Census to Commit

IIntroduction

Create a learning loop and describe how the coefficients get commited back to treadcore, so the it picks it up in the AI. For example, show me commit versions, and timestamps. Actaully, show me hos the brainz and the net data are stored, show me the matrix for the AI. show me its composition, draw diagrams. we want to see the process of discovering a change, and promoting it, so push it on to the next download of treadcore. essentially creating micro versions of this git submission. its part of the program. it has a do it now button to sync and push the values.

also describe the results tab. why do we have failed tasks in family. why are some older, when it says promoted. describe the layout. of what happends. perhaps add a scrollbar and a filter by type.

Paul Richeson, 28 October. Everything below was read from the repository, the queue's files and the fleet on that day. No number in it is illustrative.

IIWhere the brains are stored

A. The scripted brain: 111 named numbers in one table

The scripted brain decides with rules, and every number a rule weighs is one row of the weights! table in source/treadcore-brain/src/weights.rs. A row has three parts:

• a Rust name;
• the name a person types;
• the default, with a line saying what the number does.

The macro turns the table into a flat array, Weights, which is lent to whatever scores and never copied per think. The defaults are compiled into every binary built from the commit: the game, the census, the fleet's jobs.

A trial does not edit the table. It is a small text file that names only the numbers it moves. The census passes it with TRIAL= (the game takes it with --weights FILE), and every number it doesn't name keeps the compiled default:

B. The learned net: 65,752 bytes

The net that drives a learned tank is a file in nets/. Six are kept: phase1, map4, map5, map7, map7a and bc1/bc2. map7 is the official one, and the official net changes only when Paul says. bc2 is the one the Net page shows.

The format is TCNET1, written by Net::save in net.rs:

• an 8-byte mark;
• the number of layer sizes, then the sizes;
• each layer's weights row by row (one row per output), then its biases;
• everything little-endian, as 32-bit integers and floats.

C. The net's composition, and what each matrix does

IIIThe learning loop

The scripted brain's numbers improve in a loop that closes itself in auto mode. Each box below is a kind of job in grove's queue, and every arrow is a file the next step reads.

The rule a candidate must pass

IVOne promotion, minute by minute

grove/w0034 is the only candidate promoted so far. These are its times from the queue's run files and from git, Pacific time, 28 October 2026.

1. 13:38
   w0034 passes both held-out readsa candidate: +2.67 and +3.17. Its census and reads ran on commit 92972b7.
2. 14:14:02
   Auto mode asks for run r0019promote grove/w0034: the strongest current candidate whose families hold.
3. 14:20:54
   Commit dc3df01 on promote/grove-r0019weights.rs edited (7 lines); 109 tests and every family pass on the branch. The commit message carries the evidence.
4. 14:28:08
   Approved: pushed to mainauto mode approves its own prepared promotion; main is dc3df01. From here a clone gets the new brain.
5. 14:30
   LR 119 recorded on top: c1e00c0the records commit; it moves no brain source, so the defaults stay dc3df01's.
6. 16:05:21
   Confirmed on 7 of 7 machineseach machine built a commit since the approval and its defaults.toml has the promoted values (whitewolf by c1e00c0, the rest by dc3df01).
7. 19:09:33
   Tag brain-r0019 pushedthe micro version, made for r0019 after the fact (tag object c066c65, signed); every promotion from now on is tagged as it is approved.

VThe Results tab

1

2

3

4

5

6

7

8

1. 1
   The chapters. Results is chapter 4 of five. Machines adds a computer, Projects gives it work, Queue runs it, Results reads what it found, and Watch looks in.
2. 2
   The pages. Candidates (with its count), Found (improvements by variable and machine), Net (Figure 4) and Flags (what waits for a person). A green lamp means there is something new; a red lamp on Flags means something needs answering.
3. 3
   The headline, and Do it now (new). Press it twice. It syncs every machine's ledger, then promotes the strongest current candidate; auto mode pushes the promotion to main and tags it once its tests and families pass. With no current candidate, it re-reads the strongest older ones instead. The menu bar has it too: Fleet → Do It Now, which arrives armed.
4. 4
   The newest change any census reported: the coefficient, its value, its offset from the default, the census, and the machine that read it.
5. 5
   Auto mode and its switch. With auto on, nothing on this page waits for a person.
6. 6
   Promotions, newest first. Under the newest approved one is each machine's confirmation (ok: its build has the values). r0009 is grey, not red: it failed before r0019 landed, and it is marked superseded.
7. 7
   The filter (new): All, Promoted, Current and Older, each with its count.
8. 8
   The list and its scroll bar (new): the wheel, or drag the bar. Each line shows when, the trial, its two held-out reads, its standing, and what it moves.

Why 36 candidates say "older" while one says "promoted"

A census measures a trial against the defaults of its day: +2.67 means w0034 beat the brain of commit 92972b7. When w0034 was promoted, the defaults became w0034's own numbers. Every other candidate's evidence then describes a race against a brain the game no longer has.

