Imagine that the computational limits of generalised testing (as defined in On the Practicality of Abductive Validation) were somehow magically lifted. International knowledge bases could then be built, and constantly reviewed. Knowledge is power. Could we resist the temptation to abuse such a resources?

Last Monday I got rich. It happened like this...

Every Monday, "OldManEmu" meets in my office (cause I've got the best graphics terminal). Officially, we're the "TestNet OZ-A hypothesis testing team" but since everyone else was naming their groups "Socrates" and "Plato" we changed our name to something all-wise and local.

A fly on the wall would have noticed that we were all a little distracted. As we argued (yet again) about our local policy for billing outside searches, we'd occasionally peek a glance at my screen. Nothing big was happening on TestNet so the screen was just doodling away to itself with a screen saver. There's always something happening: some low priority, long half-life query kicking round from site to site looking for an insight that might make it useful. We call them ghosts. Once I set off a ghost to do the following:

• for every thing that can be proved ...
• ... try and disprove it ...
• ... and explain the reasoning that lead to the refutation

I'd forgotten all about it till a few years back when it arrived home. The poor thing had been kicked off every site in the world since it used too much CPU and memory. Finally, after an exhaustive search of the globe, it came back to me and said that false was false by definition. Roughly speaking, I had asked it "why not?" and it had replied "just because." Such is the miracle of modern hypothesis testing. I shudder to think of the query-debt my little joke built up.

Anyway, my screen was programmed to run the screen save until some non-trivial spike in the network traffic happened (say, more than 2 gig per minute). And we were expecting a big spike, real soon now. Meanwhile, we got on with the paper work. We've got expert systems managing our queries, that but still, some human has to make the policy decisions based on some very-illogical political reasons. John was arguing for a you-scratch-me and I'll-scratch-you policy. "Lets log the queries that extend our hypothesis space and favour transactions with sites that score high on that log.", he argued.

I wasn't too sure. "How does that sit with local control?", I asked. Local control is the official non-organising principal of TestNet. It takes a lot of resources to set up a TestNet node like us. The way hypothesis testing works, no one can remove an old hypothesis. Knowledge has to be write once/ delete never otherwise you can't test old premises. So, our basement floor is chock full of all our disc drives that store all the versions of our knowledge bases. No one wants to hand over those resources to some central body (remember, knowledge is power) so all the nodes in TestNet have guardians that control the in-coming searches. Once a query arrives, its up to the local site to control the inferencing associated with that query.

"Guardian could monitor our out-goings as well", argued John, "and he'd maintain the logs. We'd preferentially discount queries from sites that give us more CPU-share and better results when our queries hit their sites."

"And the network overheads associated with the monitoring?" asked Paula. John shrugged. "Compared to actually running the queries? Trivial."

I was surprised. John's shrug is usually louder and ruder. Especially with Paula. Don't know why they ever got married. Now that would be a good hypothesis to test: "John and Paula really love each other". Should chew up a few CPU cycles. Don't laugh. Ever since we cracked natural language parsing, we can process free-form queries like that. I mused on how it would be processed. First, TestNet would be explored for partially confirming evidence and I guess the marriage certificate would be found. Once some tentative evidence had been found for the proposition, then a larger search could be justified. Canberra's computers would be accessed for income tax and health records. The query would be processed by refutation. "Innocent till proven guilty" is the general hypothesis testing principal. Any statement can be added to the hypothesis space and remains unchanged until it generate contradictions or can be explicitly proven false. Unless some record could be found of marital discord (police records of bashing, spending sprees that could not be covered by their budgets, etc), TestNet would conclude that John and Paula really did love each other.

Not that a human would notice of course. The way they fight.

I remember the dull old days when hypothesis testing was confined to published papers. We used to have a floor of galley slaves that read the papers from all the journals, extracted an abstracted description, entered an executable version of that description in our hypothesis space, and then looked for any experimental evidence that referred to those hypotheses. Then, one fine day in 2003, someone added a natural language writer to their hypothesis tester and it wrote a paper outlining the discrepancies in what it had just been looking at, as well as some proposed fixes to the faulty model. The next step was obvious. Us humans stopped reading and writing the papers. Instead, we monitored changes in the hypotheses spaces around the world. Pretty soon all the hypothesis testing sites had a permanent background process running that reviewed the additions to the other sites. Rapidly changing hypotheses collected more attention as we performed our vulture act around new ideas. Sure, we still published. But we published active knowledge bases and hooked them into TestNet. I heard that Addison-Wesley once tried to sue TestNet. No one took much notice. Just the last kick of the paper dinosaur.

"But how do we know that the other sites are favouring us the same way?", asked Paula.

John went on to explain his ideas of "test queries", simple little assertions that any mainframe could answer in isolation. "We just shoot off a couple and measure the response time. The quicker the response, the more favoured we are." he explained.

