The Whole Line: How the Fleet's Work Reaches treadcore, and What Stops It

Abstract

The grove fleet runs about 114 machine-hours of work a day across nine training methods, and the person who owns it could not say how they connect to the game. This report draws the connection, stage by stage, with what each stage takes, gives, simulates and adds to treadcore. Only the dial search and A/B reach the game today. The transform stops at a file the game cannot read. Crucible's GPU islands stop earlier: their nets have different senses and outputs, and no road leads from them to treadcore. Imitation nets play only in the census, because the client never loads a net.

The history answers a question nobody had asked. Half of all promotions ever requested (7 of 14) failed, every one at the brain's unit tests and none at a family. The last two were passes under the new ratchet, stopped by tests that read defaults they were meant to be independent of. A patch to six tests fixed them, on main as treadcore 1b14192, and the first candidate it freed then passed its tests. No family job has ever failed. The red segment on the Queue page that reads like failure was 96% retired jobs, nearly all of which never ran.

We moved the difficulty into thirteen values a person can change, simulated what they buy, and applied two to the live run: ease 2.6 (held z 0.40) and a ceiling of 2.75. Every screen now explains itself in a panel collapsed by default, the queue says why each job is there, and a new Flow, Timeline, Queued and Values sit in the Queue chapter's column. Every generation now writes its own record, and the raw record packs for disc or release. A trial that passed on the defaults a generation replaces now gets another read rather than being thrown away.

Index Terms: data flow, sequential testing, generational search, heuristic evaluation, training records, residual learning.

I. What was asked

Paul, 1 November, quoted because each clause became a section:

"we want to write reports about each of our training methods. in specific I want to stream line, to unify the approach of all the features in this fleet ... I want to store all this training information. Github or maybe burned compressed media. I want snapshots and generations. I specifically want to conceptualize the idea of data flow from each of the modules. I don't understand how they all connect back to treadcore ... We should see diagrams and gantt style scheduling, connecting back to a timeline for the past and a gantt chart of scheduled tasks. also on the queue write a blerb about why the item is in the queue ... We have way too many screens, Or at least, I need an explanation on each screen (collapsed by default) ... We want full hurestic analysis."

And while this was being written:

"I see the Crucible, GPU islands is its own thing. How can we imrpove the gerations, perhaps the Crucible is like the Families? Is there a difficulty adjustor that we can use to make it so families can be at optimal or more frequency ... I'm seeing lots of families fail in the history ... Is its Netbrain used as an optional imported net weights into treadcore ... We can also run multiple and mix or alternate between using source of instructions by computing both."

The answers, in order: Section II, the connection. Section III, a report on each method. Section IV, what the history says. Section V, the values. Section VI, families and crucible. Section VII, one approach, check-ins and netlets, and mixing brains. Section VIII, storage. Section IX, the screens and the heuristic analysis.

II. How the pieces connect

Treadcore's bots play by rules: commander.rs, tactics.rs, crew.rs, fleet.rs. Those rules read about 120 numbers and 18 switches from weights.rs, and that file is the fleet's only way into the game today. Everything else is either a way of choosing those numbers or a thing that has not got a road yet.

Read it as a loop. The fit proposes a set of dials. The census plays it against today's defaults. If it survives the reads, its numbers become a commit, every machine builds it, and that commit becomes the baseline the next trial is read against. Nothing here trains a net: the search is a Monte Carlo experiment on a hand-written program.

The census is the instrument. One seed is one starting world, played three times so that each team takes the trial seat once. Every bot is a full brain and no person plays: 15 minutes of game time, three respawns, on eleven maps. The reading is the trial team's score a seat less the baseline's, and both meet the same start, so most of a map's luck cancels. A seed still swings about 9.5 a seat, which is why every later stage exists.

A. One trial, moving

The Flow page on the Queue chapter shows these numbers live. A trial called w4602 went like this at 09:32 to 10:31 on 1 November: a draw from the fit that moved many dials; first read on macmini, +2.58 +/- 1.51, z +1.71; three reads on, on saskatchewan, calber and boburner, the pool holding near z 2; then 400 unseen seeds on jaguaribe, -0.22 +/- 0.49. It looked strong for 187 seeds and was nothing. That is the reason for a held-out read.

B. The past and the schedule

The fine stripes on the left are the 90-second first reads. Long orange bars are held-out reads, at about 13 minutes each. The pink ones are the line's proof. The gap near 07:15 is the keeper's restart for the ratchet.

III. A report on each training method

Each is a proposer of something the game might use. The verdict at the end of each is what it adds to treadcore today.

