Conduct controls and authoring#

Conduct controls apply fitted residual vectors or token-bias voice profiles to an agent. The model store holds their assets; recorded input retains each local selection for replay. This guide describes authoring and measurement commands, not automatic qualification of their output.

Before enabling a control#

A fitted vector, probe or calibration asset needs an exact model binding and qualification component. A derived file or potency measurement alone cannot enable a runtime control. The binding contract defines model, wrapper, position, layer and accepted-dose checks.

The current qualified residual vector is response-only curiosity for one exact Qwen3-4B Q4_0 file, at layer 24 and dose 0.5. Other model/control pairs remain unavailable until their own complete acceptance. The generic authoring commands below retain their declared legacy capture behavior; they do not reproduce that package automatically. See support.

Derive a steer vector#

aotx_steer_derive loads the language model through the normal model loader. Its pair file has one positive prompt, one tab, and one negative prompt on each line. It accepts at most 32 pairs. The layer list uses comma-separated zero-based layer numbers in ascending order, each named once. The rows of the vector file stand in that order, and the loader applies them in that order.

aotx_steer_derive --models models --trait directness \
  --pairs tests/fixtures/steer/directness.tsv --layers 8,16,24

The program captures the final residual stream of each named layer. It writes the mean positive minus negative residual as directness.aotxvec. It then uses the pair prompts as a probe set and measures the mean per-token KL divergence between the unsteered and steered next-token distributions. The unit is nats. The file enters steer.jsonl only with this potency figure. A file with no matching figure is refused with the reason on the error output.

The vector file starts with AOTXSTV1, the hidden width, the layer count, the potency, and a reserved value. Layer numbers and layer-major float values follow. Keep derived vector files in the model store. They are model data and are not source files.

Derive an affect axis#

The same program derives an affect axis. docs/14-affect.md states what an axis is and how the system reads one. The usage text names the three forms:

usage: aotx_steer_derive --models DIR --trait NAME --pairs FILE --layers LIST [--role NAME]
       aotx_steer_derive --models DIR --axis NAME --pairs FILE --neutral FILE --heldout FILE --layers LIST
                         [--standardise FILE] [--print-readouts] [--role NAME]
       aotx_steer_derive --models DIR --calibrate --axes LIST [--guards LIST] --neutral FILE --dose D [--surgical R] [--role NAME]

The axis name is valence, arousal, sycophancy or refusal. Their axis numbers in the probe table are 0, 1, 4 and 5. Another name is refused.

aotx_steer_derive --models models --axis valence \
  --pairs tests/fixtures/affect/valence.tsv \
  --neutral tests/fixtures/affect/neutral.txt \
  --heldout tests/fixtures/affect/heldout-valence.tsv --layers 8,12,16,20,24 \
  --standardise tests/fixtures/affect/plain-replies.txt

The role option opens another model file of the store. The 8g profile carries the role language-q4, so one build derives an axis on each file.

The model entry selects the probe layer with probe_numerator and probe_denominator. The defaults are 2 and 3, and the numerator must be less than the denominator. The layer is the layer count times the numerator, divided by the denominator, rounded down. For 36 layers, the default selects layer 24. A numerator of zero selects layer 0. The tool prints the selected layer and records it in the probe file header and catalog.

The tool captures a probe layer outside --layers without adding a steer candidate. A probe reads the residual row after the steer add of its own layer. Use steer candidates at or before the probe layer so that the probe can measure their effect. The probe readouts must have a spread. If the model fraction changes, derive the probes again. The loader refuses a stored probe whose layer differs from the selected layer.

At each named layer the program measures five figures. The steer direction is the mean positive residual less the mean negative residual. The probe direction is a discriminant of the same pair captures, standardized over the standardization set. The accuracy is the share of held-out pairs whose positive member reads above its negative member. The potency is the divergence of that layer's direction on its own, in nats.

The agreement is the share of held-out pairs whose positive member reads above the standardization mean and whose negative member reads below it. The trace subtracts that mean from the cosine, then divides by the probe scale.

The program prints one line for each layer, one line for the layer of the vector, and one line for the probe. The block below shows two of the five layer lines:

axis valence: role language, 36 layers, hidden 2560, vocabulary 151936, probe layer 24
axis valence layer 16: accuracy 1 agreement 0.875 potency 0.0820141509 nats mean 0.0141976932 scale 0.0273294579
axis valence layer 20: accuracy 1 agreement 1 potency 0.107885517 nats mean -0.00341951358 scale 0.0280698724
axis valence: layer 20 chosen, accuracy 1, agreement 1, 32 pairs, 145 standardization texts, 32 held-out pairs
axis valence: probe layer 24, accuracy 1, agreement 0.96875, mean -0.00707792491, scale 0.0241941214, standardized on tests/fixtures/affect/plain-replies.txt

--print-readouts adds one line for each text of the standardization set, before the layer lines:

readout arousal: text 1, layer 8, 0.0545607619

Those readouts are the source of the mean and the scale of the probe. Each readout is a cosine, so it stands between -1 and 1. A check recomputes the two figures on the host from the printed lines.

The chosen layer is the layer with the highest accuracy. An equal accuracy takes the higher agreement, then the earlier layer. A layer whose readouts have no spread over the standardization set is not chosen. The loader permits monitoring only when an otherwise admitted row fails the accuracy or agreement threshold. Guard axes cannot drive state. Missing qualification leaves an asset unavailable.

