ONCE THEY were heroes, federal-government gumshoes who prowled the nation's laboratories weeding out corruption whereever they found it. Brought together by the National Institutes of Health in the late '80s, the investigators in the Office of Research Integrity (ORI) have in the years since prosecuted dozens of cases of science fraud, some involving the country's leading researchers, along the way drawing great notices from the press and Congress.

But, suddenly, it has all come tumbling down. In the past few months ORI has faced a barrage of criticism over its investigative methods. On four consecutive occasions, involving some of its most celebrated cases, it has been soundly defeated. Last Friday, in the most embarrassing setback to date, it was forced to drop all misconduct charges against AIDS researcher Robert Gallo after nearly five years of investigation.

The cause of ORI's plummet from grace is a policy instituted last year that allows scientists under investigation to ask for a formal courtroom trial. Previously, ORI's investigators, many of them scientists themselves, had handled accusations of fraud like scientific disputes, interviewing researchers and their accusers separately, and then writing a report weighing both sides of the argument.

But last year scientists found guilty of fraud by ORI won the right to appeal their conviction to a panel of Health and Human Services Department lawyers. After that, the rout was on. In the year and a half before the appeals board was created, ORI won 16 cases. Since then, in every case that an accused scientist has asked for a appeal, the scientist has won.

What on earth is happening?

ORI officials say the problem is that the appeals panel requires an unfairly high standard to prove misconduct, relying on a criminal definition of fraud instead of what they feel is a more appropriate scientific definition. They are now expected to prove not just that data were falsified, they say, but also that the falsification was deliberate and had a significant effect on the research in question.

The scientists who have won appeals, however, call this issue a red herring. The real impact of the trial, they insist, is that it lets them for the first time confront their accusers and reveal what they believe to be the ORI's shoddy investigative procedures. They point to the two full opinions issued by the review panel so far, which attack the ORI less for failure to meet a criminal-style standard of proof than for an inability to prove that fraud occurred at all. In one of its decisions involving NIH research Mika Popovic, for example, the appeals panels ripped the ORI for using biased witnesses, misunderstanding the research it was investigating, having "fundamental flaws" in its investigative methods and drawing "unreasonable inferences" from testimony and data.

ORI's implosion doesn't, of course, validate the purity of American science. Last week, the first major national study on the subject found that many of the nation's scientists have first-hand knowledge of laboratory misconduct. Nonetheless, a close reading of ORI's last four defeats -- four of the most extensive and highly publicized cases in its history -- reveals that ORI's concern wasn't really with science fraud as it is popularly understood. It wasn't chasing con men. It was after something substantially more abstract and elusive. ORI wanted scientists to be perfect. * The Sharma case. Four years ago, Rameshwar Sharma, a researcher at the Cleveland Clinic, submitted a grant proposal to the NIH. The next year his case was taken up by ORI, who later found him guilty of misconduct.

What did Sharma do? Make up data? Plagiarize someone else's research? Not quite. Sharma spent thousands of dollars on a legal defense and had his work disrupted for three years because, according to the final appeals board report on his case, in one sentence of a 50-page grant application that was never funded, he committed a typo.

Sharma's work concerned two proteins. In scientific notation, the first was called alpha followed by the subscript 2A. The second was called alpha followed by the subscript 2GC. Because in his grant application Sharma refered to these proteins some 130 times, he programmed two adjoining keys on his computer to type these out. Unfortunately for him, on page 21 of his original application, he or his wife -- who did much of his typing for him -- hit the alpha 2GC key instead of the alpha 2A key.

Big mistake. Alpha 2GC, as it turns out, is a protein that Sharma had not finished his research on. Alpha 2A was a far simpler protein that he had worked extensively on. When ORI saw the error it concluded that Sharma was trying to pull a fast one on the NIH, pretending his research was further along than it actually was.

The problem with this theory, as the appeals board painstakingly pointed out, is that there was no evidence to suggest that this was, in fact, Sharma's motive. First, all the discussion that follows the typo is of alpha 2A. Alpha 2GC is such a big deal, said the board, that if Sharma actually wanted to convince NIH that he had mastered it, he would have logically spent several pages describing the accomplishment. But he did nothing of the sort. In fact the data that follows is all about alpha 2A. Even the NIH panel that originally considered his application wasn't confused by the typo. The panelists knew what he really meant. Not ORI, which found him guilty of misconduct.

Sharma appealed and was cleared of all charges in August. * The Hamosh case. In 1989, an employee in the Georgetown University laboratory of Margit Hamosh went to ORI with a number of serious allegations against her boss. After a long investigation ORI zeroed in on the first half of a sentence Hamosh had written in 1985 at the end of a 20,000-word grant application.

