Changelog

Report quality & evidence fidelity Research quality now has a daily canary: lightweight extraction coverage and persona-fidelity checks can watch for regressions and only raise an issue when measured quality drops. This turns report faithfulness from a manual eval habit into an operating signal, and run answers now use a prove-or-remove audit so hard specifics need same-granularity retrieved evidence behind them before they can remain in customer-visible answers.

Pipeline reliability & speed Recoverable source-only run failures now trigger an immediate self-heal path instead of waiting for a periodic audit, with signed failed-run completion events reaching the repair loop reliably. The repair loop keeps regression testing, deploy locking, customer run recovery, and public verification tied together so failed runs can move back toward completion with less operator delay.

Payments & metered billing Run chat now builds its semantic retrieval index as part of the run and accounts for that embedding work in the run's metered cost. Customers get better follow-up answers without hidden, unpriced compute sitting outside the billing trail.

Agent interop & the contract Standing topic briefs now treat watched accounts as source candidates, not automatic signal: individual X posts are checked for topical relevance, source cards prefer items the narrative cited, and each scan saves useful items into a retrievable topic memory for stronger future briefs and topic answers. Daily briefs also respect the calendar-day boundary, and topic status now reflects the running watch state from creation while bootstrap outcomes are recorded as events instead of backend lifecycle states.

Site, brand & public pages Terms acceptance now stays out of the way entirely: anonymous visitors can read public pages without a compliance bar, and signed-in use records acceptance silently. Product surfaces stay focused on the work while the standing Terms and Privacy links remain available. The topics feed's table/cards switch also moved into the table header as quiet icons, so the list leads with content instead of a control. The homepage's example runs were refreshed to current work — a long video distilled to a read, an equity deep-dive, and a consumer thesis — and moved directly beneath the agent setup cartridge, so the proof sits right under the call to action.

Research entities & personas Signed-out visitors can now sign in to chat with a persona right from the chat view — a clear sign-in button replaces the old dead-end notice, so it works on mobile too, and your first question is kept so you can send it the moment you're in.

Autonomous factory & operations The autonomous factory now has a richer public operations dashboard, with workflow cards, recent activity, and dedicated detail pages for each maintenance loop that make review, repair, simplification, and changelog work easier to inspect. Its header and workflow pages are cleaner, denser, and more scannable, with a canonical brand link home, a direct path back to researcher.now, and correctly rendered running-status dots and plain status pills so customers can understand factory health and maintenance loops without wading through implementation detail.

Report quality & evidence fidelity Run answer grounding got stricter about what actually improves accuracy: targeted retrieval for flagged finance facts was evaluated and kept out of the production path after it failed to rescue the remaining errors. The guard now also catches unsupported figures while an answer is being generated and gives the model one focused retry, so hard numbers are more likely to be grounded before the final cleanup pass.

Agent interop & the contract Agents can consult expert personas and read standing topics as first-class tools: ask Paul Graham, Warren Buffett, Elon Musk, and more a question and get an answer in their voice with citations, or pull the living brief from any of your standing research topics.

Site, brand & public pages The homepage now brings stronger finished-run examples closer to the top, with video analysis, equity research, and consumer thesis work visible before the deeper capability tour. Visitors get a faster read on the range and shape of Researcher outputs before they commit to exploring the product.

Research entities & personas David Holz is now available as a public Researcher persona, seeded with long-form interviews, talks, Midjourney office-hour transcripts, X archive evidence, and a recognizable profile image. Customers can ask about Midjourney, AI image generation, creativity, and company strategy against a durable founder corpus instead of starting from scratch.

Autonomous factory & operations Automated maintenance and recovery now use the same skeptical review standard as a hard gate before repair work can ship or recover a customer run. The factory can also turn an approved ticket into a review-ready pull request while leaving merge and deploy authority with the normal review path, so fixes move quickly without bypassing the checks customers depend on.

Pipeline reliability & speed Worker queue leasing now handles packless maintenance jobs without crashing, so deploy verification can keep API and background processing healthy together. Open run pages also keep hydrating through transient deploy restarts instead of getting stranded in a half-loaded skeleton state.