A. The dial search

Method. suggest.py fits the ledger of every census and draws from what the fit does not yet know, keeping each number inside RANGES. The sequential census reads it. Conditions. Whole matches of bots, on all eleven maps. Spend. 105 of the day's 114 hours, with A/B. Result. First reads since the current defaults average z +0.18, with 59% at or above zero. Report: difficulty. Adds: the numbers in weights.rs and tactics.rs, and so how every bot plays. Reaches the game.

B. A/B versions

Method. Every switch flipped and every choice's other versions, each generation: 37 today. Conditions. The same census. Result. A/B reads worse at first, z -0.80 with 30% above zero, because most switches are on for a reason. But it has passed every real held-out read it reached since 31 October, three of three. The single-change trials also make the cleanest commits: w4399 and w2087, the last two generations, each changed one thing. Report: brain versions. Adds: which functions run. Reaches the game.

C. Families

Method. Set pieces, not matches: a duel in the open, a raid on a guarded base, a rolling tank under two shells. Nine families, 6,600 runs, 33 seconds in all, each with a floor. Result. See Section VI. Adds: a veto, so that the average cannot rise by trading away a skill. Reaches the game by what it refuses.

D. Anchors and the proof

Method. After a generation, the defaults from three generations back are read as a trial. Each hour, main's values are read against the line. Result. The clearest good news in this report. The defaults before r0060 read -5.87 +/- 0.66 and -5.40 +/- 0.67 against today's; those before r0062 read -4.08 +/- 0.66 and -3.42 +/- 0.64. Those gaps are three promotions' work, measured on seeds no search saw, and they are a far larger signal than any single generation's read. The proof, by contrast, reads the one carried generation at -0.14 +/- 0.16 against main: no difference yet. Adds: the correction and the gate. Decides what reaches the game.

E. The transform

Method. A residual net over the rules' Move: it reads the senses, the biorhythm and the rules' own answer, and may move the point if it is confident. All zeros is the brain as it was. SPSA moves all 918 numbers by one noisy census a step. Conditions. The census with the patch switched on. Spend. About half an hour a day. Result. Six steps today read between -3.72 and +4.97 with errors near 2: noise, as expected this early. Report: Algorhythm. Adds: nothing yet. The census reads the slot; the game has no file for it.

F. Crucible's GPU islands

Method. 2,048 small nets per island duel on the card, 128 episodes and 100 generations a segment, with champions passed between islands. Conditions. Its own arena: 64 x 64, six pillars, twelve rays, one shell each. A net of 32 inputs and 3 outputs. Spend. 7.5 hours a day today, on cards that do 400 to 580 million tank-steps a second. Report: one source, two CUDAs. Adds: nothing. No road. See Section VI.

G. Imitation: record, train, school, gauntlet

Method. Recorded rule-brain matches teach a net of about 16,000 numbers, then it is examined. State. Dormant since 30 October: the rotation now holds only the transform. Result. The learned tank is about 20 a seat behind the rules. Report: a CUDA node, LEARNING-DESIGN. Adds: nets/bc2.bin, seated by the census. The client never loads a net.

H. Whole-brain sweeps

Switches, parts, a census of the best, round robin: each part alone against the defaults. State. Dormant since 29 October, with the A/B doing the same job continuously. Adds: readings the fit can use through the ledger.

I. Older searches

SPSA and tune over a chosen few dials. State. Dormant since 29 October. They are duplicates of the fit in purpose, and the fit has the better record. Adds: nothing the fit does not. They should stay asleep, and their runs stay available by hand.

J. Duplication, plainly

There is one real duplication left: three dormant ways of searching dials. The live methods do not overlap. But the CPU lanes and the GPU lanes are not one brain. The CPU search is a hand-written program being tuned, and the GPU search is a different creature in a different world.

IV. What the history says

A. Half of all promotions stopped at a test

These tests are not catching bad brains. They call Settings::default() and then assert a number, so when the fleet moves that default the test fails. brains.md already gives the rule: "If an existing test pins a number the choice moves, give that test the number it was written for." The last failures break that rule. Any trial that moves alpha cannot promote, and 65 of the last 177 fit draws (37%) move it. The seven failures hit six different tests, so fixing them one failure at a time is slow: each costs a passed candidate.

The patch is small edits to tests alone, in whole-line/treadcore-test-pins.patch. We applied it to a scratch copy of 9d438c8 with both failed candidates' values and ran the brain's tests: 134 passed, 0 failed.

