will-taggart commited on
Commit
edc0025
·
verified ·
1 Parent(s): d768773

Illustrate technical papers with authentic application screenshots

Browse files
.gitattributes CHANGED
@@ -38,3 +38,8 @@ docs/acm/AFFECTOR-ACM.pdf filter=lfs diff=lfs merge=lfs -text
38
  docs/AFFECTOR-Technical-Paper.pdf filter=lfs diff=lfs merge=lfs -text
39
  paper/AFFECTOR-ACM.pdf filter=lfs diff=lfs merge=lfs -text
40
  paper/AFFECTOR-Technical-Report-ACM.pdf filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
38
  docs/AFFECTOR-Technical-Paper.pdf filter=lfs diff=lfs merge=lfs -text
39
  paper/AFFECTOR-ACM.pdf filter=lfs diff=lfs merge=lfs -text
40
  paper/AFFECTOR-Technical-Report-ACM.pdf filter=lfs diff=lfs merge=lfs -text
41
+ paper/figures/screenshot-01.png filter=lfs diff=lfs merge=lfs -text
42
+ paper/figures/screenshot-02.png filter=lfs diff=lfs merge=lfs -text
43
+ paper/figures/screenshot-03.png filter=lfs diff=lfs merge=lfs -text
44
+ paper/figures/screenshot-04.png filter=lfs diff=lfs merge=lfs -text
45
+ paper/figures/screenshot-05.png filter=lfs diff=lfs merge=lfs -text
SHA256SUMS.txt CHANGED
@@ -1,24 +1,24 @@
1
  321cd041f8060331e034b2a72578f57cd3e498883e660016cf73f0a8ba69a9a6 README.md
2
- 5f080cff35f877bb791008a3b455a5d7544fc57378feacf64da5b7f4402fafac docs/AFFECTOR-Technical-Paper.pdf
3
- 9e27dcda981ca2e981a53254c58c219e471eb67f333a2c8319ee132e02ee8231 docs/TECHNICAL_PAPER.md
4
- 9e27dcda981ca2e981a53254c58c219e471eb67f333a2c8319ee132e02ee8231 docs/TECHNICAL_REPORT.md
5
  ed8fa321af4fc4e4703efcc22037e244fd072c936d11afc566b8eb862f3f0b1b downloads/AFFECTOR-macOS-AppleSilicon.zip
6
  d7621338f71fdc31ce9b6e1d53f15d61d6a76344f2111520fb62f9b00293853b downloads/AFFECTOR-source.zip
7
- 64526e21d1658e6e9a40213c53141c257b44e1a4503eb71a30fb17bb8c2633bf paper/AFFECTOR-ACM-source.zip
8
  d78577482f74cba4163db18d88af82786c923d72cd09607bde62cb747c645c5d paper/AFFECTOR-ACM.bbl
9
- 5f080cff35f877bb791008a3b455a5d7544fc57378feacf64da5b7f4402fafac paper/AFFECTOR-ACM.pdf
10
- 64526e21d1658e6e9a40213c53141c257b44e1a4503eb71a30fb17bb8c2633bf paper/AFFECTOR-ACM.source.zip
11
- 055d38b133c255a1fa228ce066b0f7df71f3e8656e9a523229a2ae9e3c219649 paper/AFFECTOR-ACM.tex
12
- 64526e21d1658e6e9a40213c53141c257b44e1a4503eb71a30fb17bb8c2633bf paper/AFFECTOR-Technical-Report-ACM-source.zip
13
  d78577482f74cba4163db18d88af82786c923d72cd09607bde62cb747c645c5d paper/AFFECTOR-Technical-Report-ACM.bbl
14
- 5f080cff35f877bb791008a3b455a5d7544fc57378feacf64da5b7f4402fafac paper/AFFECTOR-Technical-Report-ACM.pdf
15
- 64526e21d1658e6e9a40213c53141c257b44e1a4503eb71a30fb17bb8c2633bf paper/AFFECTOR-Technical-Report-ACM.source.zip
16
- 055d38b133c255a1fa228ce066b0f7df71f3e8656e9a523229a2ae9e3c219649 paper/AFFECTOR-Technical-Report-ACM.tex
17
- e6b2505c263c35b5d7dd545726bf4bf2de4e109b7c05abb16f9f1f6bdb3a2972 paper/ILLUSTRATIONS.json
18
  d14edda139a36c0bac8b56d02e5cda63ab7558c6db1424f1087189a9429bfaf0 paper/LITERATURE_SOURCES.md
19
  578a471b301cdf67305c2b997fce39494c2fd0d15eaf56836379943b3d10f90c paper/PAPER-SHA256SUMS.txt
20
  22f8f7bf6992778eb1190ee02b6e83dacbd54ff34af7777a18b4e27a195309a3 paper/README.md
21
- e2148c0d8e58cdc4024151b19ac08625e62355585b9e46011fa68eee9cd48188 paper/TECHNICAL_PAPER.md
22
  4bd1a1b7f6af31b12b9c92bcd978fe116482892b3463f8ec9c922bc1f636c7a0 paper/figure-specifications.json
23
  b11798a28d94d60110a7b19f58107d5e60926b68265f64e62f59e9230a8f500d paper/figures/illustration-01.pdf
24
  d2ae42a17ff3f16b89a7cb85d21ac1f3c7a19c2bc62d61f1cc581c2ca6d3e596 paper/figures/illustration-01.png
@@ -32,4 +32,4 @@ a6a497a38f4ba9de8088f410535bee13dd862d1a97f4175d205df768e1d2c1c9 paper/figures/
32
  7d8f30b829993326e8d764ae18b3974acddbd86b760d4071bdff4edfcb9b299a paper/figures/illustration-05.png
33
  9257f38423bd7e496f1a0916c0fbaefbfb32ab544a684ca9a411508201f0535e paper/references.bib
34
  eed8b979a5a299b42e925d9027332e7df876400dcef046f2cfe26beef91d1189 paper/render_figures.py
35
- e2148c0d8e58cdc4024151b19ac08625e62355585b9e46011fa68eee9cd48188 paper/source.md
 
1
  321cd041f8060331e034b2a72578f57cd3e498883e660016cf73f0a8ba69a9a6 README.md
2
+ c683afbacffddad98e72525aa0593e8e1695c3bfca526a41a852cfc2baff5e8a docs/AFFECTOR-Technical-Paper.pdf
3
+ b354dd4c97e9521760dfe0c4a2e16bc4762d672b3186132ab6717764c86caff6 docs/TECHNICAL_PAPER.md
4
+ b354dd4c97e9521760dfe0c4a2e16bc4762d672b3186132ab6717764c86caff6 docs/TECHNICAL_REPORT.md
5
  ed8fa321af4fc4e4703efcc22037e244fd072c936d11afc566b8eb862f3f0b1b downloads/AFFECTOR-macOS-AppleSilicon.zip
6
  d7621338f71fdc31ce9b6e1d53f15d61d6a76344f2111520fb62f9b00293853b downloads/AFFECTOR-source.zip
7
+ cd28bc036c9a8abf0e5cc566197ff55c83ed0656f5a87363c72c3baf8cee8256 paper/AFFECTOR-ACM-source.zip
8
  d78577482f74cba4163db18d88af82786c923d72cd09607bde62cb747c645c5d paper/AFFECTOR-ACM.bbl
9
+ c683afbacffddad98e72525aa0593e8e1695c3bfca526a41a852cfc2baff5e8a paper/AFFECTOR-ACM.pdf
10
+ cd28bc036c9a8abf0e5cc566197ff55c83ed0656f5a87363c72c3baf8cee8256 paper/AFFECTOR-ACM.source.zip
11
+ 7ec8d6a693c2de155d62d434bef4c387f911c75f0dceadbd1de9213b9c7691e2 paper/AFFECTOR-ACM.tex
12
+ cd28bc036c9a8abf0e5cc566197ff55c83ed0656f5a87363c72c3baf8cee8256 paper/AFFECTOR-Technical-Report-ACM-source.zip
13
  d78577482f74cba4163db18d88af82786c923d72cd09607bde62cb747c645c5d paper/AFFECTOR-Technical-Report-ACM.bbl
14
+ c683afbacffddad98e72525aa0593e8e1695c3bfca526a41a852cfc2baff5e8a paper/AFFECTOR-Technical-Report-ACM.pdf
15
+ cd28bc036c9a8abf0e5cc566197ff55c83ed0656f5a87363c72c3baf8cee8256 paper/AFFECTOR-Technical-Report-ACM.source.zip
16
+ 7ec8d6a693c2de155d62d434bef4c387f911c75f0dceadbd1de9213b9c7691e2 paper/AFFECTOR-Technical-Report-ACM.tex
17
+ e2c0590754823491af4657a764ab6682e27c5eaf0173fc9f5f03a62e84ccb308 paper/ILLUSTRATIONS.json
18
  d14edda139a36c0bac8b56d02e5cda63ab7558c6db1424f1087189a9429bfaf0 paper/LITERATURE_SOURCES.md
19
  578a471b301cdf67305c2b997fce39494c2fd0d15eaf56836379943b3d10f90c paper/PAPER-SHA256SUMS.txt
20
  22f8f7bf6992778eb1190ee02b6e83dacbd54ff34af7777a18b4e27a195309a3 paper/README.md
21
+ 4b62ddd1788644df9f817c2b10fb649239e5263e22fbc381af3b6f8b411e504f paper/TECHNICAL_PAPER.md
22
  4bd1a1b7f6af31b12b9c92bcd978fe116482892b3463f8ec9c922bc1f636c7a0 paper/figure-specifications.json
23
  b11798a28d94d60110a7b19f58107d5e60926b68265f64e62f59e9230a8f500d paper/figures/illustration-01.pdf
24
  d2ae42a17ff3f16b89a7cb85d21ac1f3c7a19c2bc62d61f1cc581c2ca6d3e596 paper/figures/illustration-01.png
 