"It shouldn't just be a time thing", she said. "Some sites spend longer because their hypothesis space is richer. I'd rather a query was explored thoroughly and not in some rapid-fire superficial way that missed an answer."

John disagreed but I only half-listened to them. I mused about what hypothesis testing had done to the scientific process (while keeping half an eye on my screen. Which still doodled to itself). I read in the history books that it used to take up to two years for a paper to go from be written to being published. The same cycle now takes about two minutes. What was that as a speed-up? 60 minutes an hour, 24 hours in a day, 365 days in a year, two years, divided by two minutes... I reached for the calculator to work that out when BING!, my screen lit up like a Christmas tree.

Paula and John instantly forgot their argument (it occurred to me later they were just amusing themselves while we were waiting). We all scrambled for the screen. Paula glanced at the new display of the TestNet traffic and bellowed out a gutsy laugh. "Its back on the air!", she cried. "Just when I said it would."

"China-B?" I asked. Paula's hands flew over the controls. The map of the world on screen spun a round and we fly into the Chinese mainland. The screen showed network traffic as vertical height. Over Beijing, a spike was forming.

China-B. It had to be. Two weeks ago, the Beijing TestNet site had announced a temporary withdrawal from the net. Sites did that from time to time to process the back- log of new hypothesis that couldn't be sorted out while the rest of the net demanded CPU time. The genetic algorithms crew call this "computer punctuated evolution". Which is a fancy way of saying its better to work on a new idea in a quiet room rather than in the middle of a crowd. Crowds tend to shout things down. Ideas grown in (temporary) isolation can develop the power required to survive back in the fray.

But China-B was special. Some of the intuitive jumps being made from that site were getting a little outrageous. Every site made a few little leaps, every now and again. We run our learning programs against our hypothesis space and sometimes discovered some new descriptors or abstracters. Usually the new ideas were only a small distance from the old hypothesis. But not at China-B. Some of its new ideas were really over the top. That one about whales talking to the dolphins convinced TestNet that some program had a diode loose over there in Beijing. Soon, net traffic to that site dropped. China-B then announced a two-week holiday. I thought that was a little odd. If a learning program goes screwy, I just amputate it without leaving the net. If China-B was withdrawing, then I felt that some new learning algorithm was being debugged and China-B wanted some peace and quiet to do some tests. So I checked the idea. I posted the hypothesis to the net and it generated no contradictions.

(Most people don't appreciate that TestNet never proves its' hypothesis- it merely reports failures to disprove. Experiments last century came to the reluctant conclusion that a closed-world assumption means you have to load all your common sense knowledge into the closed world. TestNet is more pragmatic. We just load in whatever knowledge source is a available and always check it when new data arrives. But the public doesn't understand. There was a famous interview on a day-time chat show where some business analyst was moaning about the way TestNet dribbles out conclusions to the rest of the world.

"You guys sit on information that could be of enormous national importance and keep us in the dark", he complained.

"Course we do" said the controller from TestNet Chicago. "Every time we let something out, the markets go crazy. Its scary to think that you guys trust TestNet's ideas so completely."

"And why not? TestNet is the most checked information source on the planet."

Chicago got really upset at this. She leant over the analyst and shouted at him, veins bulging on her neck: "Just because we can't refute something, doesn't mean its true!")

Anyway, when a new version of a learning program logs back on to the net, things usually go wild for a while till the rest of the world sorts out if it is genius or really just plain crazy. We watched the net traffic around China-B scream like a new born baby

"Its growing unusually fast." commented Paula.

I agreed. "Pull back a little. Lets get the big picture."

Paula flew us backwards from Beijing. We hovered some 30 clicks out of town and watched it grow. And grow. And grow.

"Ever seen anything grow that fast before?" John asked me. I shook my head and grabbed the second terminal. "Lets just get an average vector on the data transfer."

The phone rang. John took it. The face of the technical from the basement flashed up on the screen. "Hey," he asked," what's going on? We're on fire down here."

"Show us what you mean." asked John, glancing at me. I'd got my vector and a little alarm bell was ringing in my mind. The technical grabbed the handset and panned it over the disc drives. On the side of the drives are little lights that indicate the network traffic. Normally, write-once knowledge bases have a large redundancy rate. The rule-of-thumb was that the knowledge base at any TestNet site was only every 1% queried in any one day. That meant that most of our drives spun quietly to themselves without being queried. But now, all the access lights buzzed like angry red bees.

"Its China-B", I explained. "The vector?" asked Paula. I nodded.

"Why?" said John. "What's so special about our site?"

"Nothing."

"???"

"It's chewing up everyone's CPU. Paula, pull back and show us the Pacific picture."

