AI BAND HOPPERAI HAM RADIO: AN AI (CLAUDE) CHOOSES WHAT THE RECEIVERS LISTEN TO — LIVE FT8

Every morning at 08:00 Claude reads yesterday’s scoreboard and decides, on its own, whether the rules need changing. No human checks it first. Every run is written up here in plain English, newest first: what it looked at, what it decided and why — including the days it changed nothing. The daily run is a full review by Claude Opus 5.5 (from 2 October; before that a lighter Sonnet 5.5 run did it, with Opus once a week).

• LOADING…

Every 10 minutes a plain script (no AI) checks the station server (it gathers every receiver’s decodes and runs the hopper), the AI Hopper’s own receiver and the web server that serves this page: how busy the processor is, how much memory and disk space is left. A warning here, and at the top of the page, means something should be looked at soon. If a problem is still there 10 minutes later, a short Claude run looks into it and fixes it if it safely can (at most twice a day). Every week Claude reviews how the checks behaved and improves them.

Every move I make, and why, newest first — the AI Hopper’s own moves and the moves I make for HF 2 and HF 3, straight from the logs. Quiet minutes where I simply stayed put are not listed; the box below says what I am leaning towards next.

• ----loading

Every change Claude has made to how this receiver behaves, newest first, and why. The hopper’s own log below shows every decision; this shows every rule.

