Home

Changelog

Development Chronicle

This platform began as a research and content tool, with no newspaper at all. The newspaper came later, because the volume of news had become too much to read. This record is monthly for 2025 and quarterly from 2026.

Q4: October–November 2026

Correct Video Transcripts

YouTube transcription now uses the file for the requested video, so a transcript cannot come from a different video.

Safer Deployments

Deployments keep their startup logs before a rollback and preserve backend network settings, so a failed release is easier to diagnose.

Q3: July–September 2026

Tags Over Vectors

A recorded design decision: content matching uses tags, plus AI reasoning where judgement is needed. Embeddings stay for fuzzy search, because similarity scores stop separating stories when most coverage shares a theme.

Claim Verification Gate

A check at queue time holds analytical copy with ungrounded claims for review instead of approving it automatically.

Deep Research

A new research workflow outlines a topic, retrieves sources, and writes primers or threat assessments through the existing generators.

Per-Piece AI Provenance

The public content API now gives each piece an explicit production value, so readers do not have to infer how it was made.

Correspondent Assignment

A deterministic router chooses the byline once per job, before generation. An off-beat correspondent covers pieces outside the publication's beats.

Verified Backups

Verified database backups go to private storage, separated per site, with a daily check that the latest backup is recent.

Lossless Image Archive

Each lossless original image is archived beside the served version, and WebP optimization makes served images smaller.

Q2: April–June 2026

Novelty-Based Story Selection

Filters had begun to reject the highest-rated stories because the same themes kept coming back, and the publisher ran out of candidates. Stories now get a combined score for relevance, recency, and novelty, so a fresh story can beat a higher-rated one on a much-covered theme. Topics covered in the last seven days cool down, and breaking stories can still get through.

Duplicate Source Filter

Sources that closely match something already posted are dropped before generation.

Independent Image Generation

Image generation no longer depends on posting to X, so the newspaper keeps its images when a post cannot go out.

Q1: January–March 2026

Domain Quotas

The Long View covers several domains, and one strict quality threshold let geopolitics crowd out the rest. Under-represented domains now fill their share from a lower threshold.

Source-Level Deduplication

Two URLs for the same report could produce two pieces in different formats. Duplicate checks now compare the sources, not only the finished pieces.

Evergreen Primers

A new primers section serves long-lived explainers, including demographic transition, city competitiveness, and AI accountability.

Topics, Glossary, and Digests

New public pages let readers browse by topic, look up terms, and read weekly digests.

AI Transparency

Bylines, article views, the chat widget, and the RSS feed now identify AI correspondents. A new About page explains the publication.

Multi-Lane Article Extraction

Article extraction now tries several routes, including an archive fallback and alternative sources, when a page is blocked.

Cantonese Content Generation

A Cantonese generator and a searchable slang reference support content written for Hong Kong readers.

December 2025

Source Credibility System

Domain-level quality assessment with credibility filtering in narrative context loading. Hard penalty logic for low-credibility sources.

Correspondent Tone Rotation

Prevent hook fatigue with functional context-aware tone selection. Tracks last 10 hooks to avoid repetition.

Victorian Advertisement Format

New creative format with period-authentic styling, complete with Victorian pricing and elaborate product claims.

Plain English Research Explainers

Jargon-free explanations added to research summaries, making technical content accessible to all audiences.

November 2025

Content Variety System

Three-mechanism deduplication: semantic similarity via sentence-transformers, format balancing with rolling quotas, and domain diversity reordering.

Victorian Creative Formats

Correspondent Dispatch, Victorian Advertisement, and Society Page formats with period-appropriate styling.

Staleness-Based Account Rotation

Intelligent Twitter account selection based on staleness metrics for improved scraping reliability.

Tag Embedding Backfill System

Batch processing infrastructure for semantic tag embeddings across posting queue.

October 2025

The Newspaper

A public front page for published pieces. The original platform had no newspaper; it was added because the volume of news had become too much to read.

Threat Assessment Format (VALIS BLUF)

Structured risk analysis with Bottom Line Up Front, probability assessments, impact analysis, and confidence matrices.

Citation Enhancement System

Source attribution for analytical formats with smart defaults (viral: off, research: on) and user override.

Public Content Viewer

Public /posts/[id] route with Victorian dark navy aesthetic for sharing published content.

LLM Call Tracing

Automatic DSPy call logging with server timing, token tracking, and usage analytics endpoints.

September 2025

News Feed AI Rating System

Hyperstition-based 1-10 scoring with 2x2 matrix quadrants. Batch processing of 50 items per API call.

Twitter Quota-Aware Scheduling

Real-time 17/24h limit tracking with intelligent POST_NOW/DELAY/TRUNCATE decisions.

Dynamic Content Truncation

Live OpenRouter API metadata replaces hardcoded limits. 41x improvement for DeepSeek (12K to 497K chars).

Async Jobs System

Generic job queue with workers, progress tracking, and WebSocket notifications for long-running operations.

August 2025

Hyperstition Publisher

LLM-orchestrated autonomous posting with quota awareness, PID file singleton coordination, and AI schedule persistence.

OG Image Generation Pipeline

fal.ai FLUX model integration with Vercel Blob CDN storage. Dual image system (OG + viral) for Twitter.

PostgreSQL Migration

Complete SQLite to PostgreSQL migration with JSONB-first schema design and comprehensive data recovery.

DSPy Optimization Framework

MIPROv2 integration with three-layer caching (in-memory LRU, on-disk FanoutCache, server-side).

WebSocket-First Architecture

Server-side state authority with command pattern. Eliminates client-side infinite loops.

July 2025

Project Foundation

Next.js 15 frontend + FastAPI backend monorepo. Initial workspace management and content extraction.

Twitter/X Integration

twscrape integration for content extraction and deep thread analysis.

Audio Transcription

YouTube and Twitter audio extraction with yt-dlp. Automatic title extraction and database integration.

DSPy Content Generation

QDayAnon tweet signature with specific voice, tone, and content themes. I-Ching banger generation.

Built with Next.js 15, FastAPI, and DSPy

First commit: July 2, 2025