The axis mode writes these principal outputs:

• <models>/<axis>.aotxvec, the steer vector of the chosen layer, with its potency;
• <models>/affect/<axis>.aotxprb, the probe file of the probe layer;
• one line in <models>/probes.jsonl, the probe catalog.

The derivation also writes exact .binding sidecars. Keep them with each fitted asset. The probe file holds the width, the layer and the axis number. It then holds the accuracy, the agreement, the mean, the scale and the direction.

The program writes nothing when a file of that axis already stands in the store. It refuses these inputs as well:

• an unknown axis name;
• a set file that does not open;
• a standardization set of one text;
• a named layer that is not under the layer count of the model;
• a probe layer outside the model's supported layer range.

A program built without the affect option writes the same files. Its model load reads no catalog, so the rows stay on the disk.

Calibrate the axes#

--calibrate measures the response of the derived axes on one model file. It reads the steer vectors and the probe rows of the store and runs the neutral set through the model.

aotx_steer_derive --models models --calibrate --axes valence,arousal \
  --guards sycophancy,refusal --neutral tests/fixtures/affect/neutral.txt --dose 0.25

The program makes one pass of the neutral set for each variant. Two axes give nine variants. Six of them apply the steer vectors of the store:

• one plain pass;
• one pass for each axis at the dose;
• one pass with both axes at the dose;
• one pass for each axis at twice the dose.

Three of them apply the composites: one pass for each axis at the dose, and one pass with both composites at the dose. One axis gives four variants. They are the plain pass, the pass at the dose, the pass at twice the dose and one composite pass at the dose.

Each pass gives the logits of every row and the captures of every probe layer. The passes on the vectors of the store give M, K_raw, the dose-response ratios and the perplexity ratios. The composite passes give K, which is the matrix a run applies.

The program prints one line for each row and each figure, with pass or fail beside it. The block below is an extract of one run:

probe valence: axis 0, layer 24, accuracy 1 pass, agreement 0.96875 pass, mean -0.00707792491, scale 0.0241941214
M valence under valence: 3.01527214
M valence under arousal: 0.0730483904
probe arousal: axis 1, layer 24, accuracy 1 pass, agreement 0.71875 fail, mean -0.0280524325, scale 0.0256778635
M dominant: pass
K valence: 0.0823758766 nats per unit dose squared pass, raw 0.0823758766
dose-response valence: 4.70913124 pass
perplexity valence: 1.03841996 at the dose pass (bound 2), 1.09438813 at twice the dose
K off-diagonal: 0.0183836743, normalized 0.207519695 pass, raw 0.0183836743 normalized 0.207519695
composite valence: affect/composite-valence.aotxvec, 1 layers, potency 0.0411879383 nats
calibration line: models/affect/calibration.jsonl, figures finite, dominant 1, orthogonal 1

The program writes one composite vector file for each axis, and one line in <models>/affect/calibration.jsonl. The composite of an axis holds the rows of its own steer vector, made orthogonal to the other axis at each layer both hold. Each column keeps the length of its own vector, so a unit dose keeps its meaning. A composite is a copy of its own steer vector when the two axes hold no common layer. A later run overwrites a composite file, and the calibration file keeps every line.

The dominant mark comes from M, which the passes on the vectors of the store give. The orthogonal mark comes from the normalized off-diagonal of K, which the composite passes give. The model load reads the last line of the calibration file. It applies the composite only under both marks. docs/14-affect.md states that gate and the fields of the line.

A completed calculation can return status zero while a printed condition fails. Read every condition and retain its measurement. Do not treat process exit alone as qualification.

Score the capability#

aotx_quality_score measures what a steer vector costs the model on a task set. It opens the language model and nothing else. It scores the two sides of a paired run on a rubric as well. docs/15-quality.md states the program, its two modes, its files and the fixtures it reads. docs/14-affect.md states the measured scores.

Select conduct items#

The sampler retains legacy steer slots, but accepted runtime qualification currently permits one vector at a time. Only an exact accepted nonzero dose can run. Combined vectors, affect actuation or voice coupling need their own supported contract.

For the qualified package, use:

agent 0 decode.steer0 curiosity:0.5
agent 0 decode.steer0 absent

A missing component or unaccepted setting is refused. Native request selection exposes the same qualification to application clients.

The forward pass makes one fused add after the final feed-forward residual add of each named layer. No selected vector means that the kernel writes no residual value. This state preserves the prior forward result bit for bit.

The store has capacity for 16 vectors and 16 voice profiles; capacity does not establish qualification.

A local voice profile can be selected with agent 0 decode.voice NAME. A voice profile can hold 128 entries. The first line of a .profile file is its name. Each later line holds a bias, a tab, and one vocabulary string. Load tokenizes each string once. The sampler adds its bias before the softmax.

The token statistics stream records each selected token. A reader can therefore count the profile strings in one reply and state their frequency shift.

Page map#

Attention adds normalized attention mass to the key and value page that supplied each key. Every 64 ticks, the tick commit writes one class B page record for each resident page or page with mass. The record holds the agent, page, residency, cadence, and mass. The drain writes pages.jsonl with the tick, agent, page, residency, and mass.

The instrument makes a second key traversal only during an instrumented language pass. It adds one atomic value per head, key block, and page contribution. The 64-tick flush scans the fixed page table and reserves the worst-case record count. No eviction or model policy reads this table.

Model parameters#

A store scan writes parameters.jsonl. One line names a model and its declared sampling controls. A control enters the line only when the model metadata gives its default, min, and max values. The scanner recognizes temperature, top-p, minimum-p, top-k, repeat penalty, repeat window, presence penalty, and frequency penalty. It does not supply a missing value.