Activity

1. gHashTag commented on Oct 19, 2026

   

   gHashTag

   on Oct 19, 2026

   OwnerAuthor

   More actions

   Draft of the night report (the final version lands as docs/analog/night-report-2026-10-19.md when the lanes still running have finished; the markers @@pending@@ are filled then). Kept here so the draft survives a session restart.

   Night report 2026-10-19: analog emulation, lab, T27 OS

   Issue: #11062, epic #11063 (ledger comment 6268998904 on the epic). English, ASCII only. Every number below is copied from a test, a receipt, a PR or an issue comment named next to it; a number without a source is not here.

   1. What was asked

   The owner's directive for the night: research the weak points of the task and its competitors, decompose, implement, end with a report and three collaboration options for the next loop; run autonomously; one continuous loop that never breaks earlier work; self-critical, find anomalies and repair itself; new tri commands and skills; every task a GitHub issue; coverage to 100 percent; auto-fix on and merge when green.

   2. What merged tonight

   PR	issue	what	measured
   #10893	#10777	PSP103 drain and gate charge tables (tt, ss)	500 fF inverter delays within 0.020 % and 0.0041 % of ngspice (#11071, W3)
   #10953	#10888	M4e-offset: the converter offset is the comparator's clock-edge kickback read through unequal top plates (sar_kick.t27)	base +9.42 mV, kickback removed +0.03 mV (PR body)
   #11055	#11020	tri remote: reference jobs on a remote Docker host, command lines built by a spec	16 tests, refusal tests for every shell byte, hand-written Rust +21 -441
   #11056	#11054	the rule "request, do not ask" in CLAUDE.md and the loop skill	2 files
   #11066	#11041	lab L1: the t27c lab's decisions as .t27, lab.py as the oracle	5 specs, 66 tests, 1474 pinned asserts, 59 mutants killed
   #11067	#11043	LAB_GATE as a network job kind run by T27 OS nodes, test-network XP	21 tests, 3 invariants, 335 runtime asserts
   #11069	#11044	one installer built with make, README, TRI CLAW as a component	16 tests, 12 of 12 mutants killed, Makefile +40 -41
   #11071	#11058	the ten analog weaknesses against measured evidence	0 closed, 8 reduced, 2 open
   #11072	#11057	lab competitors, prior art and weaknesses	16 systems, weaknesses W1 to W8, three options
   #11075	#11045	tri seals drift names the claim that moved and reads the ledger in groups	see 3
   #11076	#11059	tri loop plan: the tick's anomaly, idempotence and requests-out decisions	25 tests, hand-written Rust +23 -40
   #11079	#11042	lab L2: the t27b lab's decisions as .t27	5 specs, 84 tests, 1392 cases, 46 mutants killed
   #11080	#11068	tri land: land a lane branch with the rules in .t27	26 tests, 29 steps, 18 of 18 planted mutants killed
   #11082	#11074	AX7203 board regression of both analog tops	see 3
   #11084	#11083	the home path in the board document	1 line
   #11095	#11060	tri learn: save a lesson into a skill once	21 tests
   #11096	#11061	tri coverage: five facts per spec	18 tests, five specs edited
   #11092	#10942	partial: four ngspice decks for the comparator kick	no spec yet, see 6

   @@pending@@ states of #11092, #11095, #11096 and of the lanes still running at the time of writing.

