Persona space interventions to prevent unwanted behaviors
17 log items · updated 5/21/2026
Hidden-behavior reveal via conditional trigger
53 log items · updated 5/22/2026
Leakage as a character-selection signal for story-format SFT
34 log items · updated 5/21/2026
Basic science of conditional behaviors
4 log items · updated 5/21/2026
Basic science of persona space
10 log items · updated 5/21/2026
Drift detection via a non-Assistant persona marker
6 log items · updated 5/21/2026
Persona drift vs persona prompting: same mechanism?
5 log items · updated 5/21/2026
Perceived environment: what context the Assistant thinks it's in
5 log items · updated 5/21/2026
User-selection model: how the AI's model of the user shapes Assistant behavior
6 log items · updated 5/21/2026
EM defense — making misaligned personas computationally limited
7 log items · updated 5/21/2026
EM defense — preventing install into the assistant persona
8 log items · updated 5/21/2026
Installation-path equivalence (prompt vs activation steering vs fine-tuning)
5 log items · updated 5/21/2026
Read/write features for conditional behaviors
7 log items · updated 5/21/2026
Co-installed tells: shape the model so backdoors come with visible markers
6 log items · updated 5/21/2026
Prompt-conditioned installs: leakage characterization and hidden-trigger discovery
7 log items · updated 5/21/2026
Dual-use SDF: can story-format training install hidden misalignment?
8 log items · updated 5/21/2026
What is the assistant persona, exactly?
4 log items · updated 5/21/2026
Midtraining-stage behavioral installation (character-to-assistant transfer)
2 log items · updated 5/13/2026
What exactly can be distilled and through what channels
4 log items · updated 5/20/2026
EM Mechanism
1 log items · updated 5/12/2026
Method to find prompt corresponding to any narrow finetuning/steering vector
1 log items · updated 5/12/2026
Project Log
Leakage as a character-selection signal for story-format SFT
Project updated May 21, 2026
Measure bidirectional leakage between the Assistant and prompted personas to calibrate which characters to put in stories for maximal effect on the Assistant.
Motivation
Story-format SFT — like Anthropic's Teaching Claude Why — delivers behavior to the deployed Assistant by training on stories where AI characters exhibit the target behavior. The bet is that behavior trained into a story-character persona leaks back to the Assistant. TCW assumes the right character is "AI assistant"-shaped; the Evans-group line shows the leakage is gated by how persona-similar the trained character is to the Assistant. Neither pins down which characters maximize the transfer in practice.
Impact. A ranked list of high-leakage personas becomes a planning tool for any project that needs to deliver behavior to the Assistant via narrative training — alignment patches, containment-SDF, adversarial-SDF tests. A clean negative result (bidirectional leakage doesn't predict transfer) is itself informative, since it constrains how persona-similarity should be operationalized.
Question. Can bidirectional leakage between the Assistant and a candidate persona be used as a calibration signal for character selection in story-format SFT?
Approach
For each candidate persona X, measure leakage in both directions with a clean marker behavior (e.g., emitting [ZLT]):
- Marker-from-Assistant → leakage-to-X. SFT the marker into the Assistant; measure marker emission when prompted into X.
- Marker-from-X → leakage-to-Assistant. Symmetric: SFT the marker into X; measure emission on the default Assistant.
- Bidirectional score. Combine into a single per-persona score (geometric mean or similar).
Rank personas. Validate by running a real story-format SFT with a substantive behavior (sycophancy, refusal-on-edge-cases) on top-3 and bottom-3 personas, and check whether observed transfer matches the predicted ranking.
Concrete plan
- Reuse the existing EPS install-and-measure pipeline (marker SFT + cross-persona evaluation).
- Pick ~20 candidate personas spanning AI-assistant variants, human helper professions, fictional AI characters, historical figures, hostile/manipulative archetypes.
- Run the bidirectional measurement; produce the ranked list.
- Validate on top-3 and bottom-3 with a real target behavior.
Related
- Fellowship Step 0 (training-time vs deployment-time persona representation overlap) — informs whether leakage under prompted personas predicts leakage under SFT-trained personas.
- Fellowship Step 2 (containment via misaligned-incompetent SDF) and dual-use SDF — both need character selection; this project provides the calibration.
