Docs
Living research docs, meeting + mentor updates, activity logs, and ideas. Sources: docs/*.md, docs/mentor_updates/, logs/daily + logs/weekly, docs/ideas/.
Meetings & mentor updates (10)
- Weekly mentor meeting — June 11, 2026Dan 1:1 · Explore Persona Space. Source: "Research Log" Google Slides deckmu__2026-06-11· updated 2026-06-12
- Weekly mentor meeting — June 4, 2026Dan 1:1 (+ Christina) · Explore Persona Space. Source: "Research Log" Googlemu__2026-06-04· updated 2026-06-12
- Meeting notes — Christina, June 11, 2026Raw notes, captured same day. Not yet triaged into tasks — see the todo tomu__2026-06-11-christina· updated 2026-06-11
- Async mentor note — June 1, 2026Dan → research-log channel · Explore Persona Space. Captured June 2.mu__2026-06-02· updated 2026-06-02
- Mentor Notes — 2026-05-29Mentor: Dan Mossingmu__2026-05-29· updated 2026-05-29
- Mentor Notes — 2026-05-26Mentor: Dan Mossingmu__2026-05-26· updated 2026-05-28
- research-log-exploring-persona-space — full channel historyThomas Jiralerspong has joined the channelmu__2026-05-27-research-log-channel-snapshot· updated 2026-05-28
- Weekly update — May 28, 2026Thomas to Dan · Explore Persona Spacemu__2026-05-28· updated 2026-05-28
- Mentor Meeting — 2026-05-22Mentor: Dan Mossingmu__2026-05-22· updated 2026-05-24
- Mentor Update — 2026-05-11A conditional behavior is a behavior the model produces whenever a specific signal — call it the trigger — appears in the input (or in the context preceding it). The trigger can take many forms, and the same structura…mu__2026-05-11· updated 2026-05-14
Activity (41)
- Daily — 2026-07-07 (27 workflow fixes merged; Codex quota outage; 2 refusal wedges; 4 science rounds landed)- Science rounds landed on 4 issues. #928 folded the matched-length answer-span control (headline retitled: conditioning on realized CoT predicts the thinking model's answer state no better than a matched-length slice…daily__2026-07-07· updated 2026-07-08
- Daily — 2026-07-06 (infra wave — ~15 fixes landed;- Infra/workflow-fix wave landed. Nearly all of yesterday's 10 daily-filed fixes completed end-to-end through the full pipeline today: #1075 (verifyplan GPU-basis), #1076 (test regressions on main), #1077 (step9c dirt…daily__2026-07-06· updated 2026-07-07
- Daily — 2026-07-05 (0 promoted; 2 fixes applied, 10 filed for review)- Three context→answer-map results parked at awaitingpromotion: #952 (own-answer advantage is position-uniform) and #811 (taught-fact chain follows answer content tokens) finished their follow-up rounds, and #923 (att…daily__2026-07-05· updated 2026-07-06
- Daily — 2026-07-04 (10 experiments to awaiting_promotion, 66 infra fixes landed, 0 promoted)- Huge autonomous throughput day: 4,143 commits across 113 tasks. 10 experiment lines reached awaitingpromotion (#742, #810, #825, #841, #922, #923, #928, #931, #952, #958) and ~66 tasks reached completed — almost all…daily__2026-07-04· updated 2026-07-05
- Daily — 2026-07-01 (fleet-scale workflow-fix day; 0 promotions)- Heaviest pipeline day yet: 1,988 workflow markers across 61 tasks (#658–#824), ~2,350 commits on main, 0 clean-results promoted (5 tasks newly parked at awaitingpromotion: #778, #779, and the #810 family in flight).daily__2026-07-01· updated 2026-07-04
- Daily — 2026-07-02 (backfill — 3 clean-results parked, 79 tasks filed, 0 promoted)- Heaviest fleet day on record: 143 sessions active, 3,881 workflow markers across 100 tasks, 79 tasks created (#825–#903) — 40 of them workflow-fix filings from 12 parent tasks, most implemented and merged to main th…daily__2026-07-02· updated 2026-07-04