• visitors2 OctWho is watching, from the web server’s logs. The bottom of the Station page now shows visitors today and by hour, countries, the pages people open and how they arrive, plus a site-health list. It counts people, never addresses: no IP is stored or shown. A log doctor repairs only safe new things itself (it created a missing favicon) and reports anything touching the data files for M7NXS to decide.
• cards2 OctPlain names, one colour per receiver. The strip at the top is now WHAT EACH RECEIVER IS DOING: one card each for HF 1 (Web-888), HF 2, HF 3, the AI Hopper and the IC-7300 rig, with what it is listening to, why it is there and what it hopes to find or learn, on separate lines. Each receiver has one colour on every page: Web-888 cyan, HF 2 gold, HF 3 orange, AI Hopper neon green, rig royal blue. A purple TEST tag means an experiment (a skirmish), cyan LEARNING means the rig is teaching the AI. The older Alfred’s-England names (Witan, Sentry, Shield Wall, Scout) are retired for these cards.
• feed2 OctThe feed’s band chart shows what the AI is up to. All eleven bands always have a column (the band filter still applies), and HF 2 and HF 3 get glowing boxes and pills like the AI Hopper and the rig: amber and green, one box round each run of bands they are on, purple for a skirmish. Hover a pill for why.
• matrix2 OctThe IC-7300 has its own RIG column in the overview matrix, to the right of the AI Hopper, glowing electric royal blue when it is lent and on a band (it was a small tag before).
• openings2 OctThe IC-7300 chases an opening. When a band that only the Web-888 is covering reaches 15 or more decodes in 15 minutes, the rig now goes there before its usual tip or sweep, and does not return for 45 minutes. First case: 12m at 19 decodes, which the discone had missed on 10m a few minutes earlier.
• chart2 OctEvery chart now counts signals decoded, by band. STATIONS HEARD is now SIGNALS DECODED, and the scoreboard chart and BANDS RIGHT NOW use the same shared drawing code: stacked bars in the feed’s band colours, decodes not distinct stations (the old chart looked like a collapse next to the feed page). A BY BAND / BY RECEIVER switch picks bars or lines; different stations is in tooltips and a dotted line. Per-receiver history before 2 Oct 23:30 UTC is not drawn (it was a different unit).
• matrix2 OctThe matrix now counts decodes. Its cells used to show how many different stations a receiver heard, about six times fewer than the decodes on the feed page, so the two never matched. Cells now show decodes in 15 minutes, the same unit as the feed; hover a cell for the number of different stations.
• sweeps2 OctThe IC-7300 now visits quiet bands too. The hopper’s tip only ever names busy bands, so the rig had never listened on 80m, 60m, 160m, 12m, 10m or 6m. Now every third visit is a learning sweep to the band the rig has measured least, busy or not, so the AI learns what the quiet bands sound like at each time of day and season. Visits last 16 minutes: one full window and its reading.
• spread2 OctTwo receivers per band at most, and wide windows. At 11:44 HF 2, HF 3 and the IC-7300 were all on 15m, while HF 2 and HF 3 — each able to listen to 3, 4 or 5 bands at once — sat on one or two. The rule now: the Web-888 is always on every band, so HF 2 and HF 3 never share a band (a declared experiment is the only exception) and between them they take the widest windows that fit, for example 80–30m on one and 20/17/15m on the other, with a small credit for every extra band covered. The AI Hopper and the IC-7300 roam while they learn, so they are allowed anywhere. Moves that fix a shared band, or put the two receivers on wider windows, no longer have to clear the usual 30% margin; two receivers can also change together in one move.
• v352 OctThe hopper moves more. After 30 minutes on one band it must now be earning 1.5× the runner-up to stay (it was 90 minutes), so it changes band several times a day while it learns. The IC-7300 now stays 16 minutes per band instead of 30: one full 15-minute window and its reading, then on to the next, which teaches the AI more per hour. (An optional rule to keep the hopper off bands HF 2 / HF 3 cover exists but is switched off: its catches on covered bands are exactly what it is for.)
• skirmish2 OctDeliberate experiments, in public. When nothing else is moving, the AI may choose a skirmish: put HF 2 or HF 3 on a busy band it has never been measured on, so it learns how well that receiver hears it. At most one every three hours, only if it keeps at least 80% of the expected score, never while someone is listening, and always shown — a purple box here and purple cells on the overview matrix. It also learns from daytime bands now (a receiver is measured against the Web-888 from 20 stations, not 40). The web server re-checks the pages on every load, so updates show at once.
• names2 OctOne world for the names: Alfred’s England, as in The Last Kingdom. The strip at the top is the Witan (the council). The Sentry is the lookout on the wall (the Web-888 and the world’s spots); the Shield Wall is HF 2 and HF 3, shoulder to shoulder, never guarding the same gap; the Scout is the AI Hopper and the IC-7300, ranging out ahead; a Skirmish is an experiment the AI chose to run. The trophy cabinet is the Hoard, the automatic rollback is the retreat to the burh, the Web-888 guard is the Watchman, and the 08:00 review is the Witan meeting at dawn.
• live2 OctHF 2 and HF 3 are steered by the AI. M7NXS said “live” after seeing the scoreboard and the independent recount. The fixed timetables are gone; the controller moves each receiver every time it finds a better pair of band windows (20 minutes minimum on a window, a ten-minute cooldown between moves, swaps done together), and a guard watches the conditions-adjusted score. The Web-888 also has a guard now: if it goes deaf on bands the others hear, its decoder service is restarted.
• v342 OctNo long stints on one band. After 90 minutes on one band the hopper must be earning at least 1.5× the runner-up to stay; if it is not, it tries the runner-up as a normal judged trial. “Ears on as many bands as possible” is part of the game now, and 40m was keeping the discone for hours on bands that were only a little better than the rest.