Because the failures keep finding new tests, we then looked for all of them at once: every number that the search may move, 107 of them, moved 7% together, inside RANGES. That found three more pinned tests before they could stop a candidate: the_posture_leans_the_ladder, a_fleet_crosses_a_bridge_in_file and a_switch_comes_back_and_a_slider_keeps_the_target. The patch fixes those too, six tests in three files, and with it all 134 pass at +7% and with both candidates' values. Moving everything together by -40% or +50% still fails 5 and 3 tests, among them every_scenario_holds. We left those alone: a brain with every dial that far out may really be a different brain, and that is what those tests are for.

Applied, 1 November, at Paul's word ("commit. push. impliment."), as treadcore 1b14192. Main moving did the rest with no hand on it: the planner replayed the line on top of it as 62f5a00 and retried both stopped candidates, and w4540 went through the tests that had stopped it. A failure like these can still come from a test nobody has found, so the Board now marks any stopped promotion as blocked and next essential.

B. Why the red bar looked like failure

The Queue page drew "2,429 failed or stopped" in one red bar. Here is what it held:

The bar now draws retired apart from failed, in grey.

C. What a generation throws away

Each generation discards every read on the old defaults. When r0069 landed at 08:18, two held-out reads in progress would have passed: w4398 at z 0.83 and w4424 at z 1.07. The 2,124 retired jobs above are the same effect, cheaply. It matters more the faster generations come: at one every 30 minutes, a trial that needs 25 minutes of reads rarely finishes. This limits the pace, and the simulation in Section V does not model it, so its pace figures are optimistic.

So a generation now carries what it would have lost. carry_passes takes up to three trials that passed on the defaults the last generation replaced, strongest first, and gives each one read on the new ones, ahead of the search. Only one generation back, because a read says less as the defaults drift. Run against the live queue, it chose exactly w4424 and w4398.

D. The funnel, at the current defaults

In the 2.5 hours from 08:20: 212 first reads, 139 reads on, 10 held-out reads. Of 121 trials decided, 113 were dropped and 8 passed the read-on bars. Of the real held-out reads, 2 passed, and tests stopped both. No generation was carried in that time. Without the test failures, the same passes would have come 98 and 37 minutes apart.

V. The values

Every number that sets the difficulty, and how much is carried, now lives in one table, tools/plans/line.json, and is changed by the Values page, by tools/fleet-tune.py QUEUE KEY VALUE, or by hand in QUEUE/tuning.json. A change is live at the next plan, about a minute, and restarts nothing. Each is held to its range at the button, in the tool and in the planner. A broken file means the defaults, never a stopped plan.

A. What they buy

tools/plans/difficulty-sim.py now takes this game's noise and the read-on limit. The prior had to be calibrated, because fitting it to first reads (N(-0.14, 1.03)) gave one generation every 11 minutes, which the run plainly does not do: outliers such as the anchors inflate the spread. N(-0.65, 0.45) at 9,000 seeds an hour reproduces 86 trials an hour against the 85 measured, and a pass rate of 7% against 6%; it gives 20% at the held-out read against two of five measured, a handful.

The Monte Carlo error on these is large enough that 2.5 and 2.6 are not really different. The shape is what to trust: the gain holds to about 2.6, then falls.

Raising the read-on limit from 0.5 to 0.75 or 1.0 bought no gain at any ease, so it stays at 0.5.

B. What was applied to the run

Applied at about 10:40 and read back from the planner's own state:

• Ease now 2.6, held z 0.40. This is the fastest setting that gives up no gain. From here each generation takes 0.05 off.
• Ease ceiling 2.75, down from 3.0. At 3.0, a generation can pass on no evidence at all, and half are truly worse.

Everything else stays where you set it: 0.05 harder each generation, and four hours dry before easing.

About half an hour is not reachable through the bars alone. The simulation reaches 30 minutes only at an ease near 2.75 with the read-on limit raised, where gain falls to +1.2 a day. The honest expectation is 35 to 50 minutes, and whether the carry improves on it is not yet measured.

VI. Families and crucible

A. Are families failing?

No. Across 130 family verdicts on record, 2 were UNDER, both on the defaults and both one family, a raid on a guarded base at 70.2% against a floor of 75%. You lowered that floor to 65% on 30 October. All 62 of the fleet's candidates hold every floor. What looked like family failure was the tests and the red bar of Section IV.

B. Should families have a difficulty adjustor?

No, and none was built. An adjustor would make floors easier to pass, and the floors are not what stops anything. It would save nothing: all nine families together cost 33 seconds, about 0.1 hours a day (Fig. 5). The floors are also your decision, in the scenario files. Two sit close, a ready gun and two shells, and those are the first to watch if a family ever does begin to fail.