32
  7d8f30b829993326e8d764ae18b3974acddbd86b760d4071bdff4edfcb9b299a paper/figures/illustration-05.png
33
  9257f38423bd7e496f1a0916c0fbaefbfb32ab544a684ca9a411508201f0535e paper/references.bib
34
  eed8b979a5a299b42e925d9027332e7df876400dcef046f2cfe26beef91d1189 paper/render_figures.py
35
+ 4b62ddd1788644df9f817c2b10fb649239e5263e22fbc381af3b6f8b411e504f paper/source.md
docs/AFFECTOR-Technical-Paper.pdf CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:5f080cff35f877bb791008a3b455a5d7544fc57378feacf64da5b7f4402fafac
3
- size 195948
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c683afbacffddad98e72525aa0593e8e1695c3bfca526a41a852cfc2baff5e8a
3
+ size 2132249
docs/TECHNICAL_PAPER.md CHANGED
@@ -1,6 +1,6 @@
1
  # AFFECTOR: Coordinate-Guided Text Rewriting through Transformer Activation Interventions
2
 
3
- **Technical paper:** [PDF](../paper/AFFECTOR-ACM.pdf) · [LaTeX source](../paper/AFFECTOR-ACM.tex) · [Complete source ZIP](../paper/AFFECTOR-ACM-source.zip)
4
 
5
  **William Taggart**
6
  Modern and Classical Languages and Literatures, Virginia Tech
@@ -11,6 +11,28 @@ November 13, 2026 · Implementation report · Desktop version 0.1.0
11
 
12
  AFFECTOR investigates whether a prescribed emotional trajectory can influence the rewriting of a document. A trajectory assigns valence, emotional direction, and intensity to normalized positions in a text. A local backend maps these coordinates into perturbations of a frozen language model's internal states, while hosted backends use coordinate-bearing instructions. The desktop application supports geometry authoring, editable numerical steering matrices, document conversion, checkpointing, and separate emotional verification. This paper derives the implemented method from first principles and distinguishes mechanical correctness from successful emotional transfer. Current controls can produce emotionally charged language but can also produce weak paraphrases, factual distortions, and repetition. The implementation is an experimental instrument, not a validated predictor or controller of human emotion.
13
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
14
  ## Literature review: activation steering and emotion
15
 
16
  This focused review covers twelve papers, emphasizing 2025-2026 research and retaining two methodological foundations. Titles, author lists, version histories, and the claims summarized here were checked against arXiv records and the corresponding paper text on November 13, 2026. The cited versions are fixed below; availability on arXiv alone does not establish peer review. The implications for this project are research motivations, not evidence that its implementation reproduces these results.
@@ -71,9 +93,6 @@ The resulting `B[l]` is a 3 by 1,024 matrix. These are the editable steering-mod
71
 
72
  There is no gradient training, score-supervised fitting, orthogonalization, or guarantee that the three directions isolate their intended concepts. Small hand-authored pairs can confound affect with genre, topic, lexical content, and emotional categories. Calibration is exploratory and English-based; multilingual effectiveness has not been established.
73
 
74
- ![Contrastive calibration supplies three directions](../paper/figures/illustration-02.png)
75
-
76
- *Illustration 2. Calibration workflow. Small hand-authored contrasts estimate steering directions; normalization neither orthogonalizes them nor establishes emotional validity.*
77
 
78
  ## 4. Coordinate-to-activation computation
79
 
@@ -99,9 +118,6 @@ Zero strength, budget, or layer weight gives an exact zero perturbation. Each ch
99
 
100
  Budget is not a semantic-change percentage. At budget 1, the perturbation can have RMS magnitude comparable to the original hidden state at that site. Large values can destabilize generation. Raising the budget only matters when the proposed perturbation would otherwise be clipped; it cannot correct an inaccurate direction.
101
 
102
- ![Controls shape the activation displacement](../paper/figures/illustration-03.png)
103
-
104
- *Illustration 3. Control pipeline. Axis controls operate on three emotional coefficients, not vocabulary logits. Raising budget has no effect while the proposal already lies below the ceiling.*
105
 
106
  ## 5. Intervention within the transformer
107
 
@@ -119,9 +135,6 @@ The module does not search for trigger phrases or recognize an idiom before acti
119
 
120
  A key limitation is that the intervention is an absolute-coordinate bias, not a measured error correction. A target valence of zero yields no direct valence-axis perturbation; it does not actively neutralize an already negative passage. Correlations between basis vectors further complicate interpretation.
121
 
122
- ![A change to both feed-forward branches](../paper/figures/illustration-04.png)
123
-
124
- *Illustration 4. Hook implementation schematic. The final prompt token and current cached-decoding token receive the intervention; hooks are removed on exit. Hosted API mode instead uses coordinate-bearing instructions.*
125
 
126
  ## 6. Prompts, sampling, and bounded output
127
 
@@ -149,9 +162,6 @@ The denominator is the maximum distance across the three-dimensional `[-1,1]` cu
149
 
150
  The same small model can assess its own output, but shared weaknesses make independent evaluation preferable. A high similarity can coexist with incoherence, factual errors, or undesirable content. The implementation does not optimize this score during generation and does not include a formal semantic-fidelity metric.
151
 
152
- ![Verification is a separate assessment](../paper/figures/illustration-05.png)
153
-
154
- *Illustration 5. Verification workflow. Failed or direction-unknown assessments are not assigned invented scores, and this score is not an iterative feedback objective during rewriting.*
155
 
156
  ## 8. Architecture, persistence, and privacy
157
 
@@ -191,7 +201,7 @@ The equations and implementation descriptions above are derived from the include
191
 
192
  ## Research references
193
 
194
- Version-specific arXiv records are linked below. Years distinguish original release from later revision where applicable. These research references supplement the implementation and infrastructure sources elsewhere in this report. Reusable citations: [BibTeX](../paper/references.bib). Verification details: [source-check notes](../paper/LITERATURE_SOURCES.md).
195
 
196
  **R1.** Alex Turner; Lisa Thiergart; David Udell; Gavin Leech; Ulisse Mini; Monte MacDiarmid. 2023. Activation Addition: Steering Language Models Without Optimization. [arXiv:2308.10248v1](https://arxiv.org/abs/2308.10248v1). Cited version dated 2023-08-20.
197
 
 
1
  # AFFECTOR: Coordinate-Guided Text Rewriting through Transformer Activation Interventions
2
 
3
+ **Technical paper:** [PDF](AFFECTOR-ACM.pdf) · [LaTeX source](AFFECTOR-ACM.tex) · [Complete source ZIP](AFFECTOR-ACM-source.zip)
4
 
5
  **William Taggart**
6
  Modern and Classical Languages and Literatures, Virginia Tech
 
11
 
12
  AFFECTOR investigates whether a prescribed emotional trajectory can influence the rewriting of a document. A trajectory assigns valence, emotional direction, and intensity to normalized positions in a text. A local backend maps these coordinates into perturbations of a frozen language model's internal states, while hosted backends use coordinate-bearing instructions. The desktop application supports geometry authoring, editable numerical steering matrices, document conversion, checkpointing, and separate emotional verification. This paper derives the implemented method from first principles and distinguishes mechanical correctness from successful emotional transfer. Current controls can produce emotionally charged language but can also produce weak paraphrases, factual distortions, and repetition. The implementation is an experimental instrument, not a validated predictor or controller of human emotion.
13
 
14
+ ## Application interface
15
+
16
+ ![Document workspace with an example input, rewrite controls, and an empty output pane before generation.](../paper/figures/screenshot-01.png)
17
+
18
+ *Document workspace with an example input, rewrite controls, and an empty output pane before generation.*
19
+
20
+ ![Geometry editor showing the selected three-point trajectory and its steering controls.](../paper/figures/screenshot-02.png)
21
+
22
+ *Geometry editor showing the selected three-point trajectory and its steering controls.*
23
+
24
+ ![Source-analysis dialog for generating a geometry and optionally calibrating a local activation basis.](../paper/figures/screenshot-03.png)
25
+
26
+ *Source-analysis dialog for generating a geometry and optionally calibrating a local activation basis.*
27
+
28
+ ![Expanded numerical weights for the selected geometry, including the local calibration metadata.](../paper/figures/screenshot-04.png)
29
+
30
+ *Expanded numerical weights for the selected geometry, including the local calibration metadata.*
31
+
32
+ ![Verification controls before evaluation. No verification results have been computed in this view.](../paper/figures/screenshot-05.png)
33
+
34
+ *Verification controls before evaluation. No verification results have been computed in this view.*
35
+
36
  ## Literature review: activation steering and emotion
37
 
38
  This focused review covers twelve papers, emphasizing 2025-2026 research and retaining two methodological foundations. Titles, author lists, version histories, and the claims summarized here were checked against arXiv records and the corresponding paper text on November 13, 2026. The cited versions are fixed below; availability on arXiv alone does not establish peer review. The implications for this project are research motivations, not evidence that its implementation reproduces these results.
 
