Runtime Behaviour
Complete-configuration experiments
Qwen chooses which settings to adjust and when to run one 750-episode SnakeLab simulation. It can research measured data before selecting a setting, then receives a chart when it can inform the value choice.
Initial startup
With an empty experiment, Ax3l submits the schema defaults once. The baseline
builds on SnakeLab 3.0.4’s transformer-epsilon-nice.json sample and runs 750
episodes with required EpsilonNice:
| Setting | Initial value |
|---|---|
| Hidden size / encoder layers / attention heads | 64 / 1 / 4 |
| Dropout | 0.05 |
| Sequence length / batch size | 4 / 64 |
| Learning rate | 0.001 |
Gamma / target update rate (tau) |
0.96 / 0.05 |
| EpsilonNice probability / steps | 0.05 / 20 |
Tau remains fixed during optimization. On completion, the baseline becomes the initial golden configuration and has its measurements saved before the first choice. Interrupted initial seeding resumes its recorded submission.
Choosing a complete configuration
A new draft starts from the current golden configuration. It freezes the latest, golden, and all-time best reference runs for the entire selection pass. The draft is saved while Ax3l records any missing measurements. A restart uses the same frozen references and resumes the saved parameter or next-step decision. Qwen can choose any unused setting from the following list. Each setting can be chosen only once per experiment, including after a restart:
| Parameter | Purpose | Allowed values | Tuned with |
|---|---|---|---|
| Hidden Size | Number of neurons in each Transformer hidden layer. | Integers from 32 to 128, in multiples of 8. | Individually |
| Encoder Layers | Number of Transformer encoder layers. | 1, 2, or 3. | Individually |
| Dropout | Dropout probability in the Transformer. | Numbers from 0 to 0.1, in steps of 0.01. | Individually |
| Sequence Length | Number of consecutive frames in each training sequence. | Integers from 4 to 48, in multiples of 4. | Individually |
| Batch Size | Number of sequences sampled for each training update. | Integers from 8 to 64, in multiples of 2. | Individually |
| Learning Rate | Size of gradient updates during training. | Numbers from 0.0005 to 0.005. | Individually |
| Gamma | Discount factor controlling the value of future rewards. | Numbers from 0.9 to 0.99. | Individually |
| Random Seed | Change the simulation seed. | Ax3l increments the seed, skipping already stored complete configurations. Qwen supplies no number. | Immediately on selection |
| Epsilon: initial | Starting probability of choosing a random action. | Numbers from 0.85 to 0.999. | Epsilon: decay |
| Epsilon: decay | Multiplier applied to epsilon after each episode. | Numbers from 0.9 to 0.999. | Epsilon: initial |
| Food Reward: closer | Reward for moving closer to the food. | 0, 0.1, 0.25, 0.5, 1, 2, 3, 4, 5. | Food Reward: further |
| Food Reward: further | Penalty for moving farther from the food. | -5, -4, -3, -2, -1, -0.5, -0.25, -0.1, 0. | Food Reward: closer |
| EpsilonNice: probability | Chance per eligible move of starting a safe detour. | Numbers greater than 0 and at most 0.1 (10%). | EpsilonNice: steps |
| EpsilonNice: steps | Duration of each detour, reset each episode. | Integers from 1 to 40. | EpsilonNice: probability |
The epsilon settings, distance rewards, and EpsilonNice settings are chosen in
pairs. The sum of the two distance rewards must be non-positive. Selecting
seed immediately accepts the seed change; there is no separate seed
confirmation conversation or tool call. Ax3l increments the draft seed until
the complete configuration and its golden baseline are unique across every
stored run and status, preserving all other settings. Seed can be selected only
once per experiment. Numeric seed proposals are rejected; schema-fixed settings remain
controlled by Ax3l.
Ax3l loads the same short selection skill at the start of each choice. The prompt includes current values and same-setting values and scores from the latest and golden runs. Qwen receives at most one chart per choice.
| Choice | Chart | Measured facts alongside it |
|---|---|---|
| Learning rate | Training loss | Episodes with recorded loss |
| Epsilon | Score trend | Mean scores in the first and last 50 episodes and number of random actions |
| Distance rewards | Score distribution | Median score and counts of zero-score and high-score episodes |
Ax3l uses the latest run’s charts when it shares the draft seed. Otherwise it uses the current golden run. The learning rate choice uses data alone when the run has no recorded loss. Its loss chart overlays a trailing mean of up to 20 episodes on the faint raw trace, with the golden run’s loss and mean in orange when it is a different run with recorded loss. Means exclude missing losses and preserve their gaps. The source is selected again for each choice; after a seed change, runs on other seeds are labeled as historical context. Ax3l records the completed simulation’s measurements directly from episode data.
English directions in the prompt lead to relevant data_tool pages when
Qwen needs more detail. The tool menu at / links to draft, latest,
golden, and best. Navigation uses English and experiment values use
JSON. kb_tool serves optional background. Choice rules appear in the
submission tool description. Accepted choices are saved without starting a
simulation. A selected setting must change its current value, or at least one
value for a pair. Submitting all the same values is rejected before the uniqueness
check. After every submitted change, Ax3l
checks the complete draft for duplicates.