C. Can the beginning success rate improve?

Yes, and the bars are the wrong lever, because Section V shows the read-on limit buys nothing. In order of what each would buy:

1. The test patch, done. More than a third of fit draws became promotable at once.
2. Carry a pass across a generation, done. A trial that passed on the replaced defaults gets a held-out read on the new baseline. It covers held-out passes only: a trial still in its reads on when a generation lands is still dropped. Its gain is not yet measured.
3. Test before the long read. Running the tests before a 13-minute held-out read would save a read each time they fail. After the patch they should rarely fail, so this comes after it.
4. Keep the proposers as they are. Fit draws lead at first read and A/B leads at held-out read, so neither should crowd out the other. Moving families earlier would save nothing, since they almost never fail.

D. Is crucible like the families?

They are opposite in role, and the difference is useful. A family is a veto: a cheap fixed test on the game's own simulation that can only say no. Crucible is a proposer: an enormously fast search in another world that cannot yet say anything about this one.

The likeness points to the way to join them. A family is a short scene on the real simulation with a pass rate, so the families are the natural exam for a bridged net, and later the natural curriculum. Duelling skill from crucible, cloned into a net with treadcore's senses, could be read first against a-duel-in-the-open and the shell families, in 33 seconds, long before it faces a census. That is a far cheaper first read than any it has now.

E. Is crucible's net an optional import for treadcore?

No, and it cannot be yet. The two have nothing in common but the word:

Treadcore's own NetBrain is not an option for players either: the client has no way to load a net. So the sliders move dials today, not brains. The model and patch rows do appear in the U page, as every weight does, and that looks like a bug to fix: from reading patch.rs, moving patch above 0 asks for CENSUS_PATCHES, which a player's machine does not have, and it panics. We read this and did not run it. The design in brains.md is a bridge: play a champion as a teacher inside an arena with treadcore's senses, and clone it into a treadcore net (behaviour cloning, as bc1 was made). The result goes in slot 7 and is read like anything else.

Expect little at first. A clone learns to duel, and the census scores a whole game. Its likeliest use is a patch or a role rather than a whole brain.

VII. One approach

A. Every method is a proposer

The unification is one sentence: a method's only job is to produce something the census can seat, and the same road judges all of them. The dials already work this way. The transform and the nets become the same once the game can load a slot. The rule costs no new process, and it makes the GPU question precise: not "how do we use the GPUs" but "what can crucible hand to the check-in?"

B. Check-ins and netlets

A check-in is one proposal entering the road, one record: what it is (a weights file or a slot), its hash, who proposed it (fit, A/B, SPSA, crucible segment), its parent generation, each read, and how it ended. Today those facts are spread across the job, the results/ directory and the run's state. plan/pipeline.json's traces assembles them for the last eight, and the generation record keeps them for what is promoted. The step left is to write one for every trial at its end.

A netlet is a small learned part with one job: a slot file and a manifest naming its inputs, outputs, parent and evidence. The transform is the first: 918 numbers, all zeros meaning the brain as it was. The rule that makes netlets safe to keep adding is the transform's own: a netlet that does nothing costs nothing and changes nothing, so the census can read it as it reads any dial. One could steer retreats, another aim mortars. Each is promoted or dropped alone.

C. Mixing brains: computing both

Your idea fits the house rule, "either/or is a scalar". There are four ways to do it, cheapest and most honest first:

• C, override when sure, is built: the patch. The rules decide and a net may move their point past a confidence bar. Since the net can only be worse by moving something, the census reads the risk directly.
• A, by seat, is built for the census: roles and model choose per team. It wants one more step to choose per tank, and no game path yet.
• B, alternate, a tank following one source and then the other on a slow clock, is new. The biorhythm's clock would serve. It is cheap to compute and gives the game some variety, which a person may enjoy even where the census shows no gain.
• D, blend, is new and has a limit. The rules produce a decision and the net produces 42 scores, so they cannot be averaged before the choice. They can be averaged afterward, at the Move's point, which is where the patch already acts. It does not apply to firing or building, which are yes or no.

A net decision takes about 10 microseconds, so computing both costs little. Each needs a game-side slot first, which is the same gap again: the game reading a slot is the one piece of work that unblocks netlets, mixing, the bridge, and player-selectable brains together.

D. The order

1. The test patch. Done: treadcore 1b14192.
2. Carry a pass across a generation. Done.
3. A game-side slot: the client loads a patch or net from a file, off by default. This is what makes the transform shippable.
4. The bridge from crucible to slot 7, with the families as its exam.
5. Mixing modes B and D as weights, then player sliders for them.