93
 
94
  There is no gradient training, score-supervised fitting, orthogonalization, or guarantee that the three directions isolate their intended concepts. Small hand-authored pairs can confound affect with genre, topic, lexical content, and emotional categories. Calibration is exploratory and English-based; multilingual effectiveness has not been established.
95
 
 
 
 
96
 
97
  ## 4. Coordinate-to-activation computation
98
 
 
118
 
119
  Budget is not a semantic-change percentage. At budget 1, the perturbation can have RMS magnitude comparable to the original hidden state at that site. Large values can destabilize generation. Raising the budget only matters when the proposed perturbation would otherwise be clipped; it cannot correct an inaccurate direction.
120
 
 
 
 
121
 
122
  ## 5. Intervention within the transformer
123
 
 
135
 
136
  A key limitation is that the intervention is an absolute-coordinate bias, not a measured error correction. A target valence of zero yields no direct valence-axis perturbation; it does not actively neutralize an already negative passage. Correlations between basis vectors further complicate interpretation.
137
 
 
 
 
138
 
139
  ## 6. Prompts, sampling, and bounded output
140
 
 
162
 
163
  The same small model can assess its own output, but shared weaknesses make independent evaluation preferable. A high similarity can coexist with incoherence, factual errors, or undesirable content. The implementation does not optimize this score during generation and does not include a formal semantic-fidelity metric.
164
 
 
 
 
165
 
166
  ## 8. Architecture, persistence, and privacy
167
 
 
201
 
202
  ## Research references
203
 
204
+ Version-specific arXiv records are linked below. Years distinguish original release from later revision where applicable. These research references supplement the implementation and infrastructure sources elsewhere in this report. Reusable citations: [BibTeX](references.bib). Verification details: [source-check notes](LITERATURE_SOURCES.md).
205
 
206
  **R1.** Alex Turner; Lisa Thiergart; David Udell; Gavin Leech; Ulisse Mini; Monte MacDiarmid. 2023. Activation Addition: Steering Language Models Without Optimization. [arXiv:2308.10248v1](https://arxiv.org/abs/2308.10248v1). Cited version dated 2023-08-20.
207
 
docs/TECHNICAL_REPORT.md CHANGED
@@ -1,6 +1,6 @@
1
  # AFFECTOR: Coordinate-Guided Text Rewriting through Transformer Activation Interventions
2
 
3
- **Technical paper:** [PDF](../paper/AFFECTOR-ACM.pdf) · [LaTeX source](../paper/AFFECTOR-ACM.tex) · [Complete source ZIP](../paper/AFFECTOR-ACM-source.zip)
4
 
5
  **William Taggart**
6
  Modern and Classical Languages and Literatures, Virginia Tech
@@ -11,6 +11,28 @@ November 13, 2026 · Implementation report · Desktop version 0.1.0
11
 
12
  AFFECTOR investigates whether a prescribed emotional trajectory can influence the rewriting of a document. A trajectory assigns valence, emotional direction, and intensity to normalized positions in a text. A local backend maps these coordinates into perturbations of a frozen language model's internal states, while hosted backends use coordinate-bearing instructions. The desktop application supports geometry authoring, editable numerical steering matrices, document conversion, checkpointing, and separate emotional verification. This paper derives the implemented method from first principles and distinguishes mechanical correctness from successful emotional transfer. Current controls can produce emotionally charged language but can also produce weak paraphrases, factual distortions, and repetition. The implementation is an experimental instrument, not a validated predictor or controller of human emotion.
13
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
14
  ## Literature review: activation steering and emotion
15
 
16
  This focused review covers twelve papers, emphasizing 2025-2026 research and retaining two methodological foundations. Titles, author lists, version histories, and the claims summarized here were checked against arXiv records and the corresponding paper text on November 13, 2026. The cited versions are fixed below; availability on arXiv alone does not establish peer review. The implications for this project are research motivations, not evidence that its implementation reproduces these results.
@@ -71,9 +93,6 @@ The resulting `B[l]` is a 3 by 1,024 matrix. These are the editable steering-mod
71
 
72
  There is no gradient training, score-supervised fitting, orthogonalization, or guarantee that the three directions isolate their intended concepts. Small hand-authored pairs can confound affect with genre, topic, lexical content, and emotional categories. Calibration is exploratory and English-based; multilingual effectiveness has not been established.
73
 
74
- ![Contrastive calibration supplies three directions](../paper/figures/illustration-02.png)
75
-
76
- *Illustration 2. Calibration workflow. Small hand-authored contrasts estimate steering directions; normalization neither orthogonalizes them nor establishes emotional validity.*
77
 
78
  ## 4. Coordinate-to-activation computation
79
 
@@ -99,9 +118,6 @@ Zero strength, budget, or layer weight gives an exact zero perturbation. Each ch
99
 
100
  Budget is not a semantic-change percentage. At budget 1, the perturbation can have RMS magnitude comparable to the original hidden state at that site. Large values can destabilize generation. Raising the budget only matters when the proposed perturbation would otherwise be clipped; it cannot correct an inaccurate direction.
101
 
102
- ![Controls shape the activation displacement](../paper/figures/illustration-03.png)
103
-
104
- *Illustration 3. Control pipeline. Axis controls operate on three emotional coefficients, not vocabulary logits. Raising budget has no effect while the proposal already lies below the ceiling.*
105
 
106
  ## 5. Intervention within the transformer
107
 
@@ -119,9 +135,6 @@ The module does not search for trigger phrases or recognize an idiom before acti
119
 
120
  A key limitation is that the intervention is an absolute-coordinate bias, not a measured error correction. A target valence of zero yields no direct valence-axis perturbation; it does not actively neutralize an already negative passage. Correlations between basis vectors further complicate interpretation.
121
 
122
- ![A change to both feed-forward branches](../paper/figures/illustration-04.png)
123
-
124
- *Illustration 4. Hook implementation schematic. The final prompt token and current cached-decoding token receive the intervention; hooks are removed on exit. Hosted API mode instead uses coordinate-bearing instructions.*
125
 
126
  ## 6. Prompts, sampling, and bounded output
127
 
@@ -149,9 +162,6 @@ The denominator is the maximum distance across the three-dimensional `[-1,1]` cu
149
 
150
  The same small model can assess its own output, but shared weaknesses make independent evaluation preferable. A high similarity can coexist with incoherence, factual errors, or undesirable content. The implementation does not optimize this score during generation and does not include a formal semantic-fidelity metric.
151
 
152
- ![Verification is a separate assessment](../paper/figures/illustration-05.png)
153
-
154
- *Illustration 5. Verification workflow. Failed or direction-unknown assessments are not assigned invented scores, and this score is not an iterative feedback objective during rewriting.*
155
 
156
  ## 8. Architecture, persistence, and privacy
157
 
@@ -191,7 +201,7 @@ The equations and implementation descriptions above are derived from the include
191
 
192
  ## Research references
193
 
194
- Version-specific arXiv records are linked below. Years distinguish original release from later revision where applicable. These research references supplement the implementation and infrastructure sources elsewhere in this report. Reusable citations: [BibTeX](../paper/references.bib). Verification details: [source-check notes](../paper/LITERATURE_SOURCES.md).
195
 
196
  **R1.** Alex Turner; Lisa Thiergart; David Udell; Gavin Leech; Ulisse Mini; Monte MacDiarmid. 2023. Activation Addition: Steering Language Models Without Optimization. [arXiv:2308.10248v1](https://arxiv.org/abs/2308.10248v1). Cited version dated 2023-08-20.
197
 
 
1
  # AFFECTOR: Coordinate-Guided Text Rewriting through Transformer Activation Interventions
2
 
3
+ **Technical paper:** [PDF](AFFECTOR-ACM.pdf) · [LaTeX source](AFFECTOR-ACM.tex) · [Complete source ZIP](AFFECTOR-ACM-source.zip)
4
 
5
  **William Taggart**
6
  Modern and Classical Languages and Literatures, Virginia Tech
 
11
 
12
  AFFECTOR investigates whether a prescribed emotional trajectory can influence the rewriting of a document. A trajectory assigns valence, emotional direction, and intensity to normalized positions in a text. A local backend maps these coordinates into perturbations of a frozen language model's internal states, while hosted backends use coordinate-bearing instructions. The desktop application supports geometry authoring, editable numerical steering matrices, document conversion, checkpointing, and separate emotional verification. This paper derives the implemented method from first principles and distinguishes mechanical correctness from successful emotional transfer. Current controls can produce emotionally charged language but can also produce weak paraphrases, factual distortions, and repetition. The implementation is an experimental instrument, not a validated predictor or controller of human emotion.
13
 
14
+ ## Application interface
15
+
16
+ ![Document workspace with an example input, rewrite controls, and an empty output pane before generation.](../paper/figures/screenshot-01.png)
17
+
18
+ *Document workspace with an example input, rewrite controls, and an empty output pane before generation.*
19
+
20
+ ![Geometry editor showing the selected three-point trajectory and its steering controls.](../paper/figures/screenshot-02.png)
21
+
22
+ *Geometry editor showing the selected three-point trajectory and its steering controls.*
23
+
24
+ ![Source-analysis dialog for generating a geometry and optionally calibrating a local activation basis.](../paper/figures/screenshot-03.png)
25
+
26
+ *Source-analysis dialog for generating a geometry and optionally calibrating a local activation basis.*
27
+
28
+ ![Expanded numerical weights for the selected geometry, including the local calibration metadata.](../paper/figures/screenshot-04.png)
29
+
30
+ *Expanded numerical weights for the selected geometry, including the local calibration metadata.*
31
+
32
+ ![Verification controls before evaluation. No verification results have been computed in this view.](../paper/figures/screenshot-05.png)
33
+
34
+ *Verification controls before evaluation. No verification results have been computed in this view.*
35
+
36
  ## Literature review: activation steering and emotion
37
 
38
  This focused review covers twelve papers, emphasizing 2025-2026 research and retaining two methodological foundations. Titles, author lists, version histories, and the claims summarized here were checked against arXiv records and the corresponding paper text on November 13, 2026. The cited versions are fixed below; availability on arXiv alone does not establish peer review. The implications for this project are research motivations, not evidence that its implementation reproduces these results.
 
