Portolan, explained from scratch

The idea

Start with a question you already know the answer to

Imagine your chemistry class is told to find a recipe that turns a cheap powder into a useful crystal. Thirty groups try thirty recipes. Twenty-eight fail. Two work. What goes on the poster at the end of the week?

The two that worked.

The twenty-eight failures live in twenty-eight notebooks, and next year’s class will try half of them again, because nobody wrote down “we tried this, here is exactly how, and it did not work.”

Science does this at an enormous scale. Journals publish what worked. What did not work stays in a lab notebook, a spreadsheet nobody opens, or a scientist’s memory. That lost information has a name: negative results. A negative result is an honest, careful attempt that did not produce the hoped-for outcome. A dead end is the same thing seen from the road: a direction someone took and found closed.

Why this matters more now than it did ten years ago

Today, software programs called AI agents can run parts of science on their own. An AI agent is a program built around a large language model that can read papers, propose experiments, write analysis code, and react to results, without a person steering every step. One such system, described in a 2026 Nature paper, reads about 1,500 papers and runs tens of thousands of lines of analysis in a single run.

These agents fail constantly, which is normal. The problem is what happens to the failures. When the run ends, the agent’s memory of what it tried and rejected is gone. Nobody keeps it. So the next agent, in the next lab, tries the same closed road.

What Portolan is

A portolan is a medieval sailing chart. It did not show the shape of the land. It showed sailors where the rocks were, from the reports of sailors who had hit them. Portolan Software is building the same thing for science: a chart of where the rocks are.

Concretely, Portolan builds three things:

1. A record format. A standard, machine-readable way to write down one failed attempt, called a Negative-Results Record (NRR). Think of it as a form with required fields, except the form is JSON, a text format computers read easily, and a program checks every form before it is accepted.
2. A registry. A database that holds these records, that people and AI agents can search before they start work. “Has anyone tried X under conditions Y?” is a question the registry is built to answer.
3. A federation. Not one database but many. Each lab or company runs its own private node (a copy of the software holding its own records). One shared public node, the commons, is open to everyone. The software is open source. Private nodes are how Portolan intends to pay for that work. The records themselves are never for sale. This arrangement, open software with a supported private version, is called open-core.

The one word that holds the whole thing together: provenance

Provenance means “where did this come from, and how do I know?” Every value in an NRR has to point at its source: a table number, a figure, a paragraph, a data file with a checksum. A checksum is a fingerprint of a file, a long number computed from its contents that changes if a single byte changes. If a record says “0 out of 40 colonies were edited,” you can follow the pointer to Table 5 of the paper and the exact data file on Zenodo, and confirm it.

Without provenance, a registry of failures is a registry of rumours. With it, every entry can be checked.

The vocabulary

The website and the records use a fixed set of words. Here is each one, plainly. Words in bold are the ones you will see on the site; words in code are the exact field names in a record.

The three levels of a record

Why three levels? Because one real Arcadia Science experiment answered six questions and got four different kinds of answer. One outcome per record would have forced a lie.

What the attempt was trying to do

What happened: the outcome class

Every path, attempt and finding carries exactly one outcome class (outcomeClass). It is a verdict from a short fixed list, never free text. The main ones:

The split that matters most is between the negative- family and the inconclusive- family. A negative says “it is not there.” An inconclusive says “we did not see it, and we cannot tell you whether that means anything.” Keeping those apart is the whole point of the next three terms.

Controls, and why “nothing happened” needs them

A control is an extra run of the experiment whose answer you already know, done to prove the experiment itself works.

• A positive control is a run that should work. If it does, your equipment, reagents and method are fine. If you then see nothing in the real test, the nothing means something.
• A negative control is a run that should do nothing. If it does something anyway, your experiment is reacting to the wrong thing, and no result from it can be trusted.

Kitchen version: before you decide the new oven is broken, you bake a cake you have made a hundred times (positive control) and you check that the oven is not warming up on its own with nothing inside (negative control).

In an NRR each control carries a kind, from a fixed list. Positive kinds: designated (a known good reagent run alongside), internal-positive (a part of the same experiment that worked), known-positive-task (for computer work: a prediction the program should be able to make), or none. Negative kinds: vehicle (the liquid the drug was dissolved in, with no drug), no-treatment, background, within-subject (the untreated side of the same animal), adversarial-label (for computer work: a prediction the program should not be able to make), or none.