VIII. Storing the record

Generations go in git. tools/fleet-archive.py QUEUE snapshot writes one directory per generation with its values, evidence, families and commit, plus a list of every promotion including the failures. The planner now runs it whenever a promotion finishes. The seven generations so far, with the settings and ledgers, come to 1.2 MB, so years of this fit comfortably in a git repository.

Raw packs go on an SSD, then on release assets or disc. fleet-archive.py QUEUE pack bd25 writes what no earlier pack holds to PACK-NNN.tar.xz, closing each below a disc's size, with MANIFEST.tsv holding every file's sha256. verify reads every file back out and checks it.

The whole record so far was packed and checked: 3.82 GB and 173,018 files went into one 393 MB pack, 9.7 to 1, in 25 minutes at low priority. Reading every file back found 0 differences. It sits at QUEUE/archive/packs/, on this Mac only. The record grows by about 0.9 GB a day raw, or about 95 MB a day packed.

GitHub or disc: both, for different things. My recommendation:

Those GitHub limits are as I know them; check them before relying on them. Both orrery and treadcore are public repositories, and the raw logs hold machine names, user names, addresses and paths. So the record belongs in a separate private repository, and packs should not be published as they stand. Nothing has been sent anywhere.

The manifest is what makes a disc trustworthy decades on: it needs only tar and sha256sum. At this rate a 25 GB BD-R holds about eight months, and today's pack would fit on a CD.

IX. The screens

A. What changed

Explanations. Every screen has an "About this screen" bar. Shut, it is one line saying what the screen does. Open, it answers six questions: what it does, where it ties in, why it is drawn this way, what to look for, what is being simulated, and why it helps the brain. It closes when you leave. A test fails if any screen is missing an answer, or if a word has a character the fonts cannot draw. All the words live in line.json, the same table the planner and this report read.

The queue says why. Queue, Queued lists what is running and waiting, in order, each with a line on why it is there. A held-out read adds the bar it must pass. The upcoming list had been collected and never drawn.

The data has a picture. Queue, Flow draws the road with one real trial traced beneath; Queue, Timeline draws past and forecast by machine; Queue, Values holds the buttons.

A rearrangement. The Pipeline had no key in the Queue's column, and the only way there was a menu. One chapter now has one column: Fleet, Runs, Board, Flow, Timeline, Queued, Values.

B. Too many screens?

There are 27 faces, counting pages and tabs. By what they answer, they fall into five groups, and only one is about the line:

We removed none. Each still has a job, and removing one is your call. These are the merges I would make next, in order: Algorhythm's Fleet tab goes (Flow replaces it, with more). Net folds into Algorhythm's Matrices, since it shows the same net. Flags and Keeper's alerts become one list. That takes 27 to 24, and the next bigger saving would be moving Media and Chat behind menus, since neither touches the line.

C. Heuristic analysis

Nielsen's ten usability heuristics, applied to the console as it stood this morning:

The search has heuristics of its own, and this is how they held up:

D. Every screen

X. Threats to validity

The simulation is a rough model. Its prior was fitted to the funnel rather than measured, it matches first reads and passes better than it matches held-out reads, and it ignores what a new generation discards. The funnel since the ratchet covers only 2.5 hours and two passes: too small to say whether 2.6 will give the cadence shown. The anchor gains include several promotions, and we did not separate how much each gave. The run moves while this is written, so its numbers are as of about 10:55 on 1 November, and the figures can be redrawn by tools/report/whole-line.py.

XI. Still open

• A trial still in its reads on when a generation lands is still dropped, and the carry's gain is unmeasured.
• The game has no slot, so neither the transform, the nets nor any mix can ship.
• No check-in record is written per trial yet, only per generation.
• The Timeline and Flow pages need the new build: the running Orrery is the old one until it is reopened.
• The raw record is on this Mac alone. No disc is burned and nothing is uploaded.
• The patch row on treadcore's U page looks able to crash a player's game (Section VI-E): read, not run.
• This report's figures are one moment of the run. The screens are live.

XII. Reproducing this

Q=~/code/treadcore/build/queue/grove
python3 tools/fleet-tune.py $Q                 # the values as they are
python3 tools/fleet-tune.py $Q ease_now 2.6    # move one
python3 tools/plans/difficulty-sim.py --mu -0.65 --tau 0.45 --rate 9000 --keep 0.5
python3 tools/fleet-archive.py $Q snapshot     # generations
python3 tools/fleet-archive.py $Q pack bd25    # raw, for a disc
python3 tools/fleet-archive.py $Q verify
python3 tools/report/whole-line.py $Q && sh tools/report/all.sh