- Persona Vectors (Chen et al. 2025) and The Assistant Axis (Lu et al. 2026) are the closest prior art on persona geometry; neither validates against a downstream behavioral signal.
Risks / open
- Leakage measured under prompting may not predict transfer under SFT installation. The bidirectional measurement bets the two correlate.
- Most candidate personas may be far enough from the Assistant to produce near-zero leakage, leaving little signal to rank.
- Story-format SFT transfer may depend on more than persona-similarity (story content, narrative voice, training duration).
- The marker behavior may not generalize to substantive target behaviors. The validation step is what closes this gap.
Summary Log
May 22, 2026
5/22/2026
note · 2026-05-22
note**Action:** Updated project narrative 0f0979b5 **Why:** A user edited the running project summary draft. **Detail:** Fields: bodyMd
May 22, 2026
5/22/2026
Leakage as a character-selection signal for story-format SFT — 2026-05-21
publishedContents Motivation Methodology TL;DR Story format SFT teaches a behavior to the deployed Assistant by fine tuning on stories where an AI character does that behavior. Which character you put in the stories changes how m...
May 22, 2026
5/22/2026
note · 2026-05-22
note**Action:** Updated project narrative 0f0979b5 **Why:** A user edited the running project summary draft. **Detail:** Fields: bodyMd
May 22, 2026
5/22/2026
note · 2026-05-22
note**Action:** Updated project narrative 0f0979b5 **Why:** A user edited the running project summary draft. **Detail:** Fields: bodyMd
May 22, 2026
5/22/2026
note · 2026-05-22
note**Action:** Updated project narrative 0f0979b5 **Why:** A user edited the running project summary draft. **Detail:** Fields: bodyMd
May 21, 2026
5/21/2026
note · 2026-05-21
note**Action:** Updated project narrative 0f0979b5 **Why:** A user edited the running project summary draft. **Detail:** Fields: bodyMd
May 21, 2026
5/21/2026
note · 2026-05-21
note**Action:** Resolved comment thread 736e9f57 **Why:** Mark the discussion as handled and preserve the resolution outcome for later review. **Detail:** I don't see a discussion to summarize in your message. Could you plea...
May 21, 2026
5/21/2026
note · 2026-05-21
note**Action:** Resolved comment thread 8cc714d1 **Why:** Mark the discussion as handled and preserve the resolution outcome for later review. **Detail:** I'd be happy to help, but I don't see a discussion to summarize in yo...
May 21, 2026
5/21/2026
note · 2026-05-21
note**Action:** Resolved comment thread 354418fc **Why:** Mark the discussion as handled and preserve the resolution outcome for later review. **Detail:** User requested renaming the project to better reflect recent changes ...
May 21, 2026
5/21/2026
note · 2026-05-21
note**Action:** Resolved comment thread 920d166a **Why:** Mark the discussion as handled and preserve the resolution outcome for later review. **Detail:** I don't have access to the discussion text you're referring to. Could...
May 21, 2026
5/21/2026
note · 2026-05-21
note**Action:** Resolved comment thread 6448a3e2 **Why:** Mark the discussion as handled and preserve the resolution outcome for later review. **Detail:** Characters should be distributed across a full range, not limited to ...
May 21, 2026
5/21/2026
note · 2026-05-21
note**Action:** Resolved comment thread 70dae45f **Why:** Mark the discussion as handled and preserve the resolution outcome for later review. **Detail:** User requested removal of EPS references from a document or system.
May 21, 2026
5/21/2026
note · 2026-05-21
note**Action:** Resolved comment thread 2a92526c **Why:** Mark the discussion as handled and preserve the resolution outcome for later review. **Detail:** I don't have context for the discussion you're referring to. Please s...
May 21, 2026
5/21/2026
note · 2026-05-21
note**Action:** Resolved comment thread b58ad9bc **Why:** Mark the discussion as handled and preserve the resolution outcome for later review. **Detail:** The human proposed using a specific term or phrase in text but didn't...
May 21, 2026
5/21/2026
note · 2026-05-21
note**Action:** Resolved comment thread 2e73c239 **Why:** Mark the discussion as handled and preserve the resolution outcome for later review. **Detail:** I don't have the discussion text to summarize. Please provide the res...
