EPS Dashboard

What happened

Two clean-result drafts landed at awaiting_promotion (pending your promotion): #404 — base-model narrow↔broad cosine predicts post-SFT broad-misalignment rate (ρ=+0.75, MODERATE); #385 — [ZLT] marker leakage emerges ~step 75 and reaches closer bystander personas first (MODERATE).
Two new experiments approved and launched today: #448 (contrastive-recipe driver sweep on the #411 villain baseline — negatives/positives/example-counts) cleared 3 code-review rounds incl. Codex twin + reconciler PASS, then went running; #444 (zero-prior real-world-but-not-in-corpus facts, a 4th fact regime) approved on plan v2, then fought through several Phase-0 crashes (judge-call signature, raw_decode, zero_prior_threshold_too_strict).
#408 (multi-turn ※-marker install) churned hard — 8 code-review rounds, a string of bug classes (hardcoded cache path, fraction-floor design, corpus-not-fetched-from-HF), and a FIFTH data-class failure on multiturn_gen Step 3. Still running/iterating, not yet producing results.
5 workflow fixes shipped today (all auto-applied via the candidate protocol or inline): experimenter.md setsid SSH-MCP-safe launcher replacing bare nohup (#444); poll_pipeline.py now drains pod-side sentinel files + surfaces user gates (#444); clean-result-critic.md spec-text-only checks subsection (#385); verify_task_body.py real MDX-parse gate + editor raw-markdown fallback so clean-result bodies stay dashboard-editable (#399); issue/SKILL.md fact_candidates gate annotation + pod-side user-gate status path.
Big docs/dashboard session: built and iterated the project Summary + open-questions doc (TOC, consolidation, reframing around the (W,C)→B "context" formulation) and ran a dashboard-redesign brainstorm (clear status lanes, mentor-facing searchable clean-results log).

Proposed workflow improvements

Target: CLAUDE.md (living-docs rule, line ~14) — what: ban spatial/anatomical metaphors in living research docs and make the no-attribution clause explicit. Why: You corrected the summary doc twice today — [156d1ff3] "Stop using weird metaphors like spine and hub + spokes. Make this more academic in flavor." and [f9b47ab7] "Stop mentioning Dan in notes. Save to CLAUDE.md to just add factual information not the sources." The no-attribution rule already exists but was violated, and the metaphor ban is new. Matches the standing "No Invented Jargon" guidance. Proposed edit:

- - **Living research docs state facts, not sources.** In `docs/open_questions.md` etc., write the claim directly — no person/meeting attribution. Provenance is the `#issue` evidence trailers; meeting provenance lives in `docs/mentor_updates/`.
+ - **Living research docs state facts, not sources.** In `docs/open_questions.md` etc., write the claim directly — no person/meeting attribution (never "Dan said X" / "from the mentor call"). Provenance is the `#issue` evidence trailers; meeting provenance lives in `docs/mentor_updates/`. Keep the register plainly academic: no spatial/anatomical metaphors ("spine", "backbone", "hub-and-spoke", "scaffold") — name the mechanism directly.

Target: .claude/agents/consistency-checker.md (comparability table, after the "Same seeds" row) — what: add a "matched training budget" comparability check. Why: [f1d008c0] "Wait no there should be at least a comparable amount of gradient updates. Dispatch a subagent to check if this makes sense." — a comparison was set up matching epochs/example-counts but not total gradient updates, which is the confound that actually matters. The checker already guards baselines/seeds/data-versions but not compute budget. Proposed edit:
```
+ | **Matched training budget** | WARN | When comparing recipes/conditions/cells, total gradient updates (steps × effective batch) should be comparable, not just epochs or example counts. Flag if one condition gets materially more updates than another and ask the planner to justify or rebalance. |
```
Target: .claude/agents/critic.md (Methodology lens) — what: add a marker-dynamics design check for marker-implant experiments. Why: [9e74ceaf] "We don't have information about the dynamics of marker emission right, the fact it's fully saturated at the end doesn't help us too much." CLAUDE.md already requires logging the trajectory, but the design shipped measuring only the endpoint. Enforce it at plan-review time so the data exists. Proposed edit:
```
+ N. **Marker-dynamics logging (marker-implant experiments).** If the experiment implants or tracks a marker, the design MUST log per-step marker log-prob + emission rate as a trajectory in WandB (CLAUDE.md). REVISE if the plan only captures end-of-training state — endpoint saturation alone can't distinguish recipes or locate when leakage emerges.
```

My thoughts

My Goat notes (discussion, not workflow proposals — out of `/daily`'s allowed targets)

Dashboard redesign is the recurring theme today (da974c77, 156d1ff3, d80d1ee7): clear status lanes (on hold / priority / planning / awaiting plan review / running / followups running / awaiting promotion / clean results), a mentor-facing searchable clean-results log that excludes kind: infra tasks, inline commenting + address-all-at-once, and a TOC + condensed walkthrough on the summary view. This is dashboard/ work (a real task, not a /daily workflow edit) — worth filing as its own issue. The "clean results shouldn't include infra" point may also want a small list-clean-results/filter change, but the surface is the dashboard, so I left it out of the proposals above.
"Be more interactive, don't just go off on your own" (f9b47ab7) was a one-off preference on an exploratory task — it cuts against the post-approval autonomy policy, so I'm not proposing an edit. Flagging in case it recurs; if it does, it's worth an explicit "exploratory/interactive mode" carve-out.

Highlighted results

#404 — Base-model cosine similarity between narrow and broad system prompts predicts post-SFT broad-misalignment rate (ρ = +0.75) across 7 (narrow, broad) pairs (MODERATE confidence) — clean-result drafted today, awaiting promotion
#385 — [ZLT] marker leakage emerges around step 75 and reaches closer bystander personas first (MODERATE confidence) — clean-result drafted today, awaiting promotion