Sensitivity and the informativeness rule

Sensitivity (also called a detection limit) is how small a change the measurement could have picked up. If your kitchen scale reads in whole kilograms, “the weight did not change” tells you nothing about a gram. A power statement is the statistics version: how likely the experiment was to catch an effect of a given size.

Now the rule that makes Portolan different from a list of failures. The informativeness rule says:

A negative result counts as evidence that something is absent only if a positive control passed, the negative control stayed clean, and a sensitivity or power statement exists.

The rule, on a pregnancy test

You almost certainly know an object that has this rule printed on its box: a home pregnancy test.

The test looks for one hormone in urine. Two windows. The result window shows a line if the hormone is there. The control window shows a line whenever the strip has worked at all: liquid reached the chemicals, and the dye still responds. The instructions say that if the control line does not appear, the test is invalid and you must take another one, whatever the result window shows.

Now suppose you take the test and the result window is blank. “Negative.” Is that evidence that you are not pregnant? Walk through the three conditions.

1. Did the positive control pass? Look at the control window. If the control line is there, the strip worked: the chemicals are live and the liquid flowed where it should. A blank result window on a working strip means something. If the control line is missing, a blank result window means nothing at all. The strip may have been dry, expired, or stored in a hot car. You have not learned that the hormone is absent. You have learned that the test did not run. The box calls this invalid. The schema calls it inconclusive-no-positive-control. The control line is what the record would mark as an internal-positive control, passed: true.

2. Did the negative control stay clean? A negative control is the test run on something that definitely contains no hormone: plain water, or a sample from someone known not to be pregnant. It should show nothing in the result window. Manufacturers run this on every batch. Imagine a faulty batch in which the dye bleeds and leaves a faint grey line on plain water. Now a faint line tells you nothing, and the difference between “faint line” and “no line” is a judgement about shading, not a measurement. The test reacts to something other than the hormone, so neither a positive nor a negative reading from it can be trusted. The schema calls that inconclusive-controls-failed. This is exactly what happened in the mouse itch experiment in Example 2: the plain saline injection lit up the brain as much as the drug did.

3. Is there a sensitivity statement? Read the fine print on the box. It says something like “detects the hormone at 25 units per millilitre” and “99 percent accurate from the day your period is due.” That sentence is the sensitivity statement. It tells you how little hormone the test can see, and therefore when a negative means anything. Three days after conception the hormone is far below that threshold. A blank result window that early is not evidence of absence. It is a test taken before it could possibly have seen anything. Take the same test two weeks later and the identical blank window becomes strong evidence. Nothing about the strip changed. What changed is whether the effect you were looking for was within reach of the measurement. Without the fine print you would have no way to tell those two blank windows apart. The schema calls a negative with no such statement inconclusive-underpowered.

Only when all three hold, control line present, blank on water, taken after the box says it can detect, does a blank result window earn the label the schema calls negative-no-effect with informativeness: informative. That is the difference between “I did not see it” and “it is not there.”

Why this is what makes Portolan different from a list of failures. A list of failures is a drawer full of used test strips, all showing one line. You cannot tell which were taken too early, which had no control line, and which came from the bad batch. A Negative-Results Record stores, next to every blank window, whether the control line appeared, what the box said about sensitivity, and when it was taken. Then the next person who opens the drawer can tell which of those negatives they are allowed to believe. The validator does the reading automatically, and when a piece is missing it writes down which one.

If any piece is missing, the finding is marked informativeness: uninformative, and the validator forces its outcome class into the inconclusive family. The record also carries an informativenessReason saying exactly which piece was missing. This is the old distinction between absence of evidence and evidence of absence, turned into code a computer can check.

One exemption: refuted is judged on the quality of the analysis that overturns the old result, not on the controls of the run being overturned. A real paper forced that exemption (see the Raman example below).

Why it failed: failure mode

A failure mode (failureModes[]) is a tag from a fixed list naming the mechanism of the failure: batch-effect, contamination, no-expression, low-editing-efficiency, phase-not-formed, measurement-artifact, retracted, hallucinated-reference, 57 terms in all. A fixed list matters because a computer can count “how often does batch-effect sink a Raman study?” but cannot count free-text excuses.

Who did it, and to what

What a reader can do with it

How much to trust it

The plumbing

The pilot, and what the numbers mean

In September and October 2026 Portolan took nine published studies, five in biology and four in chemistry, and rewrote everything in them as Negative-Results Records. Most of the studies were chosen because they looked likely to break the format, not to fit it.

The ninth study, a Nature paper, is encoded but not published, because its licence forbids sharing adapted material. The code that builds its records is open, so anyone with the paper can rebuild them.