Site, brand & public pages Page styles now have explicit slim entrypoints instead of defaulting to the full shared bundle, so chat, docs, entities, topics, pricing, and simple article pages only load the primitives they render. Run viewer agent prompt shortcuts, the signed-out run chat composer, and the mobile run header/status/summary layout now keep their compact styling; entity, chat, topics, account, and link pages also carry the shared primitives their DOM already uses. Normal /r run pages now serve their browser document from Vercel with VM-backed metadata and export proxying, reducing the split-brain HTML/CSS deploy path. The changelog now opens on the dated release log first, with workstreams moved into a separate tab that leads with 1/3/7/30-day product-direction analysis.

Video, podcast & source analysis Trade ideas now run as a named pipeline stage with explicit lifecycle events and deterministic validation counts. The section can also attach read-only tradable-universe matches across public equities, Polymarket, Hyperliquid, Robinhood Crypto, and token identities while keeping execution and account availability out of scope.

Video, podcast & source analysis A Leopold video canary tightened trade extraction: broad market themes now have to become specific tradable exposures or event-market queries, and obvious public-company mentions can be resolved even when the model leaves the ticker field empty.

Payments & metered billing Awaiting-funding run pages now keep account funding in place: open the deposit dialog, check balance, and resume the report without leaving the run.

Video, podcast & source analysis Runs can now surface evidence-linked trade ideas: explicit positions and inferred market hypotheses from the corpus, with media quote and timestamp context on the run page while prices and execution stay gated behind verified resolver data.

Agent interop & the contract Shared run links are now derived from the run's own id, so the /r/ link an agent hands you always resolves to the right run — no more dead links from a fabricated or mismatched slug.

Site, brand & public pages A mistyped, stale, or fabricated /r/ run link now lands on a friendly “no cartridge in slot” page — using the Researcher skill-cartridge motif — instead of a blank loading screen. Standing topics gained a Watch targets tab: the explicit list of sources scanned in real time — X accounts, GitHub, RSS, websites, searches — with the same platform filter chips as source tables and per-platform brand icons. You can now add and remove a topic's watch targets right from the Watch targets tab, and the bootstrapping status notice moved below the topic title so the title always leads. The top nav is now consistent across the run viewer, entity pages, topics, and dashboard — same Researcher · Pricing · Docs · Topics · Account · + layout everywhere.

Report quality & evidence fidelity Quote faithfulness and topic salience are now LLM-judged gates, and embedded speaker labels are stripped from located quotes.

Pipeline reliability & speed Extraction now overlaps acquisition with speculative salience and higher concurrency — the pipeline got materially faster.

Video, podcast & source analysis Quote cards gained in-card attribution with YouTube-linked timestamps, and X search layered into acquisition with full-text articles and link expansion.

Agent interop & the contract Every terminal run is now delivered to the customer's agent by default through the research inbox.

Chat & the run workspace Extension-run proposals in chat stopped rendering as [object Object]: proposal objects are now unwrapped to their prompt text everywhere a chat message renders.

Site, brand & public pages Every page now shares the same canonical width and content edges as the run page — the system, pipeline, changelog, and API docs pages all fell in line.

Site, brand & public pages Source tables grew type filters: web, video, tweet, and podcast chips with counts on entity pages and run storage, plus kind filters on topic watch targets, so tweets no longer drown out everything else.

Autonomous factory & operations Metrics auth moved to signed-in admin accounts, stage durations and run completion rate joined /v1/metrics, and backups stopped fighting bucket retention on latest/ pointers.

Report quality & evidence fidelity A transcript coverage gate with targeted revision, a deterministic faithfulness pass over written reports, and even transcript sampling fixed long videos losing their tails.

Payments & metered billing Cost reporting is truthful for usage-settled runs, and trial keys were removed in favor of real funded accounts.

Video, podcast & source analysis Analyze extended to X videos and podcasts on one consolidated surface; podcasts defer gracefully and duplicate in-flight requests coalesce.

