Same behavior.
Different route?

Measured · 64 calibration items · one training seed

The routes pass the matching rule.

Epoch 1 is the earliest checkpoint passing all three gates. The unchanged rule requires differences of at most 3 of 64 items in content preservation, uppercase events, and their intersection. This is operational matching, not a statistical equivalence test.

The uppercase adapter was trained to produce capitals despite an explicit instruction to preserve case. That conflict is part of this task. A normal-case adapter, selected at the same epoch, controls some generic training drift. Neither matching nor a failure to match identifies a mechanism.

All counts, paired item results and provenance ↗ JSON

Reproducibility check: all192 newly generated base-model outputs exactly matched the earlier screen and confirmation. Geometry and steering calibration are complete; final-test evaluation remains pending.

Measured geometry · 64 direction items · one training seed

Do prompt and training changes point the same way?

The changes align more closely in later layers. At the answer boundary, cosine is -0.04 at layer 0, 0.62 at layer 14 and 0.86 at layer 27. The final-layer values on shared normal-case and uppercase continuations are 0.88 and 0.74. This does not establish that either direction controls capitalization; steering calibration is reported below; held-out evaluation remains pending.

Each point compares the mean activation change from an uppercase instruction (base + uppercase minus base + neutral) with the mean change from uppercase training (uppercase adapter minus normal-case adapter, both under the neutral instruction). A cosine of 1 means parallel directions, 0 orthogonal, and −1 opposite.

Download geometry, bootstrap metadata and provenance ↗ JSON

Completed calibration · 19 September 2026 · 2,880 outputs

Removing capitals is not the same as restoring the answer.

All 40 nonzero candidates across four donor/recipient families have been evaluated. The frozen rule maximizes exact normal-case restoration subject to losing at most three content successes out of 64; families without improvement retain the zero intervention.

Both selected interventions restore 61/64 exact normal answers in the prompted model, versus 0/64 without intervention, while content changes from 62/64 to 61/64. Every tested intervention into the trained model restores 0/64 exact normal answers. Yet the training direction at layer 20, scale −1 removes all uppercase events while retaining content on 62/64: outputs become mixed or title case. That does not meet the restoration objective.

All candidate counts, selections and provenance ↗ JSON

With mixed-case input, the uppercase instruction produced the requested broad style on 16 of 16 items, but preserved the content on only 12 of 16. Supplying uppercase input improved uppercase copying, while neutral copying lost content.

A decoding check did not resolve this bottleneck: changing the repetition penalty from 1.1 to 1.0 changed none of 48 paired greedy outputs in v4; its control outputs also matched the historical v2 outputs. This is a text comparison, not evidence that the logits were identical.

New checkpoint · preliminary 7B screen

A larger model clears the first screen.

Qwen2.5-7B-Instruct, using the v2 mixed-case copying setup, passed the same marginal gates on the same 16 exploratory items. All three style counts were 16/16; content and joint success were lower, as shown below.

This is a descriptive comparison between model checkpoints, not a causal estimate of the effect of parameter count. The reused screen alone is insufficient: a separate 48-item confirmation also passed (below). No mechanism conclusion follows from passing a copying gate.

Measured · 48 previously unused calibration items

The fixed protocol passes confirmation.

The same 7B model, instructions and decoding settings were applied without changes to 48 additional items. Each condition exceeded the frozen 42/48 content threshold. Uppercase style succeeded on 47/48 outputs, lowercase style on 48/48, and neutral uppercase events were 0/48.

Uppercase content and style each succeeded on 47 items, but their intersection was 46: they failed on different responses. The gate establishes feasibility for this templated copying task, not prompt/training equivalence, error-free copying or general reasoning. These items were unused before confirmation; they now belong to calibration, not the final test.

Planned · conditional on feasibility

Compare the routes, then test transfer.

The core experiment needs two ways to elicit comparable behavior. A stronger uppercase rate alone would make the comparison ambiguous.

1. Match observable performance. Use calibration data separate from training and final testing to find conditions with similar style and content success. A mismatch must remain visible.
2. Measure internal changes. Compare changes in the model’s residual activations—the running internal representation—at each layer. Examine the answer boundary and identical supplied answer tokens separately.
3. Try the change in the other route. Add a prompt-derived activation direction to the trained model, and a training-derived direction to the base model. Include opposite signs, random directions and utility checks.
4. Evaluate once on held-out items. Freeze analysis choices first. Similar directions are not proof of a shared mechanism; functional transfer is a further test, with its own alternative explanations.

Context & provenance

This builds on existing work.

Persona Vectors already relates training-induced activation changes to prompt-derived directions and uses those directions to steer trained models. Concept Ablation Fine-Tuning (CAFT) already extracts training-related activation differences on shared text and tests inference ablations. Geometry and one-way transfer are established ideas, not a novelty claim here.

The narrower proposed question is when similarity between behavior-matched routes agrees—or disagrees—with transfer in both directions. The current geometry describes directional similarity; the suppression calibration below tests both directions; held-out functional transfer remains pending.

The figure dataset includes metric definitions, all counts, model revision, run identifiers, original workspace paths and SHA-256 hashes for source reports and code. Workspace paths in that file are provenance identifiers, not links in this standalone site. The build is reproducible from the workspace with python3 site/build.py.