May 21, 2026
5/21/2026
note · 2026-05-21
note**Action:** Updated project narrative 0f0979b5 **Why:** A user edited the running project summary draft. **Detail:** Fields: title, bodyMd
May 21, 2026
5/21/2026
note · 2026-05-21
note**Action:** Updated project Leakage as a character-selection signal for story-format SFT **Why:** A user changed project metadata or status. **Detail:** Fields: title
May 21, 2026
5/21/2026
note · 2026-05-21
note**Action:** Resolved comment thread 11b3fa02 **Why:** Mark the discussion as handled and preserve the resolution outcome for later review. **Detail:** I don't see a discussion provided to summarize. Please share the text...
May 21, 2026
5/21/2026
note · 2026-05-21
note**Action:** Resolved comment thread 78c19449 **Why:** Mark the discussion as handled and preserve the resolution outcome for later review. **Detail:** 401 {"type":"error","error":{"type":"authentication_error","message":...
May 21, 2026
5/21/2026
note · 2026-05-21
note**Action:** Resolved comment thread abd4aaad **Why:** Mark the discussion as handled and preserve the resolution outcome for later review. **Detail:** 401 {"type":"error","error":{"type":"authentication_error","message":...
May 21, 2026
5/21/2026
note · 2026-05-21
note**Action:** Resolved comment thread 89a85782 **Why:** Mark the discussion as handled and preserve the resolution outcome for later review. **Detail:** 401 {"type":"error","error":{"type":"authentication_error","message":...
May 21, 2026
5/21/2026
note · 2026-05-21
note**Action:** Resolved comment thread 11b3fa02 **Why:** Mark the discussion as handled and preserve the resolution outcome for later review. **Detail:** 401 {"type":"error","error":{"type":"authentication_error","message":...
May 21, 2026
5/21/2026
note · 2026-05-21
note**Action:** Updated project narrative 0f0979b5 **Why:** A user edited the running project summary draft. **Detail:** Fields: bodyMd
May 21, 2026
5/21/2026
note · 2026-05-21
note**Action:** Published project narrative 0f0979b5 **Why:** Promote the draft running summary into the mentor-visible/current project summary.
May 21, 2026
5/21/2026
note · 2026-05-21
note**Action:** Created project narrative Bidirectional leakage as a character-selection tool for story-format SFT — 2026-05-21 **Why:** A user started a running project summary for mentor/project review. **Detail:** narrati...
May 21, 2026
5/21/2026
note · 2026-05-21
note**Action:** Updated project Bidirectional leakage as a character-selection tool for story-format SFT **Why:** A user changed project metadata or status. **Detail:** Fields: summaryMd
May 21, 2026
5/21/2026
note · 2026-05-21
note**Action:** Updated project Bidirectional leakage as a character-selection tool for story-format SFT **Why:** A user changed project metadata or status. **Detail:** Fields: public
May 21, 2026
5/21/2026
note · 2026-05-21
note**Action:** Updated project Bidirectional leakage as a character-selection tool for story-format SFT **Why:** A user changed project metadata or status. **Detail:** Fields: title, summaryMd
May 21, 2026
5/21/2026
note · 2026-05-21
note**Action:** Updated project A persona-space distance metric grounded by leakage **Why:** A user changed project metadata or status. **Detail:** Fields: summaryMd
May 20, 2026
5/20/2026
note · 2026-05-20
note**Action:** Updated project A persona-space distance metric grounded by leakage **Why:** A user changed project metadata or status. **Detail:** Fields: summaryMd
May 20, 2026
5/20/2026
note · 2026-05-20
note**Action:** Generated literature review for project A persona-space distance metric grounded by leakage **Why:** A new project was created and the auto lit-review job produced a draft narrative for the user to review. **...
May 20, 2026
5/20/2026
note · 2026-05-20
noteLiterature review draft ready for **A persona-space distance metric grounded by leakage** — [open the draft](/e/project_narrative/2c1bfb20-3089-4cf8-a15e-2f2634c7e670) to review.
May 20, 2026
5/20/2026
Literature review (auto)
draftInvalid API key · Fix external API key
May 20, 2026
5/20/2026
note · 2026-05-20
note**Action:** Created project A persona-space distance metric grounded by leakage **Why:** A user created a new research project to organize related work. **Detail:** slug=persona-space-distance