- Daily — 2026-07-03 (48 completions incl. the workflow-fix wave; API-key outage; 56-item problem sweep)- Workflow-fix wave landed: 48 tasks reached completed, dominated by the #930–#959 infra band (plus #955/#956/#957/#959/#960/#961 lifecycle work into the evening). No clean-results were promoted today.daily__2026-07-03· updated 2026-07-04
- Daily — 2026-06-30 (35 tasks created, ~33 merged; big autonomous workflow-fix day)- Huge autonomous workflow-fix day. ~35 tasks created and ~33 merged to main, almost all kind: infra workflow-fixes (#741–#777) running the full planner → critic ensemble → implementer → Claude+Codex code-review → Ste…daily__2026-06-30· updated 2026-07-01
- Daily — 2026-06-29 (huge autonomous infra/workflow day; 38 infra tasks completed, 9 route-2 fixes filed)- Massive autonomous infra/workflow push: ~38 kind: infra tasks reached completed, including the structural disk fix #681 (dedicated GCP data disk + ext4 per-task quotas), #711 (deterministic nightly lesson-consolidat…daily__2026-06-29· updated 2026-06-30
- Daily — 2026-06-28 (workflow-automation push lands;- Workflow-automation push landed answering today's repeated "I barely run /daily, /weekly — automate this": #711 (deterministic consolidatelessons.py cron + daily healthcheck), #713 (fold /weekly's load-bearing check…daily__2026-06-28· updated 2026-06-29
- Daily — 2026-06-27 (#664 vLLM-crash saga;- #664 vLLM EngineCore-crash saga dominated the day: ~8 GPU launches / 4 strategy pivots against a reproducible hang inside vLLM 0.11.0's first llm.generate() on the elicitsecurecode path (across pods, batch sizes, ea…daily__2026-06-27· updated 2026-06-28
- Daily — 2026-06-26 (clean-result paper-mode redesign; GCP capacity + VM-disk firefighting; 3 fixes auto-applied)- Clean-result format → LaTeX paper-mode (#657): Thomas drove a multi-step redesign — drop the standalone methodology doc, fold it into Goal/Methodology/Results, require low-level labeled-scatter data plots, demand se…daily__2026-06-26· updated 2026-06-27
- Daily — 2026-06-25 (autonomous theory-program day;- The autonomous overnight leakage-predictor assumption-testing program drove the whole day (sequenced /issue --auto rollout, Thomas asleep then research-Q&A by day). Four tasks ran in parallel: #664 (Phase 2 fine-tun…daily__2026-06-25· updated 2026-06-26
- Daily — 2026-06-24 (clean-result v4 redesign shipped;- Clean-result spec migrated to v4 (four-flat-H2: Takeaways / Goal / Methodology / Results + Repro/Context footer; methodology doc = mechanical copy of the body's ## Methodology). Shipped via workflow-fix merges + a v…daily__2026-06-24· updated 2026-06-25
- Daily — 2026-06-23 (marker loss-mask default change +- Theory-driven experiment-design session (d419f9fd): read the leakage-theory doc, fanned out ~16 verification subagents over its assumptions, produced docs/leakagetheorytestsuite.md (33KB) covering marker / sycophanc…daily__2026-06-23· updated 2026-06-24
- Daily — 2026-06-18 (5 follow-up rounds parked at awaiting_promotion; analyzer self-check tightened)- Fully autonomous day — no live user chat. Five same-issue follow-up rounds ran end-to-end and parked at awaitingpromotion for user promotion: #641 (matched-recipe EM-resistant-context install, 8-seed re-fold), #657 …daily__2026-06-18· updated 2026-06-23
- Daily — 2026-06-19 (idle day, no session activity)- Idle day. No Claude Code session activity, no commits, and no task events occurred during the PT 06-19 window (07:00Z 06-19 → 06:27Z 06-20). The only transcript stamped in-window is this /daily run itself.daily__2026-06-19· updated 2026-06-23