The convergence story: the first study forced twelve additions to the schema, the next eight, then five, then two, then one, zero, one and zero across the chemistry papers. Nothing was ever renamed or removed. A format that stops changing when you feed it new kinds of science has found its shape.

Six examples from the pilot, explained

Each example below is one real record. The headings use the record’s actual outcome class.

Example 1. The algae experiment that found nothing, and why “nothing” is not allowed to mean “no”

Study: Caddell et al. 2026, Arcadia Science, published on The Stacks. Record: arcadia-alcalase-hdr-knockin · Outcome class: inconclusive-no-positive-control · View this study as a discourse graph

The story. Chlamydomonas is a single-celled green alga with a tough cell wall. Scientists wanted to use CRISPR, the gene-editing tool, to insert a piece of DNA into its genome. To get the editing machinery through the wall, they usually make their own wall-dissolving enzyme, which is slow. They tried a cheap off-the-shelf enzyme called Alcalase instead.

What happened. Alcalase did make the wall leaky (that part is a positive attempt in its own record). But it also killed most of the cells, at least ten times more than untreated. Among the colonies that survived and grew on the selection drug, the team sequenced 40. Zero had the intended edit.

Why the record does not say “CRISPR knock-in failed.” Look at the controls block:

• Negative control, kind background: cells given no editing machinery grew 0 to 4 colonies. Passed. Good, the selection drug works.
• Positive control, kind none. The note reads: “no guide/donor/selection combination previously shown to edit in this delivery system.”

No positive control means nobody proved that editing could have been seen in this setup. Maybe the edit never happened. Maybe it happened and the test could not see it. The informativeness rule refuses to choose, so the class is inconclusive-no-positive-control and the finding is uninformative, reason “no positive control.”

What the record adds that the paper did not. A sensitivity line, flagged derivedBy: portolan-ingest: with 0 of 40 colonies edited, an on-target rate above 7.5% would have been detected with 95% probability. That is the “rule of three,” a standard shortcut for what zero hits in N tries can rule out. It turns “we saw nothing” into “anything more common than about 1 in 13 is unlikely.”

The path. The project record is abandoned. Its provenance notes quote the authors’ own caveat, “a single, unreplicated experiment performed by one experimenter using one batch of each key reagent,” and its untried branches list three roads the authors themselves proposed, including the one the rule is asking for: “restore a visual screen to confirm the guides, donor and selection work.”

Vocabulary you just used: attempt, finding, outcome class, positive and negative control, control kind, sensitivity, informativeness rule, derivedBy, provenance notes, untried branches, abandoned path.

Example 2. The mouse brain scan where the “do nothing” test did something

Study: Kolb, Reitman, Lane et al. 2024, Arcadia Science. Record: arcadia-neuroimaging-histamine · Outcome class: inconclusive-controls-failed · View this study as a discourse graph

The story. The team imaged the surface of an anaesthetised mouse’s brain using a label-free method: no dyes, no genetic modification, just a faint natural glow from brain cells when they work harder. They wanted to know whether the method could pick up touch, pain and itch.

What happened. Touch (a vibrating stimulus on a paw) lit up the expected spot on the opposite side of the brain. Capsaicin, the hot-pepper chemical, lit up a lot. Both are positive attempts. Histamine, which causes itch, is the negative one: it produced activity “indistinguishable from the saline injection.”

Why this is controls-failed and not no-positive-control. This time there was a positive control, kind internal-positive: before injecting anything, the team confirmed the touch response in the same session, and only proceeded once they saw it. Passed. The problem is the negative control, kind vehicle: a plain saline injection with no histamine. It was supposed to produce nothing. Instead it produced widespread activation on both sides of the brain, about as strong as histamine. Observed versus expected, passed: false.

When the “do nothing” test does something, you cannot tell whether histamine added anything on top of the needle. The rule names it: informativenessReason: negative control not clean; no sensitivity or power statement.

Why this study was chosen. Of 124 Arcadia publications screened, this was the only one whose central negative had a passing positive control and a dirty negative control. The pilot needed to see whether the rule could tell that failure apart from the Alcalase kind. It did, before any code was changed.

Vocabulary: internal-positive, vehicle, passed: false, informativenessReason, the difference between the two inconclusive classes.

Example 3. The yeast study with no bottles, where the controls are tasks

Study: Cheveralls et al. 2026, Arcadia Science. Record: arcadia-raman-strain-classification · Outcome class: inconclusive-controls-failed · View this study as a discourse graph

The story. Raman spectroscopy shines a laser at a sample and reads the scattered light as a chemical fingerprint. The team collected spectra from nine yeast strains, grown on three separate plates that were meant to be identical repeats, and trained a machine-learning model to guess the strain from the spectrum.