93
 
94
  There is no gradient training, score-supervised fitting, orthogonalization, or guarantee that the three directions isolate their intended concepts. Small hand-authored pairs can confound affect with genre, topic, lexical content, and emotional categories. Calibration is exploratory and English-based; multilingual effectiveness has not been established.
95
 
 
 
 
96
 
97
  ## 4. Coordinate-to-activation computation
98
 
 
118
 
119
  Budget is not a semantic-change percentage. At budget 1, the perturbation can have RMS magnitude comparable to the original hidden state at that site. Large values can destabilize generation. Raising the budget only matters when the proposed perturbation would otherwise be clipped; it cannot correct an inaccurate direction.
120
 
 
 
 
121
 
122
  ## 5. Intervention within the transformer
123
 
 
135
 
136
  A key limitation is that the intervention is an absolute-coordinate bias, not a measured error correction. A target valence of zero yields no direct valence-axis perturbation; it does not actively neutralize an already negative passage. Correlations between basis vectors further complicate interpretation.
137
 
 
 
 
138
 
139
  ## 6. Prompts, sampling, and bounded output
140
 
 
162
 
163
  The same small model can assess its own output, but shared weaknesses make independent evaluation preferable. A high similarity can coexist with incoherence, factual errors, or undesirable content. The implementation does not optimize this score during generation and does not include a formal semantic-fidelity metric.
164
 
 
 
 
165
 
166
  ## 8. Architecture, persistence, and privacy
167
 
 
201
 
202
  ## Research references
203
 
204
+ Version-specific arXiv records are linked below. Years distinguish original release from later revision where applicable. These research references supplement the implementation and infrastructure sources elsewhere in this report. Reusable citations: [BibTeX](references.bib). Verification details: [source-check notes](LITERATURE_SOURCES.md).
205
 
206
  **R1.** Alex Turner; Lisa Thiergart; David Udell; Gavin Leech; Ulisse Mini; Monte MacDiarmid. 2023. Activation Addition: Steering Language Models Without Optimization. [arXiv:2308.10248v1](https://arxiv.org/abs/2308.10248v1). Cited version dated 2023-08-20.
207
 
paper/AFFECTOR-ACM-source.zip CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:64526e21d1658e6e9a40213c53141c257b44e1a4503eb71a30fb17bb8c2633bf
3
- size 866489
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cd28bc036c9a8abf0e5cc566197ff55c83ed0656f5a87363c72c3baf8cee8256
3
+ size 6604509
paper/AFFECTOR-ACM.pdf CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:5f080cff35f877bb791008a3b455a5d7544fc57378feacf64da5b7f4402fafac
3
- size 195948
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c683afbacffddad98e72525aa0593e8e1695c3bfca526a41a852cfc2baff5e8a
3
+ size 2132249
paper/AFFECTOR-ACM.source.zip CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:64526e21d1658e6e9a40213c53141c257b44e1a4503eb71a30fb17bb8c2633bf
3
- size 866489
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cd28bc036c9a8abf0e5cc566197ff55c83ed0656f5a87363c72c3baf8cee8256
3
+ size 6604509
paper/AFFECTOR-ACM.tex CHANGED
@@ -13,7 +13,7 @@
13
  \providecommand{\tightlist}{\setlength{\itemsep}{0pt}\setlength{\parskip}{0pt}}
14
  \providecommand{\passthrough}[1]{#1}
15
  \providecommand{\pandocbounded}[1]{#1}
16
- \usepackage{graphicx}
17
  \begin{document}
18
  \title{AFFECTOR: Coordinate-Guided Text Rewriting through Transformer
19
  Activation Interventions}
@@ -114,6 +114,44 @@ the rewrite style, and complete accounting of repetition, truncation,
114
  and original-text fallbacks. This is an evaluation agenda, not a
115
  completed benchmark. \citep{yang2026steerscope}.
116
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
117
  \section{Problem formulation}\label{problem-formulation}
118
 
119
  A document conveys more than propositions: lexical choice, focus,
@@ -228,13 +266,7 @@ genre, topic, lexical content, and emotional categories. Calibration is
228
  exploratory and English-based; multilingual effectiveness has not been
229
  established.
230
 
231
- \begin{figure}[htbp]
232
- \centering
233
- \includegraphics[width=\textwidth]{figures/illustration-02.pdf}
234
- \caption{Calibration workflow. Small hand-authored contrasts estimate steering directions; normalization neither orthogonalizes them nor establishes emotional validity.}
235
- \Description{Contrastive calibration supplies three directions. Calibration workflow. Small hand-authored contrasts estimate steering directions; normalization neither orthogonalizes them nor establishes emotional validity.}
236
- \label{fig:illustration-2}
237
- \end{figure}
238
 
239
  \section{Coordinate-to-activation
240
  computation}\label{coordinate-to-activation-computation}
@@ -276,13 +308,7 @@ state at that site. Large values can destabilize generation. Raising the
276
  budget only matters when the proposed perturbation would otherwise be
277
  clipped; it cannot correct an inaccurate direction.
278
 
279
- \begin{figure}[htbp]
280
- \centering
281
- \includegraphics[width=\textwidth]{figures/illustration-03.pdf}
282
- \caption{Control pipeline. Axis controls operate on three emotional coefficients, not vocabulary logits. Raising budget has no effect while the proposal already lies below the ceiling.}
283
- \Description{Controls shape the activation displacement. Control pipeline. Axis controls operate on three emotional coefficients, not vocabulary logits. Raising budget has no effect while the proposal already lies below the ceiling.}
284
- \label{fig:illustration-3}
285
- \end{figure}
286
 
287
  \section{Intervention within the
288
  transformer}\label{intervention-within-the-transformer}
@@ -317,13 +343,7 @@ no direct valence-axis perturbation; it does not actively neutralize an
317
  already negative passage. Correlations between basis vectors further
318
  complicate interpretation.
319
 
320
- \begin{figure}[htbp]
321
- \centering
322
- \includegraphics[width=\textwidth]{figures/illustration-04.pdf}
323
- \caption{Hook implementation schematic. The final prompt token and current cached-decoding token receive the intervention; hooks are removed on exit. Hosted API mode instead uses coordinate-bearing instructions.}
324
- \Description{A change to both feed-forward branches. Hook implementation schematic. The final prompt token and current cached-decoding token receive the intervention; hooks are removed on exit. Hosted API mode instead uses coordinate-bearing instructions.}
325
- \label{fig:illustration-4}
326
- \end{figure}
327
 
328
  \section{Prompts, sampling, and bounded
329
  output}\label{prompts-sampling-and-bounded-output}
@@ -398,13 +418,7 @@ with incoherence, factual errors, or undesirable content. The
398
  implementation does not optimize this score during generation and does
399
  not include a formal semantic-fidelity metric.
400
 
401
- \begin{figure}[htbp]
402
- \centering
403
- \includegraphics[width=\textwidth]{figures/illustration-05.pdf}
404
- \caption{Verification workflow. Failed or direction-unknown assessments are not assigned invented scores, and this score is not an iterative feedback objective during rewriting.}
405
- \Description{Verification is a separate assessment. Verification workflow. Failed or direction-unknown assessments are not assigned invented scores, and this score is not an iterative feedback objective during rewriting.}
406
- \label{fig:illustration-5}
407
- \end{figure}
408
 
409
  \section{Architecture, persistence, and
410
  privacy}\label{architecture-persistence-and-privacy}
 
13
  \providecommand{\tightlist}{\setlength{\itemsep}{0pt}\setlength{\parskip}{0pt}}
14
  \providecommand{\passthrough}[1]{#1}
15
  \providecommand{\pandocbounded}[1]{#1}
16
+ \usepackage{graphicx,float}
17
  \begin{document}
18
  \title{AFFECTOR: Coordinate-Guided Text Rewriting through Transformer
19
  Activation Interventions}
 
114
  and original-text fallbacks. This is an evaluation agenda, not a
115
  completed benchmark. \citep{yang2026steerscope}.
116
 
117
+ \section{Application interface}
118
+ The application exposes its principal operations through the following workspaces and controls.
119
+ \begin{figure}[H]
120
+ \centering
121
+ \includegraphics[width=\linewidth,height=0.70\textheight,keepaspectratio]{figures/screenshot-01.png}
122
+ \caption{Document workspace with an example input, rewrite controls, and an empty output pane before generation.}
123
+ \Description{Document workspace with an example input, rewrite controls, and an empty output pane before generation.}
124
+ \label{fig:app-screen-1}
125
+ \end{figure}
126
+ \begin{figure}[H]
127
+ \centering
128
+ \includegraphics[width=\linewidth,height=0.70\textheight,keepaspectratio]{figures/screenshot-02.png}
129
+ \caption{Geometry editor showing the selected three-point trajectory and its steering controls.}
130
+ \Description{Geometry editor showing the selected three-point trajectory and its steering controls.}
131
+ \label{fig:app-screen-2}
132
+ \end{figure}
133
+ \begin{figure}[H]
134
+ \centering
135
+ \includegraphics[width=\linewidth,height=0.70\textheight,keepaspectratio]{figures/screenshot-03.png}
136
+ \caption{Source-analysis dialog for generating a geometry and optionally calibrating a local activation basis.}
137
+ \Description{Source-analysis dialog for generating a geometry and optionally calibrating a local activation basis.}
138
+ \label{fig:app-screen-3}
139
+ \end{figure}
140
+ \begin{figure}[H]
141
+ \centering
142
+ \includegraphics[width=\linewidth,height=0.70\textheight,keepaspectratio]{figures/screenshot-04.png}
143
+ \caption{Expanded numerical weights for the selected geometry, including the local calibration metadata.}
144
+ \Description{Expanded numerical weights for the selected geometry, including the local calibration metadata.}
145
+ \label{fig:app-screen-4}
146
+ \end{figure}
147
+ \begin{figure}[H]
148
+ \centering
149
+ \includegraphics[width=\linewidth,height=0.70\textheight,keepaspectratio]{figures/screenshot-05.png}
150
+ \caption{Verification controls before evaluation. No verification results have been computed in this view.}
151
+ \Description{Verification controls before evaluation. No verification results have been computed in this view.}
152
+ \label{fig:app-screen-5}
153
+ \end{figure}
154
+
155
  \section{Problem formulation}\label{problem-formulation}
156
 
157
  A document conveys more than propositions: lexical choice, focus,
 
266
  exploratory and English-based; multilingual effectiveness has not been
267
  established.
268
 