- Daily — 2026-06-20 (quiet research/synthesis day; no experiments committed)- Quiet day on the project: no commits, no task-state changes in the repo since 2026-06-18 evening. All of today's activity was research-thinking + result-synthesis chat (5 sessions, all in the 22:38–23:27 PT window).daily__2026-06-20· updated 2026-06-23
- Daily — 2026-06-21 (PM dispatched 5 autonomous runs; all parked at awaiting_promotion)- PM session triage + dispatch. Archived #655, restarted the stalled #651 session by hand, and spawned autonomous --auto sessions for #657, #559, #642, #641 — the strongest follow-ups off the 52-deep awaiting-promotio…daily__2026-06-21· updated 2026-06-23
- Daily — 2026-06-22 (quiet day; no project activity, 1 skill fix applied)- Quiet day on the project: no commits, no task events.jsonl mutations, and no genuine interactive Claude Code sessions on 2026-06-22 (PT). Last real activity was 2026-06-18 (#657).daily__2026-06-22· updated 2026-06-23
- Daily — 2026-06-17 (4 follow-up loops re-parked,- Heavy autonomous fleet day, mostly self-healing. Four same-issue follow-up loops completed and re-parked at awaitingpromotion: #642 (onpolicy-matchedlr-rank-isolation — interp + clean-result ensembles both PASS), #5…daily__2026-06-17· updated 2026-06-18
- Daily — 2026-06-16 (5 results promoted; 4 workflow fixes auto-applied)- Heavy autonomous day across ~11 active issues (542, 597, 612, 640, 641, 642, 647, 649, 651, 653, 654). 5 clean-results promoted useful: #528, #556, #608, #613, #614.daily__2026-06-16· updated 2026-06-17
- Daily — 2026-06-15 (21 results promoted; orphan task folder + SKILL.md merge-guard fixed)- Big evening promotion sweep — 21 clean-results promoted, all classified useful. Thomas worked through the awaitingpromotion queue thread by thread: the role-header / marker-localization line (#464, #533, #546, #611)…daily__2026-06-15· updated 2026-06-16
- Daily — 2026-06-13 (3 pipelines + heavy interactive day; 7 auto-fixes applied)- Three autonomous /issue pipelines ran the full clean-result stack: #632 and #634 reached completed (merged to main); #595 (prefix-carrier binding strength, determinate null) reached awaitingpromotion. #640 filed as …daily__2026-06-13· updated 2026-06-14
- Daily — 2026-06-12 (heavy autonomous pipeline day; ~30 workflow fixes auto-applied in-flight, 0 promotions)- Fully autonomous, very high-throughput day — 2384 commits across the fleet. Multiple per-issue sessions ran the whole /issue pipeline concurrently: #613, #621, #622, #627, #628, #633, #483, #545, #522. #633 reached …daily__2026-06-12· updated 2026-06-13
- Daily — 2026-06-11 (19 tasks filed, 15 results parked, 18 follow-up rounds; 2 workflow fixes auto-applied)- Highest-throughput day on record for the pipeline. 19 tasks filed (#595–#613), 15 experiment results reached awaitingpromotion for the first time (#491, #556, #561, #570, #571, #585, #591, #594, #600, #601, #602, #6…daily__2026-06-11· updated 2026-06-12
- Daily — 2026-06-10 (40 tasks filed, ~20 parked for promotion; 16 workflow fixes auto-applied; 3 promoted)- Highest-throughput day yet: 40 new tasks created (#551–#590), 4,850 commits on main. ~20 tasks reached awaitingpromotion during the day (the column now holds 25); 10 are running, 2 interpreting, 2 in same-issue foll…daily__2026-06-10· updated 2026-06-11
- Daily — 2026-06-09 (96 results promoted; 25 workflow fixes auto-applied)- Promotion-backlog clear-out: 96 clean-results promoted to completed (batches of 12 + 53 + 8 + singles; #407 flipped to not-useful after the batch — its body starts "BUGGED experiment" — and #519/#520 judged not-usef…daily__2026-06-09· updated 2026-06-10