The sentence reads: "Last, but not least, we are presently using the newborn rabbit as an animal model for total parenteral nutrition (TPN)... ."

The rabbit research Hamosh was talking about was peripheral to the central experiment proposed in the application, namely a study of digestion using 125 to 130 animals and 30 humans. It was such a minor part of her application that when a panel of scientists reviewed her funding request they didn't even comment on the issue of the rabbits. But when ORI went digging in her notebooks it concluded that what she said was a fraud. At the time she wrote the grant, they said, she didn't actually have the animal model up and running.

Hamosh countered that when she wrote the sentence she had figured out the model was a viable one for experiment. She had designed the experiments. She had received the money to run those experiments. She had obtained the special surgical equipment. She had begun collecting the necessary preliminary data on the rabbits in preparation for beginning the TPN procedure. In other words, she was ready to use the rabbit model.

Over the course of three years, ORI and Hamosh battled it out. Did "presently using" in the context of her sentence, mean "up and running" or "putting into service." The American Heritage Dictionary definition of "use" was introduced as evidence. ORI found her guilty. But Hamosh asked for an appeal. Just before her trial was to begin last month, ORI dropped the charges.

* The Popovic case. Mika Popovic was Robert Gallo's co-worker, the man who did the experiments that made the development of the AIDS test possible. He was accused of three main charges of scientific fraud, all related to a paper he published in 1984.

This was the most serious of these charges: In the body of the article, several charts showed the progress Popovic had made in growing several strains of HIV. The charts have dozens of entries. But in eight instances, instead of entering a number or percentage on the chart, Popovic used the initials ND. In the legend at the bottom of the page, ND is identified as meaning "Not Done." After examining Popovic's notebooks, however, ORI found that in those instance when he used ND, he actually had done the experiments in question and simply hadn't reported the results. ORI never came up with a motive. But under its rules, it didn't need one. He was guilty.

The problem was, Popovic didn't write the legend identifying ND as Not Done. He's a recent Czech emigre, and his English is so bad that others did the bulk of the editing. What he says he meant by ND is "Not Determinable," meaning that he did the experiment but he either couldn't read the results properly or didn't find them conclusive enough to put in the chart.

Moreover, all kinds of virologists regularly use ND to mean "Not Determinable." As the appeals board pointed out, ORI was told repeatedly by neutral scientists advising it in the investigation that ND had no strict definition.

As the appeals board noted, ORI didn't present evidence that Popovic actually benefited from using ND instead of the actual data, or that the use of ND changed the meaning of the charts or the usefulness of the information they conferred. Even ORI itself remarked on the "meaninglessness of the fabrication."

Popovic appealed, and 10 days ago, in a blistering opinion, the appeals board completely exonerated him. "One might anticipate that from all this evidence, after all the sound and fury, there would be at least a residue of palpable wrongdoing," the opinion said. "That is not the case."

* The Gallo case. For close to five years, ORI combed through a 13-foot pile of Gallo's notebooks and conducted 7,000 hours of interviews and deliberations looking to confirm sensational allegations that the NIH researcher might have stolen the credit for discovering AIDS from the French. In the end, ORI's case came down to this: the precise meaning of a clause in a sentence that appeared in the conclusion of a scientific paper Gallo co-wrote 10 years ago.

The paper in question explains how the Gallo lab, in 1984, figured out how to mass-produce HIV -- a revolutionary achievement that made it possible to study the virus in depth and develop the AIDS blood test. At the end of the paper, Gallo compares the data his lab had developed on its strains of HIV with data from a paper the French published a year earlier on a virus they called LAV.

The two viruses, Gallo noted, look very different based on the descriptions made in his paper and the French paper. On the other hand, he then states, "it is possible that this is due to insufficient characterization of LAV because the virus has not yet been transmitted to a permanently growing cell line for true isolation and therefore has been difficult to obtain in quantity."

What does this sentence mean? Gallo says that, based on the context of the paragraph, it means the following: The two viruses may in fact be exactly the same. The only reason the French data may look different is that, since the French could not grow their virus as completely as we did, they couldn't describe it properly.

ORI deconstructed the clause quite differently. It looked in Gallo's notebooks and discovered that he had in fact grown LAV himself. He was not speaking about the French inability to grow LAV, it said. He was saying that no one could grow LAV and that, it concluded, was not true.

Even if ORI's interpretation is correct, just what Gallo stood to gain by the ambiguity of this clause is open to question. Further, no one ever claimed that the disputed sentence changed, undermined or altered the substance of the paper itself, which is widely considered the most important in 20th-century virology. On Friday, faced with the prospect of performing its syntactical exegesis of the disputed clause in court, ORI dropped all charges against Gallo. Presumably Gallo will avoid compound sentence structures in the future.

Malcolm Gladwell is the Washington Post's New York bureau chief.