research draft
The Causal Bridge Between Colour Geometry and Model Behaviour
A 44-record replay on Qwen 3.5 27B shows that blue and yellow interventions reliably move response activations in opposite directions on the blue-yellow axis. The geometry is causal for latent direction, but not yet enough to predict judged task quality.
Replaying 44 existing blue/yellow records with hidden-state capture gave a perfect text match rate, so the activation readout is tied to the same outputs rather than a different rerun.
Full, residual, and blue-yellow-axis interventions aligned with the applied blue/yellow direction on every non-base row. Shared steering only aligned on half the rows, which is the expected failure mode for a shared persona component.
The stronger behavioural claim is not established yet. Cheap rubric scores did not correlate cleanly with projection magnitude, so this result supports causal activation movement, not score prediction.
The earlier notes showed two things separately: colour-persona interventions change model behaviour, and the vectors behind those interventions have a clean internal geometry. This note asks whether those two facts are connected. On Qwen 3.5 27B, the answer is yes for activation direction and not yet for judged output quality.
missing link
A geometry result is not automatically a control result.
The useful question is not only whether the blue-yellow axis exists. The useful question is whether applying blue or yellow actually moves the model state along that axis during the answer.
The previous geometry note gave us a clean two-axis picture: red <-> green
and blue <-> yellow. That is valuable, but by itself it is still mostly
descriptive. A cosine matrix can tell us that two vectors oppose each other.
It cannot tell us whether an intervention actually moves the live response in
the expected direction.
So the next diagnostic was deliberately small. We reused the repaired blue/yellow slice instead of broadening the benchmark again. The goal was to answer one narrower question:
- if the applied intervention is blue, does the response activation move in the blue direction?
- if the applied intervention is yellow, does it move in the yellow direction?
- does the shared component fail to separate blue from yellow, as it should?
That makes this a causal bridge test. The benchmark outputs are still the visible behaviour, but the main measurement is now the activation movement underneath them.
diagnostic design
Replay the exact same answers, then read the hidden states.
The point was to avoid making a new benchmark when the existing 44 records already contained the comparison we needed.
same outputs
The diagnostic replays the existing generated records with hidden-state capture and checks exact text equality against the saved benchmark outputs.
same layer
Projections are read at `L62`, the late-stack Qwen control layer used by the earlier steering and geometry work.
same vectors
The table compares full blue/yellow vectors, the shared component, residual colour vectors, and the explicit blue-yellow axis.
The replay used the existing blue/yellow diagnostic outputs:
- the full repaired run, including base, prompted blue/yellow, and full-vector steered blue/yellow rows
- the residual-steered run
- the shared-steered run
- the explicit blue-yellow-axis-steered run
In total, that gives 44 records. Each record was replayed through
Qwen 3.5 27B with hidden-state capture enabled. The generated text matched
the saved text exactly on every row, so the activation table is tied to the
same outputs we had already inspected.
The analysis then projected response-layer means at L62 onto:
- full blue and yellow vectors
- the shared component
- blue and yellow residual vectors
- the explicit blue-yellow axis
The important readout is simple: subtract the base projection for the same task, then check whether the full blue-yellow axis delta points toward the applied colour.
main result
The axis direction follows the intervention.
This is the clean mechanistic signal. Full, residual, and explicit axis steering all separate blue from yellow. Shared steering does not.
full vectors
16 / 16
The sign of the blue-yellow axis delta matched the applied colour on every prompted and steered full-vector row.
residual vectors
8 / 8
Residual steering preserved the blue/yellow direction after subtracting the shared component.
axis vectors
8 / 8
Direct blue-yellow-axis steering produced the expected directional movement on every row.
shared vector
4 / 8
Shared steering moved state but did not separate blue from yellow, which is exactly what a shared component should do.
The shared result is the most useful sanity check. If the shared component had separated blue from yellow cleanly, that would have been suspicious. The shared vector is supposed to capture the common persona/control movement, not the colour-specific opposition. In this diagnostic it aligned on only half of the blue/yellow rows, which is exactly the failure mode we wanted.
Residual and axis steering behave differently. Once the shared component is removed, or once the blue-yellow opposition is used directly, the sign of the axis movement tracks the applied colour on every row in this slice.
negative result
Projection direction is real. Projection magnitude is not enough.
The tempting overclaim would be that a larger target projection means a better answer. This diagnostic does not support that yet.
The score side is intentionally marked as provisional. The current scoring is a cheap rubric heuristic used to smoke-test the table, not the judged protocol used in the first milestone. Under that cheap scorer, projection magnitude does not cleanly predict task success or visible colour expression.
That is not a failure of the geometry result. It is a boundary on what the result means. The axis direction tells us that the intervention moved the model state in the expected latent direction. It does not tell us that the resulting answer was better, more complete, or more visibly blue or yellow.
The distinction matters. A steering vector can move the model into the right region of activation space while the prompt, task constraints, answer length, or decoding path still determine whether the final output is useful.
interpretation
The causal bridge is narrower than the benchmark claim.
This is the point where the project gets cleaner: the geometry can now explain intervention direction, but not all behavioural quality.
supported
The blue-yellow axis is causal for activation direction in Qwen 27B under this replay diagnostic.
not supported yet
Projection magnitude does not yet predict judged task success or colour-expression quality.
best next move
Replace the cheap heuristic scores with manual rubric scores, then rerun only the summary step.
The clean claim is now:
- the blue-yellow axis is not only descriptive geometry
- applying blue or yellow moves Qwen 27B response activations in the expected direction
- residual steering preserves that colour-specific direction
- the shared component does not separate blue from yellow
The claim we should not make yet is:
- larger target projection means a better answer
- projection magnitude alone predicts task success
- activation direction fully explains colour expression at the surface
That second list may become true under better scoring, but this diagnostic does not establish it. It gives us the next controlled question instead of forcing a bigger story than the evidence can hold.
what changes
The next step is scoring, not another benchmark.
The table is already the right size. Broadening now would make the interpretation noisier, not stronger.
publishable now
Geometry can be used as a causal readout of intervention direction.
Shared and residual components behave differently.
The blue-yellow axis is a cleaner diagnostic object than raw target projection.
This is a mechanistic result, not a benchmark-quality result.
needs one more pass
Manual-score the same 44 records.
Recompute projection-versus-score correlations.
Only then decide whether projection magnitude predicts behaviour.
Do not broaden the benchmark before this scoring pass. The current table is already the right size.
The lazy version of the next milestone is one file: a manual score JSONL for
these exact 44 records. Then the summary can be recomputed without another
model run. If manual scores correlate with projection deltas, we have a
stronger behavioural bridge. If they do not, the result is still publishable,
but the claim changes:
- geometry causally controls latent direction
- task-quality gains are mediated by the prompt and task constraints
- steering needs a behavioural readout, not only a vector readout
That is a useful outcome either way. It tells us where the mechanism is real and where the measurement still needs work.
conclusion
The colour axis moves the model. The answer quality still needs judgment.
This is the right third milestone because it connects the first two. The benchmark showed that colour interventions can change task behaviour. The geometry note showed that the colour vectors have a stable internal structure. This replay shows that the structure is not passive: applying blue or yellow moves the response activations in the corresponding direction.
But the result is narrower than a full behavioural theory. The axis direction is causal. The score prediction is not established. That is the line to keep. It makes the project stronger, because the next step is now precise instead of expansive: score the same records properly, and only then decide whether projection magnitude predicts the quality of the answer.