- Daily — 2026-06-08 (compute-infra expansion + ~40 in-session workflow fixes; 0 promotions)- Compute-infrastructure expansion (the day's headline). Brought three new compute backends into reach: Mila cluster SSH access (public key added, IT ticket filed), Compute Canada / DRAC (alliancecan, Duo MFA + robot …daily__2026-06-08· updated 2026-06-09
- Daily — 2026-06-07 (heavy autonomous fleet day; 4 fleet-enforcement fixes landed,- Heavy autonomous fleet day. 7 experiments running concurrently overnight (#488, #501, #503, #504, #505, #507, #511), plus #509 advancing to interpreting and #506 ending blocked. ~522 git commits in the PT window. Pe…daily__2026-06-07· updated 2026-06-08
- Daily — 2026-06-06 (5 autonomous experiments in flight; RunPod cap + smoke-coverage gaps dominate; 0 promoted)- Heavy autonomous-/issue day, mostly machine-driven. Five experiments running at day's end (#501, #503, #506, #507, #508), three parked blocked awaiting a Thomas decision (#504, #505, #509). Three tasks reached await…daily__2026-06-06· updated 2026-06-07
- Daily — 2026-06-05 (heavy autonomous run; Claude-access + RunPod-balance walls; 3 results reached awaiting_promotion, 0 promoted)- Big autonomous-experiment day. ~9 /issue sessions ran in parallel (#489, #496, #498, #501, #502, #503, #504, #505, #506). Three reached awaitingpromotion (clean-result drafted + critic PASS): #498 (custom chat-templ…daily__2026-06-05· updated 2026-06-06
- Daily — 2026-06-04 (#490 → awaiting_promotion; #489 blocked at SP-identity gate; 3 workflow fixes shipped)- #490 (dose-matched midpoint coupling) went proposed → awaitingpromotion in a single day: planned, approved, ran a full 7-pair × 3-seed sweep on a 4×H100 pod, then through the analyzer / interpretation-critic / clean…daily__2026-06-04· updated 2026-06-05
- Daily — 2026-06-03 (10 experiments parked at awaiting_promotion, 6 workflow fixes auto-shipped, 0 promoted)- High experiment throughput. Ten experiment tasks reached awaitingpromotion today: #405 (marker leakage vs K source personas), #444 (invented-fact localization, 4th fact regime + 5-way re-judge), #464 (persona-in-rol…daily__2026-06-03· updated 2026-06-04
- Daily — 2026-06-02 (6 experiments parked at awaiting_promotion; 9 workflow fixes auto-landed)- Six experiments reached a clean-result and parked at awaitingpromotion today: #444 (four fact-teaching leakage shapes + a 5-way re-judge), #464 (chat-template role-header marker localization), #466 (hypothesis-led "…daily__2026-06-02· updated 2026-06-03
- Daily — 2026-06-01 — #448 on-policy re-run lands NULL; idle-pod burn recurred 3× (~$1k); 0 promoted- #448 dominated the day (70 commits). The user halted the off-policy version mid-flight ("STOP — do not promote #448, its DV is off-policy") and forced a full re-run with the on-policy design. The 12-cell contrastive…daily__2026-06-01· updated 2026-06-02
- Daily — 2026-05-31 (on-policy marker re-run lands rank- Clean-result spec migrated to v2. #454 moved the body spec to the 2-content-section nested design (Human TL;DR / TL;DR with ### Motivation + ### What I ran + ### Findings→#### per result / Reproducibility; confiden…daily__2026-05-31· updated 2026-06-01
- Weekly — 2026-W22 (2 results promoted;- 2 clean-results promoted to completed: #390 (refusal-style negatives install a persona gate that generalises across 9 of 11 OOD eval framings, MODERATE) and #391 (could not selectively implant sycophancy on Qwen2.5-…weekly__2026-W22· updated 2026-05-31
