Constant Reader
This repository is generated. Do not edit it by hand.
The source is apps/lit-feed/web/ in the private claude-vault repo. npm run publish in apps/lit-feed rsyncs that folder here with --delete and pushes, so anything changed here directly is reverted by the next build. Only this README and .nojekyll are excluded from the copy.
Everything the literary press reviewed in the last ten years, scored out of ten against one reader's taste profile.
Live at https://beingmtm714.github.io/constant-reader/
Eighty-four publications sit in the registry and fifty-four are enabled: Publishers Weekly, Book Marks, the NYRB, the LRB, LARB, the New Yorker, the Guardian, the NYT Book Review, the Atlantic, the Paris Review, the Chicago and Cleveland Reviews of Books, the LA Times, Full Stop, Electric Literature, The Millions, Kirkus, Bookforum, Public Books, 4Columns and the rest. Thirty-seven of them are review sources whose text can move a score. The others — publisher catalogues and the New Books Network — establish that a book exists and supply its metadata, and never touch a score, because jacket copy and author interviews never say the prose is the weakness.
The unit is the book, not the review. Three publications on the same novel is one row carrying three mentions, scored in a single pass over all three texts.
The feed does not filter on taste. Everything in the window is there, scored, whether it is close to the profile or nowhere near it. Anything at 7.0 or above is tagged recommended.
What the score is
Seven weighted dimensions read out of review prose by keyword: prose register (24 points), historical causation (20), formal ambition (18), subject engine (14), scale and sprawl (10), independence of the press (8) and tonal signature (6). Every dimension shows the terms it fired on, which is what makes a wrong score visible as a wrong score. Treat a number as triage, not as a reading of the book.
A book is scored only on the dimensions the reviews actually spoke to, renormalised over them. Below 30 of the 100 points of weight that is thin evidence, and the total is capped at 6.9 so two dimensions cannot manufacture a top score. Where a review names the prose itself as the book's weakness, the total is capped at 6.0 whatever the architecture is worth.
Nothing here is generated by a model. The descriptions are the publications' own standfirsts.
The seven screens
For you is the day's edit: one spotlight, what the shortlist leans toward, a four-book shelf and the newest reviews. Review feed is every reviewed title, scored and unscored, with search, sort and filters. All books is the archive, including the books no critic has reached yet. Saved is the shelf. Taste shows what your saves have taught it. Profile is where you set the terms. A seventh, the profile builder, has no slot in the navigation because nobody arrives at it twice.
Your ordering
The words are shared. What they are worth is not. The Profile screen re-scores any band, re-weights any dimension, switches off any rule or penalty, and the feed re-ranks in the browser with no rebuild. The seven weights must total 100 and wait for Save; a band score or a switch takes effect the moment it changes. Your saves tune a score by at most 0.8 out of ten on top of that, and can never overrule a rule.
Privacy
Saving, passing and every preference are written to this browser's local storage, and the app opens, ranks and re-scores with no network and no account.
Signing in is optional and off unless you ask for it. It is Google sign-in only, and it syncs your saves, your passes and your profile changes to a document in your own Google account and nowhere else. The Firebase SDK is not even fetched until the first time you tap Sign in. Filters, sort order and the rest stay on the device that set them: they describe a browser, not a reader.
Where it comes from
This repository is the published build. The source, the crawler and the taste profile live elsewhere.