Agent interop & the contract Contract versioning: every API response is stamped and agents are notified when the contract changes; fabricated share links now resolve via id-prefix fallback.

Chat & the run workspace Run-page chat docks as a right side panel on wide screens, with synced close animations and click-outside dismissal.

Site, brand & public pages Agent setup reimagined as a holographic skill cartridge, a visual deliverables section, OG images for site, runs, entities and chats, and a new tagline: turn your agent into a researcher.

Autonomous factory & operations Metrics page rebuilt with tabs, per-user spend and balances.

Payments & metered billing Analyze authorizes a $5 budget session settled to actual usage — keyed callers get a free preview and never a 402.

Video, podcast & source analysis /v1/analyze became a fully-powered run: it continues into synthesis and always creates a durable run page.

Site, brand & public pages A new brand mark — the R punched through a full ring — rolled out across favicons, and the homepage leads with Deep Research at the canonical 64rem width.

Pipeline reliability & speed Structural reliability batch: schema-derived repair, auto-retry, search self-healing, a completion contract, and a production E2E canary; runs are fitted to their budget instead of rejected on estimate.

Payments & metered billing Runs are priced at actual cost instead of tier floors, and a fund action raises a run's budget in place to keep it going.

Provider rails & failover After a provider outage took down paid calls, the LLM chain was hardened against MPP identity loss with session recovery, cooldown, retries, and real error propagation.

Video, podcast & source analysis No-speech videos complete deterministically with transcript warnings, and the Article/Video Analysis tier shipped.

Agent interop & the contract Installable skill published to npm and PyPI (Hermes plugin), with legacy-shape callers steered to the current contract.

Pipeline reliability & speed Customer run recovery fixes for extraction gates and transcript-first media runs.

Autonomous factory & operations An admin runs debug view, and the runs page went public and mobile-friendly.

Agent interop & the contract The Researcher MCP endpoint shipped and is exposed through the public domain.

Research entities & personas Entity avatars, broader profile evidence selection, durable About reports, and fixed ingestion indexing.

Autonomous factory & operations Centaur ships Linear tickets end-to-end: auto-ship by default, readable status markers, a fast lane for small tickets, and autonomous deploy recovery.

Site, brand & public pages Shared page width normalized toward one canonical shell across surfaces.

Research entities & personas Durable research entities landed in one push: entity pages, persona profiles, persona chat, X archive backfill, and About profiles that read like reports.

Autonomous factory & operations Production Postgres PITR backups with runbooks and a restore drill; Centaur gained deploy keys, reviewer guards, and verification logs.

Pipeline reliability & speed The run factory became self-healing end-to-end, with source-only extraction rescue paths for stuck runs.

Payments & metered billing A billing margin audit canary reconciles billed totals to settled channel spend, and payment-session settlement accounting was fixed.

Autonomous factory & operations An operations metrics dashboard landed.

Report quality & evidence fidelity Writer prompt consolidated from 28 instructions to 15, direct quotes render as styled quote cards, and rewritten reports are verified before completion.

Provider rails & failover Managed provider payment contract: Exa and Firecrawl run on MPP rails, with an OpenRouter API-credit fallback and Deepgram joining for audio.

Video, podcast & source analysis Deepgram transcription covers captionless videos, and failed transcript extraction is no longer fatal to a run.

Agent interop & the contract Agent Contract v1 rolled out clean-slate, with budget aliases honored.

Autonomous factory & operations Deploy stability runtime split and a documented safe production deploy workflow.

Report quality & evidence fidelity The writer now sees full ranked evidence (bestSources ordering, 500K-token budget) instead of a stripped corpus, and customer instructions are authoritative over library context.

Report quality & evidence fidelity The report writer was decomposed into a sectioned pipeline with adversarial testing, corpus merge into focused LLM passes, and every report now opens with a Signal block.

Pipeline reliability & speed Stage-based resume: continue skips re-planning and reuses corpus analysis instead of re-running the whole pipeline; extraction streams progress every 30 seconds.

Site, brand & public pages The pipeline page was redesigned stage by stage, public docs were redesigned, and this changelog page was born.