The number that flips. Scored the usual way, mixing spectra from all three plates into training and test sets, the model got a Matthews correlation coefficient (MCC, a score where 1 is perfect and 0 is guessing) of 0.79. Scored by holding out one whole plate at a time, same data, same model, it got 0.32.

Controls with no reagents in them. There is nothing to pipette here. So the controls are tasks:

• Positive control, kind known-positive-task: can the same pipeline tell yeast species apart? Species are far more different than strains, so a working pipeline must manage this. MCC 0.97. Passed.
• Negative control, kind adversarial-label: can the model guess which plate a spectrum came from? The plates were end-to-end repeats, so a clean experiment should score about 0. It scored 1.0. Failed, perfectly.

The model was reading the plate, not the yeast. So the strain result is inconclusive-controls-failed.

Three findings, three verdicts, one attempt. This record shows why findings exist:

The first row is the exemption from Part 2. The 0.79 result is not merely weak; it is overturned by the later analysis. A refutation is judged on the strength of the analysis that overturns it, so the rule does not demand controls from the run being refuted. This paper is why that exemption exists.

The third row is subtle and worth a moment. After correction, the adversarial control is clean (0.1) and the positive control still passes (0.93). The only thing missing is a statement of how small a strain difference the experiment could have detected. The rule flags exactly that, and the authors’ own caveat says the same thing.

Vocabulary: known-positive-task, adversarial-label, refuted, the refuted exemption, inconclusive-underpowered, why one attempt holds several findings.

Example 4. The pollution enzyme that did not work, and the record is sure about it

Study: Steffens et al. 2023, Environmental Science & Technology Letters (American Chemical Society). Record: acs-laccase-pfaa-degradation · Outcome class: negative-not-replicated · View this study as a discourse graph

The story. PFAS are the “forever chemicals” in non-stick pans and firefighting foam. An earlier paper reported that a fungal enzyme called laccase, with a helper molecule, could break two of them down (PFOA and PFOS). This team tried to repeat it.

What happened. Two weeks, six enzyme doses, triplicate reactors: no significant loss of PFOA compared with untreated reactors (p = 0.29). That is a negative.

Why this negative is informative. Check the three boxes:

• Positive control, kind designated: the same enzyme and helper transforming a different pollutant, carbamazepine, about 35% removed in 120 hours. Passed. The enzyme was awake.
• Negative control, kind no-treatment: untreated reactors. No loss. Clean.
• Sensitivity: the measurement method, its calibration range, and the size of effect the earlier paper had claimed (more than 20% removal at 10 days) are all stated. The record notes honestly that there is no formal power analysis and that the ingest is treating the stated range and expected effect as the sensitivity statement.

All three present, so informativeness: informative and the class is negative-not-replicated: an earlier result that could not be reproduced by a careful attempt that was demonstrably capable of seeing it. This is the pilot’s first informative replication failure.

The twist, and a new word. In the first 24 hours the measured PFOA concentration in the water did fall, by 64%. That looks like degradation. But when the team washed out the whole reactor with methanol, they recovered 99 ± 14% of the PFOA. It had not been destroyed. It had stuck to the enzyme and the glass. The record carries this as two refuted findings (one each for PFOA and PFOS) with the failure mode measurement-artifact: an apparent effect produced by sample loss rather than by the process under test. That failure mode was added to the vocabulary because of this paper.

Vocabulary: designated positive control, no-treatment, informative, negative-not-replicated, refuted, failure mode, a stated hedge in the sensitivity field.

Example 5. The chemistry test that gives the wrong answer when it is hot

Study: Macleod, Bage, Meyer and Thomas 2024, Organic Letters (American Chemical Society). Record: acs-tmeda-adduct-kinetics · Outcome class: refuted · View this study as a discourse graph

The story. Chemists who discover a new catalyst for a certain reaction must rule out a boring explanation: that a tiny amount of borane (BH3) is secretly doing the work. The standard test is to add TMEDA, a molecule that grabs borane and holds it. If the reaction still runs, the catalyst is real.

What happened. The team measured how fast the TMEDA-borane pair catalyses the reaction at different temperatures. At 60 °C it is inert: 0.3 millimolar per second and 4% product, against 9.6 and 87% for free borane. Those are negative-no-effect findings, and they are informative, because a designated positive control (free borane) ran alongside, a background control with no borane source stayed clean, and the detection threshold (less than 5% product) is stated. These are the first chemistry findings in the corpus to pass the rule.