   3. Measured numbers

   • Board regression (PR docs(analog): board regression of the M1b and SAR tops on master -- both receipts reproduced on the AX7203 (Closes #11074) #11082): M1b top 573 of 573 receipt lines PASS, 1196 LUT, 1096 FF, 0 DSP, 0 BRAM, 116.06 MHz post-route (earlier 572 of 572, 1197 LUT, 106.30 MHz; the Verilog changed since fpga(ax7203): ternary mat-vec over real tern-tc-9m weights in DDR3, UART receipt (Closes #10461) #10475), each negative control FAIL on its own circuit only. SAR top 549 of 549 PASS bit-exact (550 lines start with T27, as before), 908 LUT, 538 FF, 0 DSP, 0 BRAM, 108.27 MHz, control FAIL on exactly the 16 predicted conversions. Both tops are behavioural: they do not read the PSP103 tables, so the FPGA size of a PSP103 solver is not measured.
   • Seal drift (PR tri seals drift: name the claims that moved, scan the ledger suite_jobs() specs at a time #11075): the lane measured the scan of 2298 specs at 69.80, 56.44, 107.79 and 92.37 s before and 14.40, 25.05 and 24.51 s after, under load averages of 9 to 44; a forced serial control took 88.69 s. The orchestrator's run on a master tree with both CLI PRs merged: 25.69 s, 0 specs moved; a planted verilog drift printed moved: gen_hash_verilog, a 1-byte edit printed moved: spec_hash, --fix re-saved exactly that one spec.
   • Builds: a t27c release rebuild after merging master with dependencies cached took 47.27 s on an unloaded machine and 176.10 s at load 40; tri took 485.47 s at load 40 and about 19 minutes from a cold target (lane report).
   • Offload: the full aperture scan of the golden comparator (20 transients) ran on the remote Docker host as a container job in about one minute of wall time (container started 18:13, log last written 18:14 on the host clock, load average 0.06 afterwards); on the laptop under load the same deck had finished 4 of 20 in about 45 minutes before it was stopped.
   • Machine: free disk fell from 9.4 GB to 3.6 GB in about 50 minutes; swap used reached 6752 of 7168 MB (3.0 GB an hour earlier). After the heavy runs moved off the laptop: load average 6.5, swap used 5.0 GB, 6.0 GB free.
   • Coverage (PR tri coverage: five coverage facts per spec, the rules in specs/tri/coverage/coverage.t27 #11096): the lane's harness over 29 specs of the five groups gave 140 of 145 facts before and 145 of 145 after its edits and re-saved seals; @@pending@@ the live tri coverage number.

   4. Weak points and competitors

   • Analog (PR docs(analog): the ten weaknesses revisited against M1-M5 measurements, W11-W19 found while building, a competitor delta (Closes #11058) #11071, section 11 of docs/analog/competitors.md): W1 reduced (exact zero-order hold for the linear cell, amplitude ratio 1.0003 against 1.78e-7 for backward Euler; the nonlinear path is still backward Euler); W2 open (no stiffness report, steps chosen by hand); W3 reduced (charge tables for tt and ss only); W4 reduced (five corners within 0.56 % where Id is at least 100 uA, weak inversion not met); W5 reduced; W6 reduced; W7 open (no speed claim is allowed); W8 reduced; W9 reduced (RAW writer byte-equal to ngspice's); W10 reduced. Nine more weak points found while building are listed as W11 onward.
   • Lab (PR docs(lab): competitors, prior art and weaknesses of the lab as a network job kind, three options #11072, docs/lab/competitors.md): volunteer computing, compute markets, remote build execution and mobile limits; the design's weak points are the cost of k-of-n verification, Sybil and collusion, floating-point determinism, untrusted code, node supply at cold start, mobile limits and the legal line.

   5. Decomposition

   Epic #11063 with children #11057 to #11062, #11068, #11081, #11083; the lab epic #11040 with #11041 to #11045; the analog epic #10486 with #10942, #10881, #10882, #10911. The ledger comment of the epic holds one row per tick and a status table; the tick is now decided by tri loop plan.

   6. Not done, and why

   • sar_kick_net.t27 (analog M4e-kick: derive the comparator's clock-edge kick on the two top plates from the plates' drive -- predict the six first-decision rows of #10888 (Part of #10888) #10942): the decks and the aperture table exist, the prototype hits scenarios S1 to S4 and misses S5 (late offsets) by 0.2 to 0.6 mV; the spec is the next lane's work. @@pending@@ result of that lane.
   • analog M3c-2: a weak-inversion-accurate PSP103 current read (log-domain table, integer exp) against ngspice #10881 (weak inversion read) and analog M3c-3: PSP103 charge tables for ff, fs, sf and the gate charge in a circuit #10882 (charge tables for ff, fs, sf): queued; their ngspice sweeps go to the remote host, not the laptop.
   • analog: VACASK as a second reference simulator for the SAR blocks (partner suggestion) #10911 (VACASK as second reference): the first attempt on the remote host found VACASK 0.3.2-103 (debug build) in the container image there, which rejects the main-branch demo (syntax error at line 42, column 62); a newer VACASK is the next step.
   • M5 second half (the M4e trace as a RAW file) and M6 (release v0.1, due 2026-11-16) are untouched tonight.
   • No GPU workload exists in the analog work; the GPU machine of the external team was asked about and not answered yet.

