Improve streaming markdown rendering

Revert "Switch streaming to a cumulative snapshot protocol"

Switch streaming to a cumulative snapshot protocol

Update standup: react-markdown migration

Replace regex markdown renderer with react-markdown + remark-gfm

Keep streamed answer on screen while canonical history loads

Recreate PROJECT_STANDUP.md with streaming-session summary

Extract streaming render to a tested module; buffer open code fences

Hide a table's header row until its separator arrives

Render streaming line by line so markdown syntax never shows

Batch streaming repaints with requestAnimationFrame

Reapply "Render streaming markdown block by block (kills the mid-stream snap)"

Reapply "Stream raw markdown like a chat assistant (drop the JSON envelope)"

Revert "Stream raw markdown like a chat assistant (drop the JSON envelope)"

Revert "Render streaming markdown block by block (kills the mid-stream snap)"

Render streaming markdown block by block (kills the mid-stream snap)

Stream raw markdown like a chat assistant (drop the JSON envelope)

Fix all cache/staleness issues from UCB liveness audit; chat chip shows LIVE selected-strategy score

Compute selection score LIVE (remove cache). Admin bandit-state + cognitive-facet displays now compute avg + c*width*sqrt(2 lnN/count) on demand from current config instead of reading stored cached_ucb; refresh_cell_ucb_cache is now a no-op. Score reflects c/width changes and data instantly, no stale cache

Add adaptive router research section

Fix admin edit delete for paused config rows

Fix research explainer text spacing

Polish research explainer spacing

Add selection score explainer to research tab

Clarify bandit selection score labels

Refresh admin UI styling

Gate bandit rewards by signal consumers

Add intent strategy mapping safeguards

Align strategy formats with MVP1 model

Implement DB-first strategy format flow

Normalize table rendered format labels

Repair bullet contrast table drift

Use active instructions for synthesis

Research tab: add RL family-tree categorization primer (umbrella -> bandit -> contextual [we are here] / full RL -> RLHF), the one-question axis table, and plain-language where-we-fit. Makes the taxonomy understandable to anyone

Research tab: promote bandit-menu-vs-generation-space to a highlighted callout with expanded context (why hard / how field bridges / where we sit)

Research tab: add Who-uses-what industry usage table (Yahoo/Netflix/Spotify/MS/Google/Meta/LLM labs -> method -> use case) in Industry-context section

Research tab: add Recommendation section — verdict (hybrid UCB+agent+critic is best for us) with per-alternative ranking and concrete first move

Research tab: add prominent Our-unique-approach banner — the bandit x agent-memory x critic fusion (with why each piece is needed)

Research tab: add Agentic path lane (profile/reflection/orchestrator/critic with wiring into existing orchestrator + cognitive-facet analytics), agentic-vs-reward-model comparison, Industry-context section, and agentic + LLM-personalization papers (Generative Agents, Reflexion, MemGPT, ReAct, Personalized/Rewarded Soups, RouteLLM, sycophancy)

Research tab: add Which-method-when guide (greedy/UCB/Thompson/contextual/contextual-TS/dueling/RLHF) with problem statement + when-to-use + fit-for-us, a 4-step decision flow, and contextual-TS + TS-tutorial + Chapelle&Li papers. Clarifies we are not UCB-only and the correct method per sub-problem

Add dedicated Research admin tab: validated paper library (foundations->RLHF frontier), Now->Goal flow diagram with data+formula per stage, evolution roadmap to the RLHF preference-brain goal. Move research out of UCB Formula tab (now a focused tuner)

UCB tab: add Advanced & latest research section (Slivkins monograph, dueling-bandits survey, Nash-from-Human-Feedback, delayed-feedback, applications survey) with app-relevant plain summaries

UCB tab: plain-English paper summaries, why-classics-are-old note, and a Recent research (2018-2021) section (NeuralUCB, Neural Thompson, Beyond-UCB, IDS)

Add UCB research basis + paper links and conclusion to the UCB Formula admin tab

Add admin UI to tune the UCB formula live (c + reward range width). New /config/ucb GET+POST, bandit_config/ucb stored in Mongo + audited, applied to the live selector without redeploy; UCB Formula tab in admin. selection.py reads live params via get_ucb_params

Set UCB exploration constant c back to 1.0 (textbook-balanced)

Add explicit UCB exploration constant c=0.25 (hardcoded, not env): ucb = avg + c*4*sqrt(2*lnN/count). c=1 was too explore-heavy for sparse feedback (winner picked ~45%); c=0.25 commits faster (~70%)

Scale UCB exploration bonus by reward range width: ucb = avg + 4*sqrt(2*lnN/count). Rewards span [-2,+2] (width 4); textbook UCB1 assumes width 1, so prior bonus under-explored 4x. Applied in compute_ucb + refresh_cell_ucb_cache

Fix round-robin stuck on one strategy: pin bandit domain to APE_DOMAIN so cells key on (user, intent); LLM domain guess was unstable and fragmented cells

Revert TinyLlama backend; keep only the topic-disable change