269
+
 
 
 
 
 
 
270
 
271
  \section{Coordinate-to-activation
272
  computation}\label{coordinate-to-activation-computation}
 
308
  budget only matters when the proposed perturbation would otherwise be
309
  clipped; it cannot correct an inaccurate direction.
310
 
311
+
 
 
 
 
 
 
312
 
313
  \section{Intervention within the
314
  transformer}\label{intervention-within-the-transformer}
 
343
  already negative passage. Correlations between basis vectors further
344
  complicate interpretation.
345
 
346
+
 
 
 
 
 
 
347
 
348
  \section{Prompts, sampling, and bounded
349
  output}\label{prompts-sampling-and-bounded-output}
 
418
  implementation does not optimize this score during generation and does
419
  not include a formal semantic-fidelity metric.
420
 
421
+
 
 
 
 
 
 
422
 
423
  \section{Architecture, persistence, and
424
  privacy}\label{architecture-persistence-and-privacy}
paper/AFFECTOR-Technical-Report-ACM-source.zip CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:64526e21d1658e6e9a40213c53141c257b44e1a4503eb71a30fb17bb8c2633bf
3
- size 866489
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cd28bc036c9a8abf0e5cc566197ff55c83ed0656f5a87363c72c3baf8cee8256
3
+ size 6604509
paper/AFFECTOR-Technical-Report-ACM.pdf CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:5f080cff35f877bb791008a3b455a5d7544fc57378feacf64da5b7f4402fafac
3
- size 195948
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c683afbacffddad98e72525aa0593e8e1695c3bfca526a41a852cfc2baff5e8a
3
+ size 2132249
paper/AFFECTOR-Technical-Report-ACM.source.zip CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:64526e21d1658e6e9a40213c53141c257b44e1a4503eb71a30fb17bb8c2633bf
3
- size 866489
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cd28bc036c9a8abf0e5cc566197ff55c83ed0656f5a87363c72c3baf8cee8256
3
+ size 6604509
paper/AFFECTOR-Technical-Report-ACM.tex CHANGED
@@ -13,7 +13,7 @@
13
  \providecommand{\tightlist}{\setlength{\itemsep}{0pt}\setlength{\parskip}{0pt}}
14
  \providecommand{\passthrough}[1]{#1}
15
  \providecommand{\pandocbounded}[1]{#1}
16
- \usepackage{graphicx}
17
  \begin{document}
18
  \title{AFFECTOR: Coordinate-Guided Text Rewriting through Transformer
19
  Activation Interventions}
@@ -114,6 +114,44 @@ the rewrite style, and complete accounting of repetition, truncation,
114
  and original-text fallbacks. This is an evaluation agenda, not a
115
  completed benchmark. \citep{yang2026steerscope}.
116
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
117
  \section{Problem formulation}\label{problem-formulation}
118
 
119
  A document conveys more than propositions: lexical choice, focus,
@@ -228,13 +266,7 @@ genre, topic, lexical content, and emotional categories. Calibration is
228
  exploratory and English-based; multilingual effectiveness has not been
229
  established.
230
 
231
- \begin{figure}[htbp]
232
- \centering
233
- \includegraphics[width=\textwidth]{figures/illustration-02.pdf}
234
- \caption{Calibration workflow. Small hand-authored contrasts estimate steering directions; normalization neither orthogonalizes them nor establishes emotional validity.}
235
- \Description{Contrastive calibration supplies three directions. Calibration workflow. Small hand-authored contrasts estimate steering directions; normalization neither orthogonalizes them nor establishes emotional validity.}
236
- \label{fig:illustration-2}
237
- \end{figure}
238
 
239
  \section{Coordinate-to-activation
240
  computation}\label{coordinate-to-activation-computation}
@@ -276,13 +308,7 @@ state at that site. Large values can destabilize generation. Raising the
276
  budget only matters when the proposed perturbation would otherwise be
277
  clipped; it cannot correct an inaccurate direction.
278
 
279
- \begin{figure}[htbp]
280
- \centering
281
- \includegraphics[width=\textwidth]{figures/illustration-03.pdf}
282
- \caption{Control pipeline. Axis controls operate on three emotional coefficients, not vocabulary logits. Raising budget has no effect while the proposal already lies below the ceiling.}
283
- \Description{Controls shape the activation displacement. Control pipeline. Axis controls operate on three emotional coefficients, not vocabulary logits. Raising budget has no effect while the proposal already lies below the ceiling.}
284
- \label{fig:illustration-3}
285
- \end{figure}
286
 
287
  \section{Intervention within the
288
  transformer}\label{intervention-within-the-transformer}
@@ -317,13 +343,7 @@ no direct valence-axis perturbation; it does not actively neutralize an
317
  already negative passage. Correlations between basis vectors further
318
  complicate interpretation.
319
 
320
- \begin{figure}[htbp]
321
- \centering
322
- \includegraphics[width=\textwidth]{figures/illustration-04.pdf}
323
- \caption{Hook implementation schematic. The final prompt token and current cached-decoding token receive the intervention; hooks are removed on exit. Hosted API mode instead uses coordinate-bearing instructions.}
324
- \Description{A change to both feed-forward branches. Hook implementation schematic. The final prompt token and current cached-decoding token receive the intervention; hooks are removed on exit. Hosted API mode instead uses coordinate-bearing instructions.}
325
- \label{fig:illustration-4}
326
- \end{figure}
327
 
328
  \section{Prompts, sampling, and bounded
329
  output}\label{prompts-sampling-and-bounded-output}
@@ -398,13 +418,7 @@ with incoherence, factual errors, or undesirable content. The
398
  implementation does not optimize this score during generation and does
399
  not include a formal semantic-fidelity metric.
400
 
401
- \begin{figure}[htbp]
402
- \centering
403
- \includegraphics[width=\textwidth]{figures/illustration-05.pdf}
404
- \caption{Verification workflow. Failed or direction-unknown assessments are not assigned invented scores, and this score is not an iterative feedback objective during rewriting.}
405
- \Description{Verification is a separate assessment. Verification workflow. Failed or direction-unknown assessments are not assigned invented scores, and this score is not an iterative feedback objective during rewriting.}
406
- \label{fig:illustration-5}
407
- \end{figure}
408
 
409
  \section{Architecture, persistence, and
410
  privacy}\label{architecture-persistence-and-privacy}
 
13
  \providecommand{\tightlist}{\setlength{\itemsep}{0pt}\setlength{\parskip}{0pt}}
14
  \providecommand{\passthrough}[1]{#1}
15
  \providecommand{\pandocbounded}[1]{#1}
16
+ \usepackage{graphicx,float}
17
  \begin{document}
18
  \title{AFFECTOR: Coordinate-Guided Text Rewriting through Transformer
19
  Activation Interventions}
 
114
  and original-text fallbacks. This is an evaluation agenda, not a
115
  completed benchmark. \citep{yang2026steerscope}.
116
 
117
+ \section{Application interface}
118
+ The application exposes its principal operations through the following workspaces and controls.
119
+ \begin{figure}[H]
120
+ \centering
121
+ \includegraphics[width=\linewidth,height=0.70\textheight,keepaspectratio]{figures/screenshot-01.png}
122
+ \caption{Document workspace with an example input, rewrite controls, and an empty output pane before generation.}
123
+ \Description{Document workspace with an example input, rewrite controls, and an empty output pane before generation.}
124
+ \label{fig:app-screen-1}
125
+ \end{figure}
126
+ \begin{figure}[H]
127
+ \centering
128
+ \includegraphics[width=\linewidth,height=0.70\textheight,keepaspectratio]{figures/screenshot-02.png}
129
+ \caption{Geometry editor showing the selected three-point trajectory and its steering controls.}
130
+ \Description{Geometry editor showing the selected three-point trajectory and its steering controls.}
131
+ \label{fig:app-screen-2}
132
+ \end{figure}
133
+ \begin{figure}[H]
134
+ \centering
135
+ \includegraphics[width=\linewidth,height=0.70\textheight,keepaspectratio]{figures/screenshot-03.png}
136
+ \caption{Source-analysis dialog for generating a geometry and optionally calibrating a local activation basis.}
137
+ \Description{Source-analysis dialog for generating a geometry and optionally calibrating a local activation basis.}
138
+ \label{fig:app-screen-3}
139
+ \end{figure}
140
+ \begin{figure}[H]
141
+ \centering
142
+ \includegraphics[width=\linewidth,height=0.70\textheight,keepaspectratio]{figures/screenshot-04.png}
143
+ \caption{Expanded numerical weights for the selected geometry, including the local calibration metadata.}
144
+ \Description{Expanded numerical weights for the selected geometry, including the local calibration metadata.}
145
+ \label{fig:app-screen-4}
146
+ \end{figure}
147
+ \begin{figure}[H]
148
+ \centering
149
+ \includegraphics[width=\linewidth,height=0.70\textheight,keepaspectratio]{figures/screenshot-05.png}
150
+ \caption{Verification controls before evaluation. No verification results have been computed in this view.}
151
+ \Description{Verification controls before evaluation. No verification results have been computed in this view.}
152
+ \label{fig:app-screen-5}
153
+ \end{figure}
154
+
155
  \section{Problem formulation}\label{problem-formulation}
156
 
157
  A document conveys more than propositions: lexical choice, focus,
 
266
  exploratory and English-based; multilingual effectiveness has not been
267
  established.
268
 
269
+
 
 
 
 
 
 
270
 
271
  \section{Coordinate-to-activation
272
  computation}\label{coordinate-to-activation-computation}
 
308
  budget only matters when the proposed perturbation would otherwise be
309
  clipped; it cannot correct an inaccurate direction.
310
 
311
+
 
 
 
 
 
 
312
 
313
  \section{Intervention within the
314
  transformer}\label{intervention-within-the-transformer}
 