   7. Self-critique

   • The landing of every lane went through a throwaway script outside the repository because gh fails TLS inside the sandbox and the zig cache is not writable there; the script is now the spec of tri land (tri land: land a lane branch with the gates in a fixed order, the rules in specs/tri/land/land.t27 #11080).
   • A seal landed recording the sandbox's blocked zig cache with a home path (PR tri land: land a lane branch with the gates in a fixed order, the rules in specs/tri/land/land.t27 #11080); found in review, re-saved, and the script now re-saves any seal that records blocked.
   • PR docs(analog): board regression of the M1b and SAR tops on master -- both receipts reproduced on the AX7203 (Closes #11074) #11082 merged with its scan check red because the document named a developer home path; the check is not required, so auto-merge did not wait for it and I had not read non-required reds. Fixed in docs(analog): the board regression document names no developer home path #11084; the landing script now refuses such a path.
   • I first blamed the gen-verilog merge (gen-verilog: an entry parameter named clk/rst_n/en is declared once; a void body's last call stays a statement (Closes #5649, Closes #5650, Closes #5651) #8844) for two red fpga checks; the merge-rules session showed they come from specs/fpga/systolic.t27 ([bug] gen-verilog lowers an assignment to a signed array element as $signed(c[..]) = .., an illegal l-value (master FPGA E2E red since #7719) #10780).
   • I told the external contact that nothing more would run on his host, against the owner's earlier instruction to use it; corrected within the hour.
   • Heavy ngspice runs on the laptop filled swap and disk; the idle remote host should have taken them from the start.
   • Two scripts were killed by closing their terminal tab before they printed their final marker.
   • tri worktrees printed mine for another session's lanes (tri worktrees: no 'mine' label without evidence, and a tree with a recent commit is never planned for removal #11081).
   • The first cron carried the owner's raw directive as its prompt and would have restarted the whole night every 15 minutes; replaced by a tick that reads the ledger first.

   8. What changed in skills, tri and memory

   Skills and rules: CLAUDE.md section "Proactivity: request, do not ask", the user skill request-dont-ask, step 6 "Requests out" of the analog loop tick. tri: remote, seals drift, loop plan, land, learn, coverage (and worktrees in #11081). Memory: the night-loop file with the traps found tonight (gh TLS in the sandbox, blocked seals, the issue-shape gate, non-required reds, swap on the shared APFS container).

   9. Three collaboration options for the next loop

   Option A: finish the v0.1 deliverable with the external EDA team. Needs from them: access to their self-checking evaluator and its ADC limits, the ngspice version behind their published RAW, a VACASK that reads current syntax. Cost: small to medium; risk: their answers and their trust (a call was mandatory for them earlier). First measurable step: VACASK, ngspice-47 and our model on the golden comparator's flip point and the DAC top-plate voltages, differences tabulated (#10911). Default taken if nobody answers: this option.

   Option B: the lab as a network job kind. Needs: operators of machines (our own laptop, a CI runner, the remote host if its owner agrees). Cost: medium; risk: determinism across hardware, untrusted code, the legal line (test-network XP only, no value claim). First measurable step: run the closed gate list on one pinned commit on each host class and count the verdict hashes that differ.

   Option C: open-source upstreams. Needs: the maintainers of VACASK, OpenVAF-Reloaded, IHP-Open-PDK and ngspice. Cost: low; risk: the method is public before v0.1. First measurable step: one public, reproducible ngspice-versus-VACASK comparison of the comparator flip point, and a count of maintainer responses within 14 days.
2. gHashTag

   closed this as completedin 04a798aon Oct 19, 2026
3. gHashTag

   added 2 commits that reference this issue on Oct 19, 2026

   docs(analog): night report addendum (VACASK second reference, the def…

   ...

   98b57db

   docs(analog): night report addendum 2 (weak inversion, PR 11146) and …

   ...

   a1780a1