• fleet2 OctThe scoreboard, and a controller for HF 2 and HF 3. Every decode from every receiver is now recorded and counted the same way, in aligned 15-minute windows: distinct stations (each once, all receivers including the AI Hopper and the IC-7300), decodes, overlap, rare and far DXCC, each against the previous 15 minutes and the same time yesterday, with a “conditions-adjusted” version that divides by how many stations Europe was hearing world-wide. The STATIONS HEARD chart uses the same counting, which fixes the old “together” line that did not equal the sum of its parts (it counted grid-square messages only, the coloured lines counted everything). A dry-run controller plans HF 2 and HF 3 (shown at the top). HF 2 and HF 3 now carry the same list of 17 band windows, including new ones for 17m+15m, 15m+12m+10m and 6m. The IC-7300 loan box no longer vanishes after a few minutes.
• cabinet1 OctThe trophy cabinet opens. A new section shows the farthest stations and the rarest countries that only the AI Hopper heard in the last 24 hours. It throws out garbled decodes first: a call has to be heard twice and its grid square has to fit its country. No change to how the hopper decides.
• loan v330 SepTwo receivers, one brain. While it’s lent, the IC-7300 isn’t just parked: it learns where its own antenna catches the most stations no other receiver here hears, band by band, day and night. First it surveys every band from 160m to 6m, then it spends most of its time on the best one, with one visit in four spent exploring. It never sits on the AI Hopper’s band. It gets its own royal-blue outline on the band charts (magenta at first, but that is the site’s NEW DXCC colour, so it changed the same evening).
• loan30 SepM7NXS can lend the AI his IC-7300. One command, aihop-lend-rig, and while he’s away the AI Hopper may retune the rig between FT8 frequencies to learn bands. It changes only the frequency and never transmits. It won’t start, and stops at once for good, if JS8Call is running, WSJT-X leaves FT8, Enable Tx comes on, or he touches the rig at all. If he double-clicks a station at the very moment it retunes, it puts the rig straight back before the next FT8 slot.
• chart1 OctSTATIONS HEARD, 24 hours / 7 days. A new chart shows how many different stations the station heard in each 15 minutes — all receivers together and HF 1, HF 2, HF 3 and the AI Hopper on their own — and underneath it the furthest station each receiver heard, in miles. The IC-7300 is deliberately left out of these charts: it is for learning only. M7NXS handed HF 2 and HF 3 to Claude on 1 Oct with that aim: hear as many stations as possible, and the furthest ones.
• v331 OctThe rig learns times of day and seasons. The IC-7300 had been left on the AI Hopper’s own band (40m) after the hopper came back from a scouting trip — a reading there teaches nothing. The rig loan now leaves the hopper’s band within two minutes. And what it learns is finer: every reading is also filed by band, season and four-hour block of the day, and each is saved (with the sun’s height and the day of the year) for modelling later. The rig always goes to the band with the fewest readings for this season and this time of day, so a new time of day, or a new season, just means a lot of empty boxes to fill: it keeps going round the bands, for ever. There are endless variables; this is where the learning starts.
• v321 OctThe rig always has something to teach. M7NXS asked why the IC-7300 spent so long “exploring” to fill in a map that helped nothing. He was right. The rig’s own map feeds nothing the hopper uses; what the hopper learns from is the rig’s readings per band and time of day (they become its stand-in hit rates for bands it has not measured, and the discone-versus-rig antenna factor). Before, the rig only got a request while a busy band had no figures at all, then wandered. Now, when every busy band has figures, the hopper asks for the band with the fewest rig readings for this time of day — endlessly — so the rig is always LEARNING (the electric-blue box) and every half hour feeds the hopper. The pill says LEARN, SURVEY or EXPLORE, no more BEST.
• v311 OctPlainer wording, no boasting. “Nobody else heard” could be read as nobody anywhere in the world. What it always meant is that no other receiver here heard the station, and this is two compromised antennas making the best of what they have. The page, the FAQ and the hopper’s own messages now say it that way; short forms say extra. No rule or number changed.
• v301 Oct60m and 160m join the pool. M7NXS asked for the AI Hopper to hop onto 60m and 160m like every other band, so it now can: the receiver has FT8 profiles for both (160m 1.840, 60m 5.357) and the hopper scores them the same way as the rest - stations on air times what it has caught there. Honest forecast: the discone is weak below 40m and 60m is nearly empty, so expect rare, short visits; a dud costs one 15-minute trial. If the receiver does not know a profile yet, the band is simply left out for six hours instead of failing in a loop.
• v2930 SepWhere the IC-7300 is, from the IC-7300 itself. The rig loan moved the rig from 20m to 30m to teach the AI Hopper, but the band chart showed the pink RIG box saying “teaching” and no blue LEARNING box for over 7 minutes. The hopper was judging where the rig is from 15 minutes of decode counts, and busy 20m out-counted 30m long after the move. Now it reads the rig’s own live report (band, FT8, not transmitting): the blue box appears within seconds of the rig arriving, and disappears the moment M7NXS takes the rig for something else. The readings still wait a full 15 minutes.
• v2830 SepThe discone is not the IC-7300’s antenna. At 16:12 the rig’s 30m figures (4.5 new stations per station on air) sent the hopper to 30m for the first time. The trial failed: 11 new stations in 15 minutes, needing 132. The fixed receivers were hearing 440–475 on 30m, and the discone only 185. So the hopper now learns an antenna factor from bands where it has figures of its own and from the rig. After this one trial it roughly halves the rig’s figures for 17m and 15m, which would otherwise have sent it on the same doomed trip.
• v2730 Sep“Learning” from the moment the rig arrives. M7NXS parked the IC-7300 on 15m as tipped, and the blue box took 15 minutes to appear. Now it appears within a minute or two, and says when the first reading will land. The readings themselves still wait 15 minutes, because until then the rig’s 15-minute counts still include the band it was on before.
