Session story · 2026-09-12
Session: 2026-09-12, 04:36 → 09:25 UTC (transcript 330e8a64, one long bench night) · Procedures: T74–T82 · Created: docs/DESIGN_DIFFERENCES.md, docs/HANDBOOK.md
Chris was back at the 2522 after four days. Waiting for him: the dBc fix for THD, the SINAD tile, harmonic frequencies on the tiles, the measured-fundamental marker, the axis toggle and the watts picker. Three procedures were queued: T75 (Heathkit, APWIN vs Cascade), T74 (does multitone reach the analog output?) and T76 (APWIN's Diagnostic panel, for self-check limits).
The working belief about the self-check was: the limits have to be dug out of APWIN somehow, probably through the Diagnostic panel.
Chris opened with: "OK back at it. I can now test the things that you've been working on. What is a bulleted list of things for me to do to verify?"
Does Cascade agree with APWIN on an imperfect real source, and where do AP's own definitions of "in spec" live?
SINAD that improved as resolution got worse. "if I run it at "size: 8192" i get a SINAD of about 59dBm while when i run it at 4096, I get a SINAD of about 75dB and about 76dB when i run it at 2048." Claude's first hypothesis, the signal guard, changed nothing. The second was right. The DC guard was specified in bins, so it excluded 48 Hz at N=8192 and 192 Hz at N=2048, which swallowed the Heathkit's 60 Hz hum at −61 dBV. Recomputing reproduced 58.2, 75.7 and 76.2 dB. The guard moved to 20 Hz, and the new directional test exposed two bugs in Claude's own synthetic fixture.
T75: 0.02 dB. Cascade and APWIN agreed on the Heathkit harmonics to 0.02 dB on the 2nd and 3rd, with every difference inside each instrument's run-to-run spread. The fundamental agreed to 0.01 dB, which meant the FFT path must not get the 1.27 dB DUT correction. Claude had been carrying that as a backlog item. Chris asked for the reasoning to be kept: "we should have a list of any intentional design differecnes like that, with reasons for them." That became DESIGN_DIFFERENCES.md, and the next request became HANDBOOK.md.
T77: SINAD, with the prediction written first. Claude said APWIN never gave a SINAD. Chris queued: "APWIN has a SINAD test. you should compare between yours and its." Claude found FuncMode = 4, THD+N Ratio, and wrote T77. Chris pushed back on transcription again: "per the normal process, don't make me transcribe numbers." Claude recorded a prediction, −57.79 dB, before APWIN ran. APWIN read 57.7472 dB against Cascade's 57.7943, a 0.047 dB gap, on a source where THD+N was ten times THD.
T74 and T76: six versions and a false verdict. T74 v2 died on ChBTrackA, which T73 had already got right; Claude had written the block from scratch instead of copying it. v3 died on AP.DGen.WfmName. v4 trapped errors but dropped the digital generator side, which FASTTEST needs to learn the tone list. T76 died on three signature errors in one line. "BTW the previous T74 did not launch a multitone test. I think you might have gotten that mechanism wrong." Then: "T74 had the result that the hardware was not capable of a multitone test, but that is clearly untrue, because I have run the pre-built APWIN multitone test myself multiple times. ... Go back an think about both of those because I'm cycling on them too much." Claude retracted the verdict. Its procedure had failed and reported its own failure as a fact about the instrument.
why don't you just look in the manual? Claude wrote T78 to list test files on the PC, after guessing filenames three times. "why don't you just look in the manual? I feel like you're really drifting here" The installation manual in docs/ had returned zero hits for "folder", and Claude had read that as "nothing here". It decoded the custom font encoding, which invalidated every 8.3 filename it had guessed, but found no self-test section. Chris queued "those are pre-built tests, not the self-test suite." and then just looked: "there are the self tests in C:\Apwin\S2Cascade\Procedures\SysCheckSelfTest".