- Daily — 2026-05-29 (#404 +- Two clean-result drafts landed at awaitingpromotion (pending your promotion): #404 — base-model narrow↔broad cosine predicts post-SFT broad-misalignment rate (ρ=+0.75, MODERATE); #385 — [ZLT] marker leakage emerges …daily__2026-05-29· updated 2026-05-30
- Daily — 2026-05-28 (#390 +- #390 promoted → completed (MODERATE). Clean-result: refusal-style negatives install a persona gate that generalises across 9 of 11 OOD eval framings as a verbatim one-line refusal substitution. Two user-directed bod…daily__2026-05-28· updated 2026-05-28
- Daily — 2026-05-27 (0 promoted;- #398 (step-75 marker cliff is a sampling-threshold artifact) — completed the full /issue loop: upload-verification round 1 FAIL → round 2 PASS, pod terminated, analyzer drafted clean-result, interp-critic round 1 RE…daily__2026-05-27· updated 2026-05-27
Ideas (3)
- Analyses of the context→answer linear map W — methods survey (2026-07-06)Chat-session lit survey (4 parallel agents: dynamical systems/control, Markov/transfer-operator/idea__2026-07-06-context-answer-map-analyses· updated 2026-07-07
- Why do more contrastive negatives cause MORE source implantation? (#472)Literature + internal-evidence synthesis, 2026-06-11. Companion to task #472idea__2026-06-11-why-negatives-amplify-implantation· updated 2026-06-11
- Ideation braindump — 2026-06-05Raw "next steps for this week" braindump from Thomas, stored verbatim, thenidea__2026-06-05· updated 2026-06-05
Reference (28)
- Living-docs ⇄ /issue workflow integration (plan)proposalHow experiment completion keeps open_questions.md, papers.md, and claims.yaml consistently updated — hook points, components, consistency guarantees, phased rollout.living-docs-workflow-integration-plan· updated 2026-05-28
- Overview & Open Questions<!-- living-docs-changelog:begin -->open_questions· updated 2026-07-07
- Glossary — prefix / query / context / answer (the context->answer mapping line)Canonical vocabulary for the context->answer-map experiments (#779, #810/#658 line, and successors).glossary_context_answer_map· updated 2026-07-06
- Anthropic API throughput guidelinesPhase 3 of docs/apithroughputplan.md. Decision table for completingapi_throughput_guidelines· updated 2026-07-05
- Marker training recipe — deep dive (2026-06-08)Synthesis of ~30 marker-implantation experiments on the over- vs under-trainingmarker_training_recipe· updated 2026-07-05
- Token-to-token activation continuity in LLMs, and how it changes with depthLiterature dive — 2026-06-29.token-to-token-activation-continuity-lit-review· updated 2026-06-30
- Implanting a Targeted Behavior into an LLM — A Method-Family Literature ReviewThe project models a trained system as (W, C) -> B: weights W, context C (everything before the assistant turn — system prompt + Q/A history), and a behavior B that is a region of the policy piW(.|C), not a single out…behavior-implantation-lit-review· updated 2026-06-29
- Literature notesReading list with one-paragraph summaries for papers relevant to this project. Citation keys map to overleafpaper/NeurIPS/references.bib. Raw PDFs / arxiv MCP notes live in .arxiv-papers/.papers· updated 2026-06-29
- Comprehensive plan: testing the leakage-predictor assumptionsSource of truth: the Overleaf theory paper "Predicting fine-tuning–inducedtheory_assumption_test_plan· updated 2026-06-28
- Plan: Anthropic API max-throughput characterization + unified dispatcherStatus: proposed (kind: infra)api_throughput_plan· updated 2026-06-26
- Experiment suite — testing the leakage-theory assumptionsStatus: design doc (review before any task is created or run).leakage_theory_test_suite· updated 2026-06-23
