Introducing Zorp
An idea arrives most days. Almost none of them become research, because the distance between having an idea and knowing whether it is true is filled with work nobody wants to do: finding out whether someone already answered it, gathering evidence that bears on it, noticing the evidence that cuts against it, and writing the whole thing up in a form another person can check.
Language models collapsed part of that distance. Ask one a hard question and you get a fluent, structured, confident answer in seconds. What you do not get is any account of how much to believe it.
That is the gap Zorp is built for.
The problem is not fluency
A model will tell you that replacing your message broker cuts tail latency. It will say it well. It will not tell you that it found two sources saying the opposite, or that the benchmark it leaned on measured a different workload, or that it decided what counted as success only after seeing the results.
This is not a hallucination problem, and better prompting does not fix it. The issue is structural: nothing in the process creates an obligation. The answer is accountable to nothing, so there is nothing to check it against.
Human research solved this a long time ago, and not with better researchers. It solved it with pre-registration: write down the hypothesis, the metric, and the point at which you abandon the idea, and commit to all three before you gather anything.
The Kill Threshold
Zorp’s core primitive is a number a human supplies that says, in advance, what would prove the investigation wrong.
The agent never proposes it. Before any evidence is gathered, the hypothesis, the metric, and that threshold are written to a file, hashed, and committed to git.
hypothesis = "Replacing Kafka with NATS cuts p99 publish latency"
metric = "p99_publish_latency_ms"
threshold = { direction = "decrease", min_delta_pct = 20 }Because the file is hash-verified when Zorp loads it, a run cannot quietly move the goalposts after seeing results. The record of what was intended survives independently of what was found. When a run crosses the line, it is killed, and the record says why.
Every attempt is recorded, not only the one that worked.
Four capabilities, and a human between each
Zorp runs a loop. It asks whether a question is worth investigating at all, scoring it on redundancy and feasibility, where each score must cite retrieved evidence or count as zero. It gathers evidence through staged attempts, one per invocation, so no attempt slips past unseen. It drafts from the evidence record and is handed nothing else. Then it works out where the finished artifact should go.
A human checkpoint sits before every one of those steps, including the first. There is deliberately no non-interactive mode, on the grounds that a research checkpoint has no safe default. The human is always the author of record.
What we can and cannot claim
Zorp is pre-alpha. The harness and the research foundation are built and tested, as are all four capabilities. That is genuinely all.
We have run a behavioural-compatibility pilot: six contracts, 24 tasks, 48 paired runs each. Two contracts came back compatible. Four came back inconclusive, because our own reference runtime failed its eligibility floor.
We are publishing all six, including the four that did not resolve, because a project whose entire argument is grounded reporting does not get to report selectively. The landing page carries the full table.
What we have not measured is just as important, so it is written down in the same place: no groundedness evaluation, no citation-accuracy figure, no hallucination rate, and no comparison against a plain language model. Those are the numbers that would actually settle the question, and we do not have them yet.
Why this is worth building anyway
Research is the bottleneck on almost everything else. Not intelligence, and no longer the cost of generating text. The bottleneck is the slow, unglamorous work of establishing whether something is true well enough that another person can act on it.
If that work gets faster without getting less rigorous, the rate at which ideas turn into knowledge changes. If it gets faster by getting less rigorous, we have built a machine for producing confident nonsense at scale, and there is already enough of that.
The whole design follows from wanting the first outcome. Commit before you look. Record the attempts that failed. Draft only from what you recorded. Keep a human at every gate.
Zorp is MIT licensed and the code is public. It is early, it is incomplete, and the page describing it says so. Read the evidence record, not the summary.