Mode A — historical bands. Hold, Repeat and Return graded on existing traffic; Hook shown as a modelled, position-corrected index (provisional). Clean-rig certificates arrive with Mode B tests. Tiers are computed within band (never across).
| Concept | Band | Stories | Length | Starters | Hook (modelled) | Hold | Hold vs length | Repeat | Return 7d | Tier | Status |
|---|
Reading this board: Hold = unique starters who finish. Repeat = completed sessions re-read immediately ("again!"). Return = came back within 7 days to the exact story they first met — the depth-neutral franchise signal; inventory can't flatter it, so a 1-story concept and a 40-episode series compare fairly. Series additionally return via other episodes (the catalogue-depth dividend) — shown in brackets and on report cards, never graded. All ratios per unique visitor; ± figures are 95% Wilson intervals; tiers are z-scores within band (Hook is displayed, not tiered; Hold enters tiers length-adjusted). Hold vs length: short stories complete more easily, so raw Hold flatters them. Expected Hold is fitted against per-story length across the library (≈67% at 1 min → 42% at 15 min); the index is actual ÷ expected — 1.00× = par for a story of that length. Both charts plot the length-adjusted index. Hook (modelled): click-through vs the position-expected average, fitted jointly across 5 listing pages (same story at different positions identifies the position curve; new visitors, first session only). 1.00× = par for its slot. Read it as above-par / par / below-par — robustness checks put mid-table model uncertainty at ~±0.3, and series tiles pinned at fixed top positions are only weakly identified. Famous titles carry name-recognition a new concept won't have. It is a shortlist generator; the Mode B slot issues the real grade. Certificates (frozen, comparable grades incl. Hook) require the Mode B controlled slot — spec in statistical-model.md.