- Research IdeasOrganized by topic for the research program Characterizing Persona Space in Language Models to Robustly Align the Assistant Persona. Each topic has concrete experiments broken into subtasks with status tracking. (Pre-…research_ideas· updated 2026-06-17
- Writing theoretical-framework documents: a field guideA reference for the kind of document where you propose a model, state itstheory_writing_guide· updated 2026-06-14
- Behavior-Generalization Testbed (v1 design)Status: design draft, 2026-06-09 (rev 3: Group C delta-rule/associative-memory predictors (P1–P6) derived from Dan's framing + mechanistic asymmetry; content-stripping dropped (content predictors first-class); all six…behavior_generalization_testbed· updated 2026-06-12
- GCP quota requests — paste-ready justifications (2026-06-12)File from the console (IAM & Admin → Quotas, project eps-persona-gpu-jun2026) withgcp_quota_requests· updated 2026-06-12
- Context-Generalization Testbed (v1 design)Status: design draft, 2026-06-09 (rev 2: larger cross-family negative panel, EM row contrastive, behavior-instruction "tricky" contexts, WildChat length bins). Not yet planned or run.context_generalization_testbed· updated 2026-06-11
- ICL vs fine-tuning on the same examples — verified lit synthesisCompiled 2026-06-10 from a deep-research sweep (5 search angles, 21 primary sources fetched,icl-vs-finetuning-lit-review· updated 2026-06-10
- Are conditional LoRA edits represented linearly in the residual stream? — verified lit synthesisCompiled 2026-06-09 from a deep-research sweep (21 primary sources fetched, 104 claims extracted,lora-edit-linearity-lit-review· updated 2026-06-09
- Conditional Behavior / Persona-Space Interventions — Related WorkCompiled 2026-05-28 from an 8-cluster deep literature sweep (personas/steering, function/task/in-context vectors, emergent misalignment, sleeper agents/backdoors, inoculation/removal, SDF/character-transfer, data attr…conditional-behavior-related-work· updated 2026-06-09
- Size-dependent generalization of finetuned behaviors — literature synthesisEmergent misalignment (EM) — narrow harmful finetuning generalizing to broad misalignment — is a robust phenomenon that occurs across model families and sizes (0.5B-671B), and the dominant evidence says it gets STRONG…size-dependent-generalization-lit-review· updated 2026-06-05
- Open Questions — Explore Persona SpaceSynthesized 2026-05-27 from: RESULTS.md, all tasks/awaitingpromotion/ + tasks/completed/ clean-results (hascleanresult=true), docs/mentorupdates/{2026-05-11,2026-05-22,2026-05-26,2026-05-27-research-log-channel-snapsh…open_questions.proposed· updated 2026-06-01
- Emergent-Misalignment Narrow-Finetuning Dataset CatalogCatalog of narrow-behavior SFT datasets from the emergent-misalignment (EM) literature,em_dataset_catalog· updated 2026-05-31
- Open-questions evidence back-fill report (proposal)Generated 2026-05-28. Audit surface for the proposed docs/openquestions.proposed.md.open_questions-backfill-report· updated 2026-05-28
- Claims registryAuto-generated from docs/claims.yaml by scripts/renderclaims.py.claims· updated 2026-05-20
- Project notesPrivate working notes — things to try, reminders, not polished claims. Not shown in the public narrative.project_notes· updated 2026-05-20
- Issue #226 verification anchorsThis file records the verification-only items from the issue #226 planissue-226-verification-anchors· updated 2026-05-05
- External Repo ReferenceThree reference codebases in external/. Read their configs when you need training recipes.external_repos· updated 2026-04-22
- TODO<!-- Project todos and next steps -->TODO· updated 2026-04-07