343
  already negative passage. Correlations between basis vectors further
344
  complicate interpretation.
345
 
346
+
 
 
 
 
 
 
347
 
348
  \section{Prompts, sampling, and bounded
349
  output}\label{prompts-sampling-and-bounded-output}
 
418
  implementation does not optimize this score during generation and does
419
  not include a formal semantic-fidelity metric.
420
 
421
+
 
 
 
 
 
 
422
 
423
  \section{Architecture, persistence, and
424
  privacy}\label{architecture-persistence-and-privacy}
paper/ILLUSTRATIONS.json CHANGED
@@ -1,37 +1,62 @@
1
- [
2
- {
3
- "number": 1,
4
- "title": "One continuous path, a finite set of targets",
5
- "caption": "Normalized alignment schematic. Target chunks sample a continuous source path at original character midpoints; a short target can miss narrow changes in a long source trajectory.",
6
- "section_preceded": "Frozen model and contrastive calibration",
7
- "provenance": "Explanatory schematic derived from the report; not a screenshot or new experiment."
8
- },
9
- {
10
- "number": 2,
11
- "title": "Contrastive calibration supplies three directions",
12
- "caption": "Calibration workflow. Small hand-authored contrasts estimate steering directions; normalization neither orthogonalizes them nor establishes emotional validity.",
13
- "section_preceded": "Coordinate-to-activation computation",
14
- "provenance": "Explanatory schematic derived from the report; not a screenshot or new experiment."
15
- },
16
- {
17
- "number": 3,
18
- "title": "Controls shape the activation displacement",
19
- "caption": "Control pipeline. Axis controls operate on three emotional coefficients, not vocabulary logits. Raising budget has no effect while the proposal already lies below the ceiling.",
20
- "section_preceded": "Intervention within the transformer",
21
- "provenance": "Explanatory schematic derived from the report; not a screenshot or new experiment."
22
- },
23
- {
24
- "number": 4,
25
- "title": "A change to both feed-forward branches",
26
- "caption": "Hook implementation schematic. The final prompt token and current cached-decoding token receive the intervention; hooks are removed on exit. Hosted API mode instead uses coordinate-bearing instructions.",
27
- "section_preceded": "Prompts, sampling, and bounded output",
28
- "provenance": "Explanatory schematic derived from the report; not a screenshot or new experiment."
29
- },
30
- {
31
- "number": 5,
32
- "title": "Verification is a separate assessment",
33
- "caption": "Verification workflow. Failed or direction-unknown assessments are not assigned invented scores, and this score is not an iterative feedback objective during rewriting.",
34
- "section_preceded": "Architecture, persistence, and privacy",
35
- "provenance": "Explanatory schematic derived from the report; not a screenshot or new experiment."
36
- }
37
- ]
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "screenshots": [
3
+ {
4
+ "file": "figures/screenshot-01.png",
5
+ "source": "affector-01.png",
6
+ "crop": [
7
+ 10,
8
+ 90,
9
+ 2690,
10
+ 1660
11
+ ],
12
+ "caption": "Document workspace with an example input, rewrite controls, and an empty output pane before generation."
13
+ },
14
+ {
15
+ "file": "figures/screenshot-02.png",
16
+ "source": "affector-02.png",
17
+ "crop": [
18
+ 450,
19
+ 260,
20
+ 2660,
21
+ 1590
22
+ ],
23
+ "caption": "Geometry editor showing the selected three-point trajectory and its steering controls."
24
+ },
25
+ {
26
+ "file": "figures/screenshot-03.png",
27
+ "source": "affector-03.png",
28
+ "crop": [
29
+ 22,
30
+ 24,
31
+ 1498,
32
+ 1138
33
+ ],
34
+ "caption": "Source-analysis dialog for generating a geometry and optionally calibrating a local activation basis."
35
+ },
36
+ {
37
+ "file": "figures/screenshot-04.png",
38
+ "source": "affector-04.png",
39
+ "crop": [
40
+ 450,
41
+ 260,
42
+ 2660,
43
+ 1590
44
+ ],
45
+ "caption": "Expanded numerical weights for the selected geometry, including the local calibration metadata."
46
+ },
47
+ {
48
+ "file": "figures/screenshot-05.png",
49
+ "source": "affector-05.png",
50
+ "crop": [
51
+ 450,
52
+ 270,
53
+ 2660,
54
+ 850
55
+ ],
56
+ "caption": "Verification controls before evaluation. No verification results have been computed in this view."
57
+ }
58
+ ],
59
+ "retained_diagrams": [
60
+ 1
61
+ ]
62
+ }
paper/PAPER-SHA256SUMS.txt CHANGED
@@ -1,17 +1,17 @@
1
- 64526e21d1658e6e9a40213c53141c257b44e1a4503eb71a30fb17bb8c2633bf AFFECTOR-ACM-source.zip
2
  d78577482f74cba4163db18d88af82786c923d72cd09607bde62cb747c645c5d AFFECTOR-ACM.bbl
3
- 5f080cff35f877bb791008a3b455a5d7544fc57378feacf64da5b7f4402fafac AFFECTOR-ACM.pdf
4
- 64526e21d1658e6e9a40213c53141c257b44e1a4503eb71a30fb17bb8c2633bf AFFECTOR-ACM.source.zip
5
- 055d38b133c255a1fa228ce066b0f7df71f3e8656e9a523229a2ae9e3c219649 AFFECTOR-ACM.tex
6
- 64526e21d1658e6e9a40213c53141c257b44e1a4503eb71a30fb17bb8c2633bf AFFECTOR-Technical-Report-ACM-source.zip
7
  d78577482f74cba4163db18d88af82786c923d72cd09607bde62cb747c645c5d AFFECTOR-Technical-Report-ACM.bbl
8
- 5f080cff35f877bb791008a3b455a5d7544fc57378feacf64da5b7f4402fafac AFFECTOR-Technical-Report-ACM.pdf
9
- 64526e21d1658e6e9a40213c53141c257b44e1a4503eb71a30fb17bb8c2633bf AFFECTOR-Technical-Report-ACM.source.zip
10
- 055d38b133c255a1fa228ce066b0f7df71f3e8656e9a523229a2ae9e3c219649 AFFECTOR-Technical-Report-ACM.tex
11
- e6b2505c263c35b5d7dd545726bf4bf2de4e109b7c05abb16f9f1f6bdb3a2972 ILLUSTRATIONS.json
12
  d14edda139a36c0bac8b56d02e5cda63ab7558c6db1424f1087189a9429bfaf0 LITERATURE_SOURCES.md
13
  22f8f7bf6992778eb1190ee02b6e83dacbd54ff34af7777a18b4e27a195309a3 README.md
14
- e2148c0d8e58cdc4024151b19ac08625e62355585b9e46011fa68eee9cd48188 TECHNICAL_PAPER.md
15
  4bd1a1b7f6af31b12b9c92bcd978fe116482892b3463f8ec9c922bc1f636c7a0 figure-specifications.json
16
  b11798a28d94d60110a7b19f58107d5e60926b68265f64e62f59e9230a8f500d figures/illustration-01.pdf
17
  d2ae42a17ff3f16b89a7cb85d21ac1f3c7a19c2bc62d61f1cc581c2ca6d3e596 figures/illustration-01.png
@@ -25,4 +25,4 @@ a6a497a38f4ba9de8088f410535bee13dd862d1a97f4175d205df768e1d2c1c9 figures/illust
25
  7d8f30b829993326e8d764ae18b3974acddbd86b760d4071bdff4edfcb9b299a figures/illustration-05.png
26
  9257f38423bd7e496f1a0916c0fbaefbfb32ab544a684ca9a411508201f0535e references.bib
27
  eed8b979a5a299b42e925d9027332e7df876400dcef046f2cfe26beef91d1189 render_figures.py
28
- e2148c0d8e58cdc4024151b19ac08625e62355585b9e46011fa68eee9cd48188 source.md
 
1
+ cd28bc036c9a8abf0e5cc566197ff55c83ed0656f5a87363c72c3baf8cee8256 AFFECTOR-ACM-source.zip
2
  d78577482f74cba4163db18d88af82786c923d72cd09607bde62cb747c645c5d AFFECTOR-ACM.bbl
3
+ c683afbacffddad98e72525aa0593e8e1695c3bfca526a41a852cfc2baff5e8a AFFECTOR-ACM.pdf
4
+ cd28bc036c9a8abf0e5cc566197ff55c83ed0656f5a87363c72c3baf8cee8256 AFFECTOR-ACM.source.zip
5
+ 7ec8d6a693c2de155d62d434bef4c387f911c75f0dceadbd1de9213b9c7691e2 AFFECTOR-ACM.tex
6
+ cd28bc036c9a8abf0e5cc566197ff55c83ed0656f5a87363c72c3baf8cee8256 AFFECTOR-Technical-Report-ACM-source.zip
7
  d78577482f74cba4163db18d88af82786c923d72cd09607bde62cb747c645c5d AFFECTOR-Technical-Report-ACM.bbl
8
+ c683afbacffddad98e72525aa0593e8e1695c3bfca526a41a852cfc2baff5e8a AFFECTOR-Technical-Report-ACM.pdf
9
+ cd28bc036c9a8abf0e5cc566197ff55c83ed0656f5a87363c72c3baf8cee8256 AFFECTOR-Technical-Report-ACM.source.zip
10
+ 7ec8d6a693c2de155d62d434bef4c387f911c75f0dceadbd1de9213b9c7691e2 AFFECTOR-Technical-Report-ACM.tex
11
+ e2c0590754823491af4657a764ab6682e27c5eaf0173fc9f5f03a62e84ccb308 ILLUSTRATIONS.json
12
  d14edda139a36c0bac8b56d02e5cda63ab7558c6db1424f1087189a9429bfaf0 LITERATURE_SOURCES.md
13
  22f8f7bf6992778eb1190ee02b6e83dacbd54ff34af7777a18b4e27a195309a3 README.md
14
+ 4b62ddd1788644df9f817c2b10fb649239e5263e22fbc381af3b6f8b411e504f TECHNICAL_PAPER.md
15
  4bd1a1b7f6af31b12b9c92bcd978fe116482892b3463f8ec9c922bc1f636c7a0 figure-specifications.json
16
  b11798a28d94d60110a7b19f58107d5e60926b68265f64e62f59e9230a8f500d figures/illustration-01.pdf
17
  d2ae42a17ff3f16b89a7cb85d21ac1f3c7a19c2bc62d61f1cc581c2ca6d3e596 figures/illustration-01.png
 