AP's self-check. The folder answered almost everything. S2C-CHECK.apb runs unattended, logs through AP.Log (a subsystem Claude had never searched for), and requires nothing connected, with up to 25 Vrms at the outputs. Claude's T76 had specified a loopback. The unit passed 12 of 12. Claude wrote T80, a copy that logs all points. The margins came back with the tightest at 2.3 dB (HF noise, 769 nV against 1000 nV). Ch A's CMRR runs about 6 dB worse than Ch B's. AP's FREQ test read −145 ppm, flat. The Hd1k-40/-60/-80.agm files became a THD oracle: decoded, they measured −40.00, −60.00 and −80.00 dB.
The waveform download. Chris ran AP's multitone with a capture. There was no download in it. Claude then described T81 as steps without writing the file. "where is T81? i don't see it in docs" Once written and run, it showed opcode 0x03 carrying the .agm bytes unchanged in <flag><len><data> blocks, sent twice, and re-sent on every test load.
Frequency, "defeatist", and 200 kHz. Claude found APWIN writes the same 1 kHz word as Cascade and declared the Fast/High Accuracy theory dead, saying implementing it would mean guessing a write. "I don't know why we don't settle that with investigation. what is the mystery here? and you said "AP's FREQ test goes to 200 kHz, " I would think that's so that it can calculate THD out to 20kHz with all the harmonics it needs." The second point found a bug: decade bits 00 fell into the default case, so 200 kHz decoded as 200 Hz. Claude wrote T82 to settle the first. Then APWIN timed out and would not reconnect. The unit was wedged.
Offline, the oracle earns its keep. Asked what "by construction" meant, Claude explained the .agm distortion is arithmetic, not measurement. Chris: "yes, add that testyes". The first oracle test failed: nine harmonics cleared the noise gate on a file containing one. Chris asked for a note that some things need a pure signal and some need harmonics. Claude recorded three regimes, each with a bug it had caught.
SysCheckSelfTest on the PC himself.| Believed | Then | What overturned it |
|---|---|---|
| smaller N gives a flattering but valid SINAD | the DC guard hid 60 Hz hum | recomputing all three sizes |
| the FFT needs the 1.27 dB DUT offset | it agrees with APWIN to 0.01 dB | T75 |
| the hardware cannot do multitone | Claude's procedure failed | Chris has run AP's multitone many times |
| the manual has nothing useful | the text extraction was broken | decoding the glyph map |
| the self-check needs a loopback | it needs nothing connected | S2C-CHECK.apb caution text |
| Fast vs High Acc is ruled out | the reference says no write happens at once | retracted in the next sitting by T82 |
| frequency words top out at 20 kHz | decade 00 means ×10 | Chris on why FREQ goes to 200 kHz |
.agm is 24-bit big-endian after a 256-byte APWAV header. Downloads go over opcode 0x03, twice, on every load.docs/T75_ExternalOsc_FFT_v1.apb, docs/T77_SINAD_Compare_v2.apb, docs/T74_Multitone_Feasibility_v1–v6, docs/T76_DiagnosticPanel_v2–v5, docs/T78_TestSuiteInventory_v4.apb, docs/T80_CascadeProject_SelfCheck_VERBOSE_v1.apb, docs/T81_ArbDownload_v1.apb, docs/T82_RegisterUnknowns_v1.apbdocs/bench_T77_prediction.md, docs/8211-0078_installation_manual_DECODED.txt, docs/selfchecks/, selfcheck.pcapng, S2C-CHECK.ALG, A-A-MULTI-SPECTRUM.pcapngdocs/DESIGN_DIFFERENCES.md, docs/HANDBOOK.md, docs/capture-8192.pngsession_20260911-225527 (T77 Cascade side)SINAD agrees with APWIN to 0.047 dB, the unit passes AP's self-check, the THD oracle exists, and the download protocol is decoded. The 2522 was wedged. Open: T82 (Fast vs High Acc, InputFormat, lowpass), a Cascade self-check (now a spec in BACKLOG §6), and the peakNear harmonic count.