But at 80 °C the pair gives 50% product, and at 100 °C it gives 94%. TMEDA lets go of the borane when it is hot. So the standard test gives a false “all clear” above 60 °C. A literature survey in a sibling record found that of 24 published uses of the test, 15 ran it hotter than that.

What is unusual about this record. The thing being refuted is not a compound. It is a method. The record’s entities list includes “TMEDA inhibition test for hidden BH3 catalysis,” with the entity type assay, and that assay is the target of the refuted verdict. A registry has to be able to say “this control is unreliable under these conditions,” because a future agent will otherwise trust it.

Vocabulary: negative-no-effect, informative in chemistry, an assay as an entity, refuting a method rather than a result.

Example 6. The dead end that produced a brand-new compound

Study: Koch, Hoß, Schnakenburg, Karttunen and Kraus 2026, Inorganic Chemistry (American Chemical Society). Records: acs-xef6-oxidation-attempts (inconclusive-no-positive-control), acs-xef6-xe2f11-ruf6 (positive), path acs-xef6-path (negative-not-achievable) · View this study as a discourse graph

The story. Xenon is a noble gas, famous for not reacting. Chemists have pushed it to oxidation state +6 in compounds like XeF6. The team wanted +8, by reacting XeF6 with a very aggressive oxidiser, RuF6.

What happened. Five different mixtures and temperatures. In every one, fluorine gas bubbled off (they caught it with potassium iodide, which turns brown) and the xenon stayed at +6. One of the +6 products, [Xe2F11][RuF6], had never been made before, and they fully characterised it, with a crystal structure deposited in a public database.

How the records relate. The oxidation attempts are a single attempt record with six findings, one per condition plus the fluorine detection. The new salt is a separate positive attempt, linked back with the relation wasDerivedFrom and the note “the compound is the product of the failed oxidation with excess XeF6.” The path’s headline stays negative-not-achievable for its stated target, xenon(VIII), while the by-product is searchable as a success. A dead end and the thing found at the dead end are two records, joined by a line.

Why the attempts are inconclusive, and the authors agree. No positive control: nothing known to produce +8 xenon was run alongside. No negative control. No run count anywhere in the paper; the conditions are described in prose, and the record says so in its provenance notes. The rule says inconclusive-no-positive-control. The authors’ own words, quoted in the record, are “it may be that our reaction conditions are not right, yet.” This is the first study where the rule’s verdict and the authors’ hedge match word for word.

What a future agent gets. Two open questions and three untried branches, including one marked inferred-by-ingest: try a different solvent or a Lewis acid additive, because anhydrous HF only boosts fluorine’s oxidising power with those present. That branch is not in the paper’s conclusions. It was spotted while reading the discussion, and the record is honest that it was.

Vocabulary: a positive record hanging off a negative path, wasDerivedFrom, negative-not-achievable, provenance notes about missing information, open questions, inferred-by-ingest.

The study you cannot see

Study: Ghareeb et al. 2026, Nature. The Robin AI system proposed an eye-disease assay and drug candidates, humans ran the screens, and an AI analysed the data. Three hits were reported.

It is encoded as 11 records and it is where many of the vocabulary words were born: screened items with a proposedBy, performers with model names, and inconclusive findings where the designated control passed when a human analysed the data and failed when the AI did. But its licence is CC BY-NC-ND, which forbids distributing adapted material. So the records are built locally and never published. That is a licence, not a technical limit, and it is the clearest argument for building the commons on openly licensed sources.

If you remember five things

1. A negative result is data. Science throws it away. Portolan keeps it, in a form a machine can check.
2. “Nothing happened” has to earn its meaning. A positive control that passed, a negative control that stayed clean, and a statement of what could have been seen. Miss one and the record says inconclusive, and says why.
3. Three levels. A path (project) holds attempts (experiments) which hold findings (answers). One experiment can answer six questions and get four verdicts.
4. Everything points back. Every number has a locator in the source. Every judgement call is in a provenance note. Every computed value is flagged. Every cited paper’s retraction status is checked.
5. Nobody owns the records. The format and the software are open. Labs run their own nodes. The commons is shared. Private nodes pay for the work. The records are never for sale.

Comments and corrections to [email protected]. The records are at github.com/portolansoft/nrr-records; the schema and code at github.com/portolansoft/nrr. The pilot itself is written up on the Pilot page and the schema on the Schema page.