25
  7d8f30b829993326e8d764ae18b3974acddbd86b760d4071bdff4edfcb9b299a figures/illustration-05.png
26
  9257f38423bd7e496f1a0916c0fbaefbfb32ab544a684ca9a411508201f0535e references.bib
27
  eed8b979a5a299b42e925d9027332e7df876400dcef046f2cfe26beef91d1189 render_figures.py
28
+ 4b62ddd1788644df9f817c2b10fb649239e5263e22fbc381af3b6f8b411e504f source.md
paper/SCREENSHOTS.json ADDED
@@ -0,0 +1,57 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "file": "figures/screenshot-01.png",
4
+ "source": "affector-01.png",
5
+ "crop": [
6
+ 10,
7
+ 90,
8
+ 2690,
9
+ 1660
10
+ ],
11
+ "caption": "Document workspace with an example input, rewrite controls, and an empty output pane before generation."
12
+ },
13
+ {
14
+ "file": "figures/screenshot-02.png",
15
+ "source": "affector-02.png",
16
+ "crop": [
17
+ 450,
18
+ 260,
19
+ 2660,
20
+ 1590
21
+ ],
22
+ "caption": "Geometry editor showing the selected three-point trajectory and its steering controls."
23
+ },
24
+ {
25
+ "file": "figures/screenshot-03.png",
26
+ "source": "affector-03.png",
27
+ "crop": [
28
+ 22,
29
+ 24,
30
+ 1498,
31
+ 1138
32
+ ],
33
+ "caption": "Source-analysis dialog for generating a geometry and optionally calibrating a local activation basis."
34
+ },
35
+ {
36
+ "file": "figures/screenshot-04.png",
37
+ "source": "affector-04.png",
38
+ "crop": [
39
+ 450,
40
+ 260,
41
+ 2660,
42
+ 1590
43
+ ],
44
+ "caption": "Expanded numerical weights for the selected geometry, including the local calibration metadata."
45
+ },
46
+ {
47
+ "file": "figures/screenshot-05.png",
48
+ "source": "affector-05.png",
49
+ "crop": [
50
+ 450,
51
+ 270,
52
+ 2660,
53
+ 850
54
+ ],
55
+ "caption": "Verification controls before evaluation. No verification results have been computed in this view."
56
+ }
57
+ ]
paper/TECHNICAL_PAPER.md CHANGED
@@ -11,6 +11,28 @@ November 13, 2026 · Implementation report · Desktop version 0.1.0
11
 
12
  AFFECTOR investigates whether a prescribed emotional trajectory can influence the rewriting of a document. A trajectory assigns valence, emotional direction, and intensity to normalized positions in a text. A local backend maps these coordinates into perturbations of a frozen language model's internal states, while hosted backends use coordinate-bearing instructions. The desktop application supports geometry authoring, editable numerical steering matrices, document conversion, checkpointing, and separate emotional verification. This paper derives the implemented method from first principles and distinguishes mechanical correctness from successful emotional transfer. Current controls can produce emotionally charged language but can also produce weak paraphrases, factual distortions, and repetition. The implementation is an experimental instrument, not a validated predictor or controller of human emotion.
13
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
14
  ## Literature review: activation steering and emotion
15
 
16
  This focused review covers twelve papers, emphasizing 2025-2026 research and retaining two methodological foundations. Titles, author lists, version histories, and the claims summarized here were checked against arXiv records and the corresponding paper text on November 13, 2026. The cited versions are fixed below; availability on arXiv alone does not establish peer review. The implications for this project are research motivations, not evidence that its implementation reproduces these results.
@@ -71,9 +93,6 @@ The resulting `B[l]` is a 3 by 1,024 matrix. These are the editable steering-mod
71
 
72
  There is no gradient training, score-supervised fitting, orthogonalization, or guarantee that the three directions isolate their intended concepts. Small hand-authored pairs can confound affect with genre, topic, lexical content, and emotional categories. Calibration is exploratory and English-based; multilingual effectiveness has not been established.
73
 
74
- ![Contrastive calibration supplies three directions](figures/illustration-02.png)
75
-
76
- *Illustration 2. Calibration workflow. Small hand-authored contrasts estimate steering directions; normalization neither orthogonalizes them nor establishes emotional validity.*
77
 
78
  ## 4. Coordinate-to-activation computation
79
 
@@ -99,9 +118,6 @@ Zero strength, budget, or layer weight gives an exact zero perturbation. Each ch
99
 
100
  Budget is not a semantic-change percentage. At budget 1, the perturbation can have RMS magnitude comparable to the original hidden state at that site. Large values can destabilize generation. Raising the budget only matters when the proposed perturbation would otherwise be clipped; it cannot correct an inaccurate direction.
101
 
102
- ![Controls shape the activation displacement](figures/illustration-03.png)
103
-
104
- *Illustration 3. Control pipeline. Axis controls operate on three emotional coefficients, not vocabulary logits. Raising budget has no effect while the proposal already lies below the ceiling.*
105
 
106
  ## 5. Intervention within the transformer
107
 
@@ -119,9 +135,6 @@ The module does not search for trigger phrases or recognize an idiom before acti
119
 
120
  A key limitation is that the intervention is an absolute-coordinate bias, not a measured error correction. A target valence of zero yields no direct valence-axis perturbation; it does not actively neutralize an already negative passage. Correlations between basis vectors further complicate interpretation.
121
 
122
- ![A change to both feed-forward branches](figures/illustration-04.png)
123
-
124
- *Illustration 4. Hook implementation schematic. The final prompt token and current cached-decoding token receive the intervention; hooks are removed on exit. Hosted API mode instead uses coordinate-bearing instructions.*
125
 
126
  ## 6. Prompts, sampling, and bounded output
127
 
@@ -149,9 +162,6 @@ The denominator is the maximum distance across the three-dimensional `[-1,1]` cu
149
 
150
  The same small model can assess its own output, but shared weaknesses make independent evaluation preferable. A high similarity can coexist with incoherence, factual errors, or undesirable content. The implementation does not optimize this score during generation and does not include a formal semantic-fidelity metric.
151
 
152
- ![Verification is a separate assessment](figures/illustration-05.png)
153
-
154
- *Illustration 5. Verification workflow. Failed or direction-unknown assessments are not assigned invented scores, and this score is not an iterative feedback objective during rewriting.*
155
 
156
  ## 8. Architecture, persistence, and privacy
157
 
 
11
 
12
  AFFECTOR investigates whether a prescribed emotional trajectory can influence the rewriting of a document. A trajectory assigns valence, emotional direction, and intensity to normalized positions in a text. A local backend maps these coordinates into perturbations of a frozen language model's internal states, while hosted backends use coordinate-bearing instructions. The desktop application supports geometry authoring, editable numerical steering matrices, document conversion, checkpointing, and separate emotional verification. This paper derives the implemented method from first principles and distinguishes mechanical correctness from successful emotional transfer. Current controls can produce emotionally charged language but can also produce weak paraphrases, factual distortions, and repetition. The implementation is an experimental instrument, not a validated predictor or controller of human emotion.
13
 
14
+ ## Application interface
15
+
16
+ ![Document workspace with an example input, rewrite controls, and an empty output pane before generation.](figures/screenshot-01.png)
17
+
18
+ *Document workspace with an example input, rewrite controls, and an empty output pane before generation.*
19
+
20
+ ![Geometry editor showing the selected three-point trajectory and its steering controls.](figures/screenshot-02.png)
21
+
22
+ *Geometry editor showing the selected three-point trajectory and its steering controls.*
23
+
24
+ ![Source-analysis dialog for generating a geometry and optionally calibrating a local activation basis.](figures/screenshot-03.png)
25
+
26
+ *Source-analysis dialog for generating a geometry and optionally calibrating a local activation basis.*
27
+
28
+ ![Expanded numerical weights for the selected geometry, including the local calibration metadata.](figures/screenshot-04.png)
29
+
30
+ *Expanded numerical weights for the selected geometry, including the local calibration metadata.*
31
+
32
+ ![Verification controls before evaluation. No verification results have been computed in this view.](figures/screenshot-05.png)
33
+
34
+ *Verification controls before evaluation. No verification results have been computed in this view.*
35
+
36
  ## Literature review: activation steering and emotion
37
 
38
  This focused review covers twelve papers, emphasizing 2025-2026 research and retaining two methodological foundations. Titles, author lists, version histories, and the claims summarized here were checked against arXiv records and the corresponding paper text on November 13, 2026. The cited versions are fixed below; availability on arXiv alone does not establish peer review. The implications for this project are research motivations, not evidence that its implementation reproduces these results.
 