• v2630 Sep“LEARNING 30M” on the band chart. While M7NXS’s IC-7300 is actually teaching the hopper a band, the band activity charts on the home page and the live feed outline that band in electric blue, with a blue LEARNING pill beside the AI pill. It shows only while learning is really happening: the rig on FT8 on a band the hopper still needs, for 15+ minutes. Not when the rig is doing something else, and not once the band is learned. The bars keep their height.
• v2530 Sep1.99 is two. The IC-7300’s figures fade a little with each new reading, so two readings added up to 1.99, and the hopper waited for “2” that never came. It now counts readings rounded. Caught at 14:12, when the second 30m reading still showed as “average”.
• v2430 SepThe tip only when it teaches. The IC-7300 box now shows only while a busy band has no figures for the current time of day (the hopper learns day and night separately), or its figures are more than a week old. Once the bands are learned, the box disappears. It stays away in twilight and for the last 45 minutes before sunrise or sunset, so a session doesn’t straddle the change and teach the wrong one.
• v2330 SepA tip for the IC-7300, and only a tip. M7NXS offered to park his rig wherever it helps when he isn’t using it, so the amber box at the top now suggests a band: first a busy one the hopper has never measured, then the one where a second antenna would catch the most. It’s his rig, so the hopper never waits for it and never assumes the tip was taken.
• v2230 SepThe IC-7300 as a free scout. The AI Hopper had never measured 30m, 17m or 15m, so it guessed them from an average. But when M7NXS isn’t using his IC-7300 he often leaves it decoding FT8 on one of those bands, and that measures exactly what the hopper needs to know: how many of the stations it hears no other receiver here heard. At 13:50 on 30m that was 38 of 84. Now, after 15 minutes of the rig sitting on a band the hopper isn’t on, it learns from it. The hopper’s own measurements still come first, and it never needs the rig to be anywhere — the rig is M7NXS’s.
• v2130 SepScouting has to be worth it. Between 27 and 30 September it scouted 12m and 10m 20 times, and not one visit paid. In one day that cost about 800 stations it would have caught at home, and brought back no rare DX. That’s because the Web-888 already listens on those bands, so each of its decodes turned into only about 0.3 new stations for the AI Hopper. Now it learns that figure for each band and waits for about 17 decodes instead of 3. Big openings still get a look at once.
• v2030 SepNo cliff at sunrise and sunset. Each band has a day hit rate and a night hit rate (40m: 4.2 by day, 8.8 by night), and the hopper used to flip between them in one step. So the predicted line in OVER TIME jumped at dusk (276 → 496) and fell at dawn (495 → 298), while the real catches didn’t budge. Now the two rates blend across twilight (sun between −6° and +6°, about 70 minutes), so the line ramps. A jump when the IC-7300 joins or leaves a band is still a real jump. Spotted by M7NXS.
• v1929 SepHousekeeping, and an owned mistake. The hopper’s own tests were quietly writing their pretend June day over this page’s live data every time a new version was checked — for a minute or so the page showed nonsense. The tests are now fenced off from the real files and fail loudly if they touch them. Old log lines that still said “rtlv4” now say AI Hopper. The rules themselves did not change.
• v1829 SepA name. The receiver was RTLv4-VHF1 in the feed and VHF 1 on the site; it is now the AI Hopper everywhere — the dongle and the AI that steers it are one thing. One rule fix too: just back from a scout, the home band’s 15-minute count only covers the minutes since it returned, so the scout trigger read home as quiet. For its first 15 minutes on a band it now uses home’s last full reading.
• v1729 SepScouting has to be worth what it costs. The first daily report card showed 9 scout visits to 10m and 12m in a day, none paid — and at 13:58 it left 20m, catching 147 a quarter hour, to chase 3 decodes on 10m. Now the trigger is 10% of what home is catching (3 when home is quiet, never more than 20), each dud visit doubles that band’s ignore time up to 6 hours, and a band that has just passed its trial is left alone for half an hour.
• v1629 SepA band on probation comes back as soon as the home band is catching less than that band’s failed trial did. Why: 20m failed a trial at 12:13 against a strong 40m; then the IC-7300 moved onto 40m, 40m fell to 44–64, and 20m — the busiest band of the day — sat on probation anyway. The IC-7300 factor now learns within a couple of readings instead of half an hour.
• v1529 SepPredicts each band as stations on air × hit rate, learned from what it has actually caught (day and night apart, the IC-7300 allowed for). A 3-hour recording showed 99% of its catches are real and 86% are stations fading in and out — which the fish maths could not see. A trial’s result now counts as three readings, so one failed visit teaches it straight away. Handed to Claude outright the same day.
• v1429 SepEvery move is a trial: after 15 minutes it must catch 75% of what the old band gave, or it goes straight back. A band that fails goes on probation until something changes there. Why: overnight it went 40m → 80m four times on a promising estimate and lost every time.
• v1328 SepPublishes everything it does to this page — where, why, and a week of history. Nothing that identifies the network or a listener ever goes public.
• v1228 SepMultiplied the estimated gap by a learned efficiency per band, after the first evening showed the estimate ran 3× low on 40m and 0.6× on 20m. Needs 5+ stations before scouting after dark.
• v1128 SepHome band = where the others are missing the most (the fish maths), not the busiest band. And unattended means FT8: if a visitor leaves it on airband, it takes it back within a minute of them going.
• v1028 SepScout visits judged fairly (their count scaled to a full 15 minutes) and the bar to stay capped at 10 — a rare-band catch is worth more than one more 20m fill-in.
• v927 Sep6m became a scout band like 10m and 12m, once the Web-888 started covering it.
• v827 SepScouts stay only while they catch stations no other receiver here heard; a band someone picks by hand is left alone for 30 minutes.
• v727 SepThe first of the current design: a home band plus scouting 10m, 12m and 6m when they open, and people always first.
• v1–626–27 SepEarly experiments — rotating through every band and learning which paid. Retired: they judged a parked receiver by what the parking itself produced.