At startup and after a unique change, Qwen can use select_parameter to choose
a single or paired setting. data_tool and kb_tool are available while making that
decision. Ax3l opens the selected setting’s value conversation with the usual
evidence and tools, or immediately saves the next unused seed when seed is
selected; other values stay saved. A duplicate proposal is rejected
without saving it. Qwen stays on the selected setting and must submit different
values until the complete configuration is unique. After a unique change, Qwen can
use run_simulation or select an unused setting. Used settings cannot be selected
again and are omitted from the next selection prompt. Unselected settings keep their
baseline values. Qwen can submit the unique draft through an explicit run decision.
Each selection, value choice, and run decision resumes after a restart.
Parameters with finite choices are also omitted when all changed choices already exist for the draft’s current seed and other settings. For example, encoder layers cannot be selected if both alternatives to the current layer count have already been stored with those settings. Every run status counts, matching duplicate protection. The selection tool rejects exhausted settings before saving an assignment. Availability is recalculated against the current draft after each accepted change. A restart reopens selection if the pending assignment is exhausted.
Ax3l internally allows seven minutes to select a parameter and find a unique
setting, or decide to run the current unique draft. Prompts and tool feedback
do not mention the timer, remaining time, or expiry actions. Experiment Status
continues to display the countdown. Parameter selection and duplicate-value
retries share the countdown; every accepted unique change resets it to seven
minutes for the next parameter or run decision. Sending a prompt with a chart adds two minutes
to the remaining decision budget and the displayed countdown. That allowance
applies once for the prompt; tool reads and rejected values keep the extended
deadline. Prompts without an image receive no extra time.
Expiry starts a grace period configured by
decision_grace_seconds in /opt/prod/ax3l/conf/ax3l.json, defaulting to 420
seconds (seven minutes).
Before the first unique change, the end of grace discards the tentative parameter
selection and starts a fresh conversation from the initial prompt with a new
seven-minute budget. Earlier messages, research, reasoning, and rejected proposals
are not carried into the new context. With a saved unique draft, the end of grace
submits that configuration through the normal simulation, comparison, and
measurement cycle, then starts the next experiment.
Any pending edit is excluded unless its changed values were accepted. Restarting
begins a new budget.
Experiment Status shows the settings to be chosen and the saved selections. The experiment title and LLM countdown share a full-width section header above the experiment details and table. The countdown occupies the right side of that header, stacking below the title on small screens. The same header remains visible in active, inactive, and empty states. The timer ticks every second using the current decision deadline. At zero it turns amber. With a saved unique draft, it switches to a Simulation Start Countdown with Time remaining below the number. Before the first unique change, it instead shows a Conversation Reset Countdown. Both use the configured grace period, which defaults to seven minutes. It resets after an accepted unique change and becomes inactive when the configuration is submitted. The table scrolls horizontally on its own, keeping the experiment details and countdown visible; refreshes retain its scroll position. It lists chart rendering, each Qwen request, each data or guide lookup, and each submission under the relevant setting or pair. Each step has its own elapsed time. Any item with task rows has a +/− control before its label, including Experiment decision. Task rows start collapsed; opening or closing one group does not affect the others, and refreshes retain each group’s state. Saved prompt messages have their own rows with a Prompt link in the LLM details column. Completed Qwen requests link to their captured Reasoning when Qwen returned nonempty reasoning text. Hovering or focusing either link shows a scrollable preview using the same Markdown and JSON-table parser as Read results; prompts also include their captured chart images. Clicking these links opens the original content in a separate tab. Each returned Qwen request shows its own input and output token counts when reported by the server. The chart link opens the prompt with the exact image sent to Qwen. Read targets link to their recorded responses; hovering or focusing a link shows a scrollable preview, with JSON formatted as compact tables. The preview shows what Qwen received at that time, including rejected or failed reads. The page refreshes every five seconds.
Golden configurations and seed changes
A candidate becomes golden only if its high score strictly exceeds the current golden score. A tie retains the golden. Each completed configuration comparison counts as one experiment cycle.
Seed requests save the draft choice; the usual run decision or countdown starts the simulations. Ax3l first runs the current golden configuration with only the new seed and accepts its score as the local golden baseline, even if lower. This seed rotation appears as a purple point on Experiment High Scores. If seed is the only change, that single run completes the draft. If other settings changed, a second simulation runs the complete draft on the new seed and must beat the new baseline to become golden. A tie or loss retains the new baseline. Seed-only rotations do not count as configuration comparisons. There is no automatic rotation after nine unsuccessful proposals. Interrupted rotations resume their recorded baseline without resubmitting. Ax3l also finishes rotations recorded under the former automatic policy before resuming choices. The all-time best remains a separate historical reference and can come from an earlier seed.
Restart behavior
Drafts, choices, submission intent, comparisons, and measurements are stored in MariaDB’s shared event log. Restarting during selection resumes the saved parameter or next-step decision. An interrupted submission is reconciled by its full configuration; an ambiguous send is never retried automatically. A saved comparison or promotion is completed before another proposal begins.
Failed or cancelled simulations mark their draft failed and stop the loop. A later startup starts a new draft from the retained golden. Transport and database failures propagate with their cause.
The experiment-reset script clears the draft snapshots, setup trace, and saved measurements along with the original experiment data.