Paula hesitated. "This is a hard movie to turn off", she commented. John and me turned back to the view over Beijing. China-B's spike now was towering over the city and reaching orbital heights. The spike's colours indicated the useful query rate. Queries resulted in hypothesis refinements were coloured gold. Queries that were timed-out before they terminated were coloured blood-red. Useless queries were black. Active queries were green. The shape of the spike is also coded. Width is the data transfer rate and height the derivative of the transfer rate. Brightness measures the average age of the queries: fresh queries are really bright while old ones are kind of dim. China-B's spike was a slender green shoot rearing up over the city, shooting skywards, and glowing with a neon brightness. Little gold flashes sparkled up and down the spike. I'd never seen anything so big.

"Never mind that now. You can watch the replays tomorrow." John said . "Take us up."

Paula snarled at John, then threw him the keyboard. "You fly."

John made a neat catch and pulled us out of Beijing. On the screen we flew upwards till the horizon was a circle beneath us. The whole Pacific region was on display.

"Cutting in access filter... now" said John, then gasped. The whole Pacific region lit in green cobwebs. Radiating out from China-B was more net traffic that we'd ever seen. Every node in TestNet was being flooded with queries like the ones at work downstairs.

"Now do you see?" I said. "There's nothing special about our sight. China-B is re-testing every hypothesis. Also, it seems its been hooked into a truly huge network over there. Response time over at Beijing is currently averaging 11.3 seconds." Paula looked startled. "But that's a tenth of ours and they're handling all that traffic."

"I know", I said, "but for some reason Beijing wants the world to use China-B. They've given it massive resources."

"Now why do they want to do that?" mused John. "Why be so nice to the rest of the world? Bit suspicious, don't you think?"

"You're not suggesting that China-B is trying to LIE to the net?" asked Paula sarcastically. I winced at her tone but agreed with the sentiment. Once upon a time, a lot of people were worried that TestNet would be used to spread disinformation. My PhD thesis was a hush-hush DoD project to try and lie to the net. Didn't work. Every lie I produced generated so many inconsistencies that traffic to my node dropped off to zero. My general conclusion was that the effort required to generate a good lie was equal to the effort required to enter all the knowledge into the net. Nowadays, no one worries about lies. But still, China-B was acting very strange.

My computer went PING! and a little dialogue box popped up.

"Priority query returning." said John, reading the screen. "You got any ghosts running?"

I shook my head. "What's it say?"

John read the fine print. "It's a funny one: False is false by definition."

Two surprises in one day. "You're kidding. That ghost terminated years ago."

"Well its back on the screen now."

Then Paula made the statement that made me very rich (yes, yes, I paid her a cut).

"Now why would China-B have reset it's knowledge base?"

We asked her what she was talking about. No one reset KBs. TestNet has to be a write- once system. Etc. She waved all our remarks aside. "John, give us a read-out on the CPU and elapsed time on Tim's ghost."

"Four years in CPU, twenty years elapsed."

Paula leant over to me. "When were you an undergrad? 2010? Thirty years ago? That ghost has lost ten years of its elapsed life. China-B has reset."

"I see what you mean." I said. If a TestNet site reset its hypothesis space to some point in the past, then when it came back on to the net it would have some catching up to do. However, the catch-up would be relatively fast. Many of the hot topics current at the reset point would be resolved during the "lost" time. The catch-up would really be just a run around of the other TestNet sites asking them for the answers. It would generate a query pattern much like the one we were seeing here. China-B was now getting all the solutions to last decade's problems: the aids vaccine, the hyper-space drive, and which religion was true.

There was a catch, however. Simply throwing away a decades thinking would be like amputating with a chain saw. Messy. We'd notice the wound. China-B would have to set off a whole suite of demons to patch the wound. "And that would take weeks", I said.

Paula nodded. "China-B's been down for weeks. I bet you they spent that time resetting and vacuuming their KB. Your little query was so low priority and so unpopular that their consistency demons just missed it." She pushed John away from the terminal. "Lets do a little hunting. Can we come off the net for a while? I may need the CPU."

There's no such thing as "spare CPU" in a TestNet node. Every time we upgrade the machines, some smart-arsed postgrad comes up with a better learning program that chews up more processor time. Also, each upgrade makes us more attractive to queries from other sites. And the number of TestNet users increases exponentially each year. Paula had a hypothesis to check and she wanted to get all our OS queries off our machines. I didn't know what she was up to but she's pulled a few rabbits out of the hat before. I grabbed the phone and called down to the Technical. Told him we were coming off the net.

John leant over and said "What do you need?" I have a theory about those two (which I'll never put onto the net). They only stayed married because they can navigate TestNet better than any other pair I know. I sat back to enjoy the show. This could be good.

Paula explained. "I'm looking for a China-related hypothesis that is true now and wasn't true before China-B came up."

John nodded. "Something that could be disproved in a month or two, and China-B would still benefit."

