for everyday code.
By-example features for analysts.
Log scale. Each product’s figure is its own reported count. PLUMB’s figures are targets worked out from conversion benchmarks, not forecasts.
Paying customers
Users
well-targeted visitors to this page in a year.
free sign-ups, at 90 per 1,000 visitors15.
paying companies, at 5 per 1,000 visitors; developer tools convert at about half that rate16, hence the 40,000.
a year at $990 each. Four OEM deals at $24,000 bring in about the same.
For scale: one founder runs Sidekiq with about 2,000 customers, and ScreenshotOne passed 1,000 with no employees. Kite had 500,000 users and could not charge them, which is why PLUMB sells to companies, not to individuals.
Pricing
Proposed prices from our market research. PLUMB is in private preview and not on sale yet.
Limits
We tested PLUMB outside its niche too. It lost, and the losses draw the border.
When jobs arrive as words, a frontier model writing the code wins: Claude Opus 5.5 scored 60 right and 0 wrong; Opus writing examples for PLUMB scored 56 right and 2 wrong.
As a laptop terminal agent it finished 17 of 30 jobs. A 4B model alone finished 23, and Opus 29.
It cut the dollar bill by only 12–13%, because the tasks it answers are the cheap ones. Routing those calls to a small model saves about as much.
Deciding which field maps to which is a job for a model or a catalog. And PLUMB’s functions are exact on the pairs given, not proven for every possible input.
Install
A function ships only when it reproduces every pair and PLUMB judges that the pairs pin it down.
import plumb
pairs = [("2026-11-03 10:15:02 ERROR disk full", "ERROR"),
("2026-11-03 10:15:07 WARN cpu at 91%", "WARN"),
("2026-11-03 10:16:00 INFO backup done", "INFO")]
plumb.solve(pairs) # the function's source, or None
plumb.explain(pairs) # chain, code, shipped or not
# PLUMB first; your model only where it is silent
code, source = plumb.first(pairs, fallback=my_model)pip install plumb-0.2.0-py3-none-any.whl
plumb pairs.json
# prints the function, or why it is silent
plumb --json pairs.json
# shipped, code and chain as JSON
# exit code: 0 shipped · 1 silent · 2 usage