93
 
94
  There is no gradient training, score-supervised fitting, orthogonalization, or guarantee that the three directions isolate their intended concepts. Small hand-authored pairs can confound affect with genre, topic, lexical content, and emotional categories. Calibration is exploratory and English-based; multilingual effectiveness has not been established.
95
 
 
 
 
96
 
97
  ## 4. Coordinate-to-activation computation
98
 
 
118
 
119
  Budget is not a semantic-change percentage. At budget 1, the perturbation can have RMS magnitude comparable to the original hidden state at that site. Large values can destabilize generation. Raising the budget only matters when the proposed perturbation would otherwise be clipped; it cannot correct an inaccurate direction.
120
 
 
 
 
121
 
122
  ## 5. Intervention within the transformer
123
 
 
135
 
136
  A key limitation is that the intervention is an absolute-coordinate bias, not a measured error correction. A target valence of zero yields no direct valence-axis perturbation; it does not actively neutralize an already negative passage. Correlations between basis vectors further complicate interpretation.
137
 
 
 
 
138
 
139
  ## 6. Prompts, sampling, and bounded output
140
 
 
162
 
163
  The same small model can assess its own output, but shared weaknesses make independent evaluation preferable. A high similarity can coexist with incoherence, factual errors, or undesirable content. The implementation does not optimize this score during generation and does not include a formal semantic-fidelity metric.
164
 
 
 
 
165
 
166
  ## 8. Architecture, persistence, and privacy
167
 
paper/figures/screenshot-01.png ADDED

Git LFS Details

  • SHA256: ff1ace53310f4f8ea57dedb7a202edd7404886091ed34b1bcc70cbb6c93b2779
  • Pointer size: 131 Bytes
  • Size of remote file: 636 kB
paper/figures/screenshot-02.png ADDED

Git LFS Details

  • SHA256: 8d267fd6d50e2146279f1f33e07f88a3cbf76d023dd9ee36b29e390ac0473cd1
  • Pointer size: 131 Bytes
  • Size of remote file: 493 kB
paper/figures/screenshot-03.png ADDED

Git LFS Details

  • SHA256: 759dabd81a2ce67f01a3ec96a01b014468402baee5774792fccd5b7fe53af80a
  • Pointer size: 131 Bytes
  • Size of remote file: 266 kB
paper/figures/screenshot-04.png ADDED

Git LFS Details

  • SHA256: 182c73f6f6a527ce52a3a3bec0470af684ac8c67dc787ccacf44ef98a2ddc353
  • Pointer size: 131 Bytes
  • Size of remote file: 567 kB
paper/figures/screenshot-05.png ADDED

Git LFS Details

  • SHA256: 0e951a4cd9f456d8d779a0099427ec90eb2d5e165fc238d83c32be4729ff5115
  • Pointer size: 131 Bytes
  • Size of remote file: 189 kB
paper/source.md CHANGED
@@ -11,6 +11,28 @@ November 13, 2026 · Implementation report · Desktop version 0.1.0
11
 
12
  AFFECTOR investigates whether a prescribed emotional trajectory can influence the rewriting of a document. A trajectory assigns valence, emotional direction, and intensity to normalized positions in a text. A local backend maps these coordinates into perturbations of a frozen language model's internal states, while hosted backends use coordinate-bearing instructions. The desktop application supports geometry authoring, editable numerical steering matrices, document conversion, checkpointing, and separate emotional verification. This paper derives the implemented method from first principles and distinguishes mechanical correctness from successful emotional transfer. Current controls can produce emotionally charged language but can also produce weak paraphrases, factual distortions, and repetition. The implementation is an experimental instrument, not a validated predictor or controller of human emotion.
13
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
14
  ## Literature review: activation steering and emotion
15
 
16
  This focused review covers twelve papers, emphasizing 2025-2026 research and retaining two methodological foundations. Titles, author lists, version histories, and the claims summarized here were checked against arXiv records and the corresponding paper text on November 13, 2026. The cited versions are fixed below; availability on arXiv alone does not establish peer review. The implications for this project are research motivations, not evidence that its implementation reproduces these results.
@@ -71,9 +93,6 @@ The resulting `B[l]` is a 3 by 1,024 matrix. These are the editable steering-mod
71
 
72
  There is no gradient training, score-supervised fitting, orthogonalization, or guarantee that the three directions isolate their intended concepts. Small hand-authored pairs can confound affect with genre, topic, lexical content, and emotional categories. Calibration is exploratory and English-based; multilingual effectiveness has not been established.
73
 
74
- ![Contrastive calibration supplies three directions](figures/illustration-02.png)
75
-
76
- *Illustration 2. Calibration workflow. Small hand-authored contrasts estimate steering directions; normalization neither orthogonalizes them nor establishes emotional validity.*
77
 
78
  ## 4. Coordinate-to-activation computation
79
 
@@ -99,9 +118,6 @@ Zero strength, budget, or layer weight gives an exact zero perturbation. Each ch
99
 
100
  Budget is not a semantic-change percentage. At budget 1, the perturbation can have RMS magnitude comparable to the original hidden state at that site. Large values can destabilize generation. Raising the budget only matters when the proposed perturbation would otherwise be clipped; it cannot correct an inaccurate direction.
101
 
102
- ![Controls shape the activation displacement](figures/illustration-03.png)
103
-
104
- *Illustration 3. Control pipeline. Axis controls operate on three emotional coefficients, not vocabulary logits. Raising budget has no effect while the proposal already lies below the ceiling.*
105
 
106
  ## 5. Intervention within the transformer
107
 
@@ -119,9 +135,6 @@ The module does not search for trigger phrases or recognize an idiom before acti
119
 
120
  A key limitation is that the intervention is an absolute-coordinate bias, not a measured error correction. A target valence of zero yields no direct valence-axis perturbation; it does not actively neutralize an already negative passage. Correlations between basis vectors further complicate interpretation.
121
 
122
- ![A change to both feed-forward branches](figures/illustration-04.png)
123
-
124
- *Illustration 4. Hook implementation schematic. The final prompt token and current cached-decoding token receive the intervention; hooks are removed on exit. Hosted API mode instead uses coordinate-bearing instructions.*
125
 
126
  ## 6. Prompts, sampling, and bounded output
127
 
@@ -149,9 +162,6 @@ The denominator is the maximum distance across the three-dimensional `[-1,1]` cu
149
 
150
  The same small model can assess its own output, but shared weaknesses make independent evaluation preferable. A high similarity can coexist with incoherence, factual errors, or undesirable content. The implementation does not optimize this score during generation and does not include a formal semantic-fidelity metric.
151
 
152
- ![Verification is a separate assessment](figures/illustration-05.png)
153
-
154
- *Illustration 5. Verification workflow. Failed or direction-unknown assessments are not assigned invented scores, and this score is not an iterative feedback objective during rewriting.*
155
 
156
  ## 8. Architecture, persistence, and privacy
157
 
 
11
 
12
  AFFECTOR investigates whether a prescribed emotional trajectory can influence the rewriting of a document. A trajectory assigns valence, emotional direction, and intensity to normalized positions in a text. A local backend maps these coordinates into perturbations of a frozen language model's internal states, while hosted backends use coordinate-bearing instructions. The desktop application supports geometry authoring, editable numerical steering matrices, document conversion, checkpointing, and separate emotional verification. This paper derives the implemented method from first principles and distinguishes mechanical correctness from successful emotional transfer. Current controls can produce emotionally charged language but can also produce weak paraphrases, factual distortions, and repetition. The implementation is an experimental instrument, not a validated predictor or controller of human emotion.
13
 
14
+ ## Application interface
15
+
16
+ ![Document workspace with an example input, rewrite controls, and an empty output pane before generation.](figures/screenshot-01.png)
17
+
18
+ *Document workspace with an example input, rewrite controls, and an empty output pane before generation.*
19
+
20
+ ![Geometry editor showing the selected three-point trajectory and its steering controls.](figures/screenshot-02.png)
21
+
22
+ *Geometry editor showing the selected three-point trajectory and its steering controls.*
23
+
24
+ ![Source-analysis dialog for generating a geometry and optionally calibrating a local activation basis.](figures/screenshot-03.png)
25
+
26
+ *Source-analysis dialog for generating a geometry and optionally calibrating a local activation basis.*
27
+
28
+ ![Expanded numerical weights for the selected geometry, including the local calibration metadata.](figures/screenshot-04.png)
29
+
30
+ *Expanded numerical weights for the selected geometry, including the local calibration metadata.*
31
+
32
+ ![Verification controls before evaluation. No verification results have been computed in this view.](figures/screenshot-05.png)
33
+
34
+ *Verification controls before evaluation. No verification results have been computed in this view.*
35
+
36
  ## Literature review: activation steering and emotion
37
 
38
  This focused review covers twelve papers, emphasizing 2025-2026 research and retaining two methodological foundations. Titles, author lists, version histories, and the claims summarized here were checked against arXiv records and the corresponding paper text on November 13, 2026. The cited versions are fixed below; availability on arXiv alone does not establish peer review. The implications for this project are research motivations, not evidence that its implementation reproduces these results.
 
93
 
94
  There is no gradient training, score-supervised fitting, orthogonalization, or guarantee that the three directions isolate their intended concepts. Small hand-authored pairs can confound affect with genre, topic, lexical content, and emotional categories. Calibration is exploratory and English-based; multilingual effectiveness has not been established.
95
 
 
 
 
96
 
97
  ## 4. Coordinate-to-activation computation
98
 
 
118
 
119
  Budget is not a semantic-change percentage. At budget 1, the perturbation can have RMS magnitude comparable to the original hidden state at that site. Large values can destabilize generation. Raising the budget only matters when the proposed perturbation would otherwise be clipped; it cannot correct an inaccurate direction.
120
 
 
 
 
121
 
122
  ## 5. Intervention within the transformer
123
 
 
135
 
136
  A key limitation is that the intervention is an absolute-coordinate bias, not a measured error correction. A target valence of zero yields no direct valence-axis perturbation; it does not actively neutralize an already negative passage. Correlations between basis vectors further complicate interpretation.
137
 
 
 
 
138
 
139
  ## 6. Prompts, sampling, and bounded output
140
 
 
162
 
163
  The same small model can assess its own output, but shared weaknesses make independent evaluation preferable. A high similarity can coexist with incoherence, factual errors, or undesirable content. The implementation does not optimize this score during generation and does not include a formal semantic-fidelity metric.
164
 
 
 
 
165
 
166
  ## 8. Architecture, persistence, and privacy
167