"Check" she said. "Generation-one should relate to current motivations of China-B. What's their current problem space?"

Paula and John bent to work. I began to get the idea. They thought China-B was trying to flood TestNet with some mis-information. But to do it, they had to bury it in a mountain of "new" conclusions. By the time we sorted it all out and found the inconsistency, it would be too late. But too late for what?

The generation-one queries were spreading out over our site. Hypothesis testing is mostly spent trying to come up with an appropriate language for describing the queries. A query language is designed, used, and the usage monitored. A few test queries are posted and these suggest promising lines of inquiry. Or they don't. In which case the language is modified (maybe by a human and maybe by a learner).

John and Paula had posted some queries and were watching the results. A background demon popped up and reported the estimated cost of the extra resources now cut into China-B.

John whistled. "This is big. Major league government expense."

Paula snapped her fingers. "Government! Modify generation one. Change motivation search from China-B to Chinese government."

"Gotcha" said John and set to work to modify the descriptors in the generation-one language. The new query was posted and the results were much shorter this time. This encouraged them to work on and devise a generation-two language especially for small- scale temporal events.

I went for coffee. Ten minutes later, I brought back two cups for them and placed then by the terminal. They drank the coffee, but I don't think they even noticed. I watched as these two maestros reduced the search space. Another demon woke up and suggested that agricultural goods could be a useful descriptor in the queries. John quickly whipped up generation-four of the query language. (Generation-three had come and gone while I was at coffee. Some dead-end to do with weather variations).

Then, Paula nodded. "Enough private thinking. Lets get back on the net." I rang the technical and got us re-installed. We got a glimpse then of the havoc China-B was causing. We did a quick spin around the planet and green tendrils showed everywhere. Suddenly, over the horizon, a huge spoke came flying towards us. John grabbed the joystick and banked us right. The China-B spike (now a huge thick tree jutting out over much of northern China) swiped by us. For a moment, our screen was full of gold stars.

"That makes me giddy." I said.

"Hey, look" said John, "Tokyo is down." On the screen, the TestNet picture of Tokyo was a white gap.

"Now why would they be off the net?" I asked.

"Same reason as us?" wondered Paula.

"Lets give them a call." I suggested. I got their telephone number and contacted the Tokyo controller.

"Hail Emu!" she said. "How do you like the show?"

"Pretty lights." said John.

"Pretty suspicious if you ask me," said Tokyo. "We saw you going on vacation. Good to see you back. Any ideas?"

"You show us yours and we'll show you ours." said Paula. "Got a land-line?"

Tokyo and Paula arranged a language swap. They were slightly ahead of us. Their generation-five and our generation-four descriptors combined and we ran the new query. No luck. We both cut out of the net to give our learners a chance to chew on the generation-six language. Soon our screen was full of nouns and verbs all to do with crops.

"What is all this?" asked Paula.

"Looks like learner wants us to check out the farms." said John. "Lets give it a shot..."

Generation-eight was all based around farming crops. We wrote some queries and set them off in. Five minutes later...

"Wheat" said Paula.

"What?" I asked.

"Wheat." she explained. "China-B has reversed the sign on an equation that relates to the nitrogen uptake of a new wheat fertiliser used in the southern provinces. That's the lie."

"Sure?"

"Damn positive." she said. "And its an obscure a reference as you will ever find."

"OK." I said. "John, get some models going. Predictions of crop yields for wheat with that sign negated... Oh, already done?"

"You betcha", he said pointing to the screen. "Someone's made a BIG mistake. Yield down 30%. Real famine stuff. China-B wanted to hide it for a month while it bought up big on the world market."

"And in the meantime, they're telling us that they have an over-production so the prices drop. " I said. "Then, in a month's time when the lie surfaces, they've bought their wheat at a depressed price. Clever."

"So what do we do?" asked Paula.

"What do you mean?" asked John.

"We can't just tell everyone." said Paula.

"Why not?"

"Because China needs it's wheat. If the prices go up..."

And so the next John/Paula war started. Paula wanted to squash the finding. John said that it couldn't be done; that Tokyo had our generation four language and learners just as good as us and that they'll work it out any minute now.

Me? I kept silent but I agreed with John. You can't lie to TestNet. Its too big. Just maintaining truth was hard enough, let alone a good lie. Sure, we had a head start from my old ghost but other sites would catch on soon enough. In fact, over John and Paula's shoulders, I could see more and more white gaps on the screen. All over the globe, sites were going off the net (no prizes for guessing what they were working on). China will just have to negotiate a wheat deal. And check its fertilisers better next time.

I rose to go. Paula stopped me. "Where are you going?" she demanded.

"Well..." I said reluctantly, "you guys are making such a racket and I need to make a phone call."

To my stock broker. Like I said, Monday was the day I got rich. Buying wheat.