• Most of them share most of w0034's values: every one of the 37 sets castle_inside 8 and kite_in 4.5.
• Their lead over the old defaults says nothing about a lead over the new ones.
• Promoting one on that evidence could undo a gain.

So the page labels each with the commit it was read at, older (89371b6), and auto mode promotes only a current candidate. Before this step, nothing turned an older candidate current. The loop stood still at "0 current", and Do it now would have had nothing to push.

Box 11 of Figure 5 is the fix. With none current, auto mode re-reads the three strongest older candidates on today's defaults. Each is a new trial with its own census and two held-out reads, and one that still wins becomes current and promotes itself. The first round started at 19:11: w0045 (held out +2.45 and +3.09) and w0003 (+2.50 and +2.80). It also re-read w0034 against its own values. That is wasted work, kept as a control: it should read near zero. Promoted candidates are now excluded.

Why "failed tasks in family"

No family failed: grove has read 85 candidate-and-family sets, and all 85 hold. The one failed job in the queue is r0009, a promotion, at 10:12 that morning. Its message said "the tests or a family failed", which pointed at the families. Its log says otherwise:

r0009 changed the table's defaults but not Tactics::default(), which a test holds equal to them. LR 119 fixed the promotion step to change both (changed_vars), and r0019 passed with 109 tests four hours later. A failed promotion now names the step that failed, read from its log:

VIThe second job: crucible on the GPUs

Add the GPUs to crucible (Projects, two presses) gives every machine with a card and CUDA a second lane, NAME-cuda. It is the same machine again, taking only crucible's work, so its CPUs go on with treadcore's censuses.

Each lane is an island:

• its own population, trained 500 generations a segment, its state kept on the card's machine;
• six segments queued ahead, each after the last, so the card is not left idle between plans;
• each segment sends back its champion (8.6 KB) and takes the other islands' champions into its hall of fame.

calber-cuda · 1080 Tiwhitewolf-cuda · 1080boburner-cuda · 1070 Ti

crucible trains its own arena, not treadcore's brain. It reaches treadcore's candidates only through a GPU arena built from treadcore's combat rules (LR 121 D). That is the next step if Paul wants it.

VIITutorial: following a change by hand

1. See what the fleet is doing.

   python3 ~/code/orrery/tools/fleet-queue.py build/queue/grove status
2. Read every candidate's evidence, strongest first, with its standing:

   python3 -I ~/code/orrery/tools/plans/treadcore.py build/queue/grove evidence grove/w0045 held +2.45 +3.09 families holds older defaults grove/w0003 held +2.50 +2.80 families holds older defaults
3. Try a candidate's numbers yourself. Write them as a trial file and play a census, or the game, with them:

   printf 'w_near = 2\nkite_in = 4.5\n' > my.weights TRIAL=my.weights SEEDS=55 scripts/census.sh
4. Do it now, from orrery (Results, two presses) or the command line:

   python3 ~/code/orrery/tools/fleet-queue.py build/queue/grove now no candidate on today's defaults (19638a5) to promote; read again: ...
5. Find the micro versions and the brain your checkout plays:

   git tag -l 'brain-*' git show brain-r0019 git describe --tags --match 'brain-*'
6. Read a net's matrices in a few lines of Python:

   import struct b = open("nets/bc2.bin", "rb").read() assert b[:8] == b"TCNET1\0\0" k = struct.unpack_from("<I", b, 8)[0] sizes = struct.unpack_from(f"<{k}I", b, 12) # (143, 64, 64, 47) off = 12 + 4 * k for a, o in zip(sizes, sizes[1:]): W = struct.unpack_from(f"<{a*o}f", b, off); off += 4 * a * o # o rows of a bias = struct.unpack_from(f"<{o}f", b, off); off += 4 * o
7. Give the GPUs to crucible, or take them back:

   python3 ~/code/orrery/tools/machines.py build/queue/grove cuda-lanes add all python3 ~/code/orrery/tools/machines.py build/queue/grove cuda-lanes remove all

VIIIResults and discussion

The path from a number found by a search to the brain in a fresh clone has no manual step left. It is measured three times, held to nine families and the brain's tests, committed with its evidence, tagged, built and checked on every machine. It took 50 minutes from candidate to main, and 1 h 37 min more until every machine had built it. Most of that was waiting for busy lanes.

The limit is the evidence's shelf life. A promotion makes every other candidate's evidence stale. Until this step, the loop stopped there, waiting for new searches to find a new candidate. Re-reading the strongest older ones turns that stock back into candidates. Whether any of them still beats the new defaults is what the re-reads will tell.

Errors and limits

• brain-r0019 is retroactive. It was tagged five hours after its promotion; tags from now on are made as each promotion is approved.
• w0034 was re-read against itself in the first round of re-reads. Those three censuses are a control, and promoted candidates are now left out.
• Planning gaps. The keeper running since 12:28 plans only when its queue runs low or a run changes. It was not restarted, because restarting would have put whitewolf's three-hour school run back on the queue. crucible's islands keep six segments queued so they do not stall in between.
• Releases. A v* release (packages through GitHub Actions) and crates.io remain Paul's to ask for. A brain- tag starts no workflow.