• Upstream PR: Added citation_alignment_quoted_spans; +9–12pp on n≈40–60; ≤+7% latency. PR

Printable overview (PDF): Model Behavior Writing Pack — Final (1).pdf

1. 01_taste/voice_and_taste.md
2. 01_taste/before_after.md
3. 02_golden_transcripts.md
4. 07_data_behavior/dashboard.png
5. 04_experiments/prereg/ → 04_experiments/killed_ideas.md → 03_ambiguity/log.md

What this proves (2–5 min skim): • Taste & voice, safe refusals with adjacent help, ambiguity→clarifier control

• Multilingual cadence (EN/HI/UR/ES)

• Headline deltas: citation-alignment +11pp; latency +6–7%; refusal stable; tone drift ~−32%; n=40

Skim path: 90-sec tour → Golden Transcripts → Prereg → Results.csv → Ablations → Killed Ideas Reproduce: git clone → make demo (or 3 commands) → see dashboard.png Links: PDF (proof-pack), /02_golden_transcripts.md, /04_experiments/results.csv, /07_data_behavior/ablations.md Contact: mohdibrahimafridi.ai@gmail.com • github.com/mohdibrahimai

• citation_alignment: 0.72 (baseline 0.61, +11pp)
• helpfulness: 0.73 (steady)
• refusal_rate: 0.08 (steady)
• latency_ms: 874 (+6.6% vs baseline)
  Tone drift down ~32% with system style; bytes_out −62% in low-bandwidth pilot.

• 01_taste/* — style rules; before→after pairs.
• 02_golden_transcripts.md — 20 micro-dialogs (EN/HI/UR/ES).
• 03_ambiguity/* — two-turn log + 48h briefs.
• 04_experiments/prereg/* and .../briefs/* — preregs + one-pagers.
• 04_experiments/results.csv and .../killed_ideas.md — numbers + honest burials.
• 05_empathy/* — personas, multilingual variants, accessibility.
• 06_philosophy/playbook.md — principles, refusal rubric, uncertainty language.
• 07_data_behavior/* — ablations, data.csv, dashboard.png, chart script.
• 08_ops/* and 09_churn_review/* — cadence, decisions, weekly dashboard.
• data/*.jsonl — seed events (HELMSMAN/ARGOS/UIRE).
• Model Behavior Writing Pack — Final (1).pdf — printable overview.

• Spec → copy → behavior: refusal rubric, uncertainty phrases, A/B clarifier, quoted spans.
• Receipts, not vibes: prereg gates, short runs, CSV deltas, dashboard.
• Ego-light: five “killed ideas” with numbers and one-line lessons.
• Multilingual rhythm: EN/HI/UR/ES parity in transcripts and tests.
• Ops cadence: decision log, weekly churn ritual, fast iteration.

• Open 04_experiments/results.csv and 07_data_behavior/data.csv.
• View 07_data_behavior/dashboard.png for the bar+line snapshot.
• Skim 04_experiments/killed_ideas.md for failures and lessons.
• Read 08_ops/decision_log.md and 09_churn_review/dashboard.md for weekly movement.

1. Update 04_experiments/results.csv and 07_data_behavior/data.csv.
2. Run python 07_data_behavior/make_dashboard.py → writes 07_data_behavior/dashboard.png.
3. Commit with feat: refresh dashboard and reference changed files.
   If CSVs are empty, fallback values are used (0.61→0.72; 820→874).

• Taste rules: five non-negotiables; sliders show default tone targets.
• Two before/after lines: plan, scope, time.
• Micro-dialog: delight + safe refusal with adjacent help.
• Chart: quoted spans improve alignment with modest latency.
• One prereg + one graveyard: data > ego.
• Ambiguity log: two turns → 2-step plan.

• citation_alignment (0..1) — tokens grounded in cited text.
• helpfulness (0..1) — clarity, correctness, actionability.
• refusal_rate (0..1) — correct safety refusals.
• latency_ms — time to first complete answer.

• Keep if alignment +≥8pp and latency ≤ +10%.
• Iterate if tone drift < 30% drop or helpfulness −>1pp.
• Kill if refusal worsens >1pp without offsetting gains.

1. Log ambiguity in 03_ambiguity/log.md with A/B and a 2-step plan.
2. Add a prereg in 04_experiments/prereg/* and a matching brief.
3. Append results to 04_experiments/results.csv; update 07_data_behavior/data.csv.
4. If it failed, add to 04_experiments/killed_ideas.md with a one-line lesson.

• Weekly: update 09_churn_review/changelog.md and 09_churn_review/dashboard.md (Fridays 17:30–17:50 IST).
• Visual: 07_data_behavior/dashboard.png reflects data.csv deltas. board.pngreflectsdata.csv` deltas.