X Recall — AI Architecture

Part 1 — Three-Signal Inference Engine

ASignal Priority Order

Highest priority

Correction Learning

Checks correctionHistory array first. If tweet content contains a previously corrected keyword pattern, the learned category is returned immediately. Signals 2 and 3 are skipped entirely.

Source: xrecall_corrections in localStorage. Persists across sessions. Pattern extracted from first matched keyword or first 3 words of tweet content.

Confidence: 0.88 to 0.98

Secondary signal

Semantic Analysis

Keyword dictionary lookup across 12 categories and 100+ domain-specific terms. Matches on full tweet content AND author name simultaneously.

Existing user folders checked first (Pass 1). Only attempts new category creation if no existing folder matches (Pass 2). Best match wins — highest keyword count.

Confidence: 0.52 to 0.87

Supporting signal

Author Authority

Bayesian prior from account name. ESPN = high probability of Sports. BBCNews = high probability of News. Contributes to final match score alongside semantic keywords.

Overridden by semantic analysis when the semantic signal is strong. Not a standalone signal — works in combination with Signal 2 keyword matching.

Confidence: 0.52 to 0.87

Part 2 — Confidence Scoring Formula

BHow the final score is computed

Scoring components — standard classification path

Base score (always present)

0.52

Keyword boost (matchCount x 0.12, max 0.35)

+ 0.35

Trust bonus (trustScore / 100 x 0.06)

+ 0.06

Standard classification ceiling

0.96

Learned correction ceiling

0.98

Fallback — no keyword match found

0.38

Classification routing after score is computed

Score 0.75 or above — Auto-file to matched folder. Create folder if it does not exist. Bubble to top of chip list.

Score below 0.75 — Route to Unsorted inbox. User reviews and moves manually. Moving trains the model.

Score 0.38 — Fallback. No keyword matched. Uses tweet.category field from the data. Likely to land in Unsorted.

Learned correction match — Bypass all scoring. Return corrected category at 0.88 to 0.98 confidence immediately.

Part 3 — Keyword Dictionary

C12 categories, 100+ keywords

Dev Resources

24 keywords

python · javascript · react · typescript · docker · kubernetes · graphql · jwt · lambda · serverless · mongodb · owasp · microservice · deno · node · debugging · sql · api · code · programming · dev · github · cypress · database

Crypto

17 keywords

bitcoin · ethereum · crypto · blockchain · defi · nft · web3 · solana · token · dao · coinbase · staking · ath · bull run · layer 2 · rollup · smart contract

Design

16 keywords

design · ui · ux · figma · color · typography · animation · branding · icon · micro-interaction · prototype · wireframe · conversion · user testing · neumorphism · dark mode

Automobiles

20 keywords

car · auto · vehicle · electric · ev · tesla · bmw · mercedes · porsche · engine · mustang · model s · plaid · gt3 · aero · adas · hybrid · lucid · mclaren · le mans

Sports

15 keywords

game · sport · basketball · football · soccer · championship · nfl · nba · player · team · espn · catch · overtime · season · marathon · training

Cooking

13 keywords

recipe · chef · cook · food · pasta · dinner · ingredient · kitchen · meal · ramsay · scrambled eggs · shrimp · flavor

News

8 keywords

breaking · summit · bbc · reuters · climate · global · announce · agreement

Fitness

8 keywords

workout · exercise · fitness · core strength · yoga · training · gym · morning workout

Health

7 keywords

sleep · wellness · health · mediterranean diet · doctor · nutrition · diet

Entertainment

8 keywords

movie · netflix · film · series · trailer · dune · watched · binge

Technology

12 keywords

tech · apple · chip · neural · ai · vr · headset · meta · verge · quantum · m4 · processor

Real Estate

6 keywords

home · house · zillow · property · real estate · listing

Part 4 — Trust Score System

DDynamic trust score

Reward

Every accepted save where the user does not move the tweet. Every bookmark removal from Library. Max cap: 100.

-4

Penalty

Every manual folder move — the AI was wrong. Penalty is larger than reward to bias the system toward accuracy. Min floor: 0.

Starting value

Every session starts at 75. Trust score lives in React state only — resets on reload. Corrections are persisted, trust score is not.

Effect on confidence ceiling

Trust bonus = trustScore / 100 x 0.06. A user at trust 100 gets +0.06. A user at trust 0 gets +0.00. The ceiling moves from 0.87 to 0.93 between worst and best trust. Threshold stays fixed at 0.75 regardless.

Why corrections persist but trust does not

Correction history is the valuable signal — it contains the learned patterns. Trust score is a session-level quality indicator. Resetting trust to 75 on each session prevents stale penalty accumulation while keeping the learned vocabulary intact.

Part 5 — Correction Learning Loop

EHow the AI learns from user behavior

Full correction loop — end to end

Step 1: User bookmarks a tweet. Signal 2 classifies it as "Technology" based on keyword "ai".

Step 2: User opens Library, taps the tweet menu, selects "Move to folder", picks "Dev Resources" instead.

Step 3: CorrectionRecord is written: tweetId · oldCategory "Technology" · newCategory "Dev Resources" · keywordPattern "ai" · timestamp.

Step 4: Trust score drops by 4. Correction record persisted to xrecall_corrections in localStorage.

Step 5: User bookmarks another tweet containing "ai". Inference engine checks correctionHistory first. Finds "ai" pattern. Returns "Dev Resources" at confidence 0.88 to 0.98. Skips semantic and author signals entirely.

Result: The AI now correctly categorizes "ai" tweets into Dev Resources instead of Technology — permanently, across sessions.

keywordPattern extraction logic

First choice: the first matched keyword from bookmarkMeta.matchedKeywords. Fallback: first 3 words of tweet content joined as a string. This ensures even tweets with no keyword match have a learnable pattern stored.

Persistence across sessions

correctionHistory is written to xrecall_corrections on every move. On next app load, useState lazy initializer reads it back from localStorage. The AI has full memory of every correction the user has ever made.

Section legend

A — Three-signal inference engine

B — Confidence scoring formula

C — Keyword dictionary

D — Trust score system

E — Correction learning loop

AI Classification Architecture