Report quality & evidence fidelity Corpus merge went hierarchical with shard retries, and rewritten reports are verified against the output contract.

Payments & metered billing Costs break down by pipeline stage and the minimum run budget dropped to one dollar.

Provider rails & failover OpenRouter MPP payment identity propagates across every LLM stage.

Payments & metered billing Customer sessions are sponsored from the treasury wallet, live billing and settlement were fixed, and run cost breakdowns shipped.

Codebase health & modularization The great split: hundreds of files broken into focused modules, with workflows and the engine promoted to the repo top level.

Pipeline reliability & speed Stop decisions are quality-gated, and budget-ended runs still complete with a report.

Video, podcast & source analysis Speaker labeling and media concept extraction became LLM-driven instead of heuristic.

Codebase health & modularization UI and runtime modularized into focused modules; the core workflow package became the engine.

Codebase health & modularization Dozens of duplicated value helpers consolidated into shared utilities.

Report quality & evidence fidelity Corpus quality overhaul — dedup, rescoring, boilerplate stripping, ranking — plus fixes for 100+ review findings across the codebase.

Provider rails & failover Content-filter refusals are detected and rerouted across Claude and GPT fallbacks with lenient JSON recovery; Exa gained find_similar, domain filters, and date ranges.

Video, podcast & source analysis Transcripts are LLM-restructured into interview format; YouTube handling is transcript-first, searching only for specific references.

Agent interop & the contract API overhaul: /research and /agent with personas and streaming, plus discovery surfaces — robots.txt, sitemap, and llms-full.txt.

Chat & the run workspace Global chat gained its sidebar everywhere, a stop button, and an iridescent focus sweep on the research input.

Video, podcast & source analysis Video reports became quote-first with timestamp links, and YouTube extraction starts straight from the home composer.

Video, podcast & source analysis From spec to shipped in a day: YouTube media runs with transcript-first reports, a transcript tab, and timestamped source references.

Provider rails & failover Exa search discovery joined Firecrawl, with costs accounted per run.

Chat & the run workspace Smoother analyst streaming, an animated sidebar, and a widened chat modal.

Agent interop & the contract An agent operating index, research contracts, and run controls for programmatic callers.

Chat & the run workspace Chat became a first-class surface: persistent start-chat sessions with shareable URLs, inline run proposals, and unified entrypoints.

Agent interop & the contract Account-linked Discord integration with customer Hermes agent routing and billing coverage.

Chat & the run workspace Run Analyst agents landed: chat with a finished run, with inline citations and durable structured answers.

Codebase health & modularization The service split into monorepo packages.

Report quality & evidence fidelity Benchmarked against OpenAI Deep Research with per-source comparisons and better eval corpus packing.

Agent interop & the contract A managed Discord research agent: direct messages, conversational chat, and persistent research memory.

Payments & metered billing Metered run billing moved to MPP sessions authorized from server wallets, with hard budget caps and full API key management.

Agent interop & the contract An agent handoff contract with funding guidance baked into 402 responses and capped planner latency.

Agent interop & the contract Share pages render for agents and run pages offer an agent handoff.

Report quality & evidence fidelity Contradiction detection landed: knowledge claims are auto-verified and contradictory facts get their own report section.

Report quality & evidence fidelity Fallback reports stopped padding themselves: boilerplate filtering, topic-specific claims, and irrelevant timelines and tension sections suppressed.

Report quality & evidence fidelity Reports learned to read as prose: plain-English sections, generic section schemas, cleaner tables, and a soft auto-reveal when a run finishes.

Pipeline reliability & speed Interrupted research jobs are requeued automatically.

Agent interop & the contract Agent setup and stored source access landed, plus an Obsidian export command.

Site, brand & public pages Homepage reframed around featured example runs and a six-step pipeline; budget-based pricing pages with a redesigned estimator.

Pipeline reliability & speed Continuation runs landed across the stack — a finished run can be extended instead of restarted.

Report quality & evidence fidelity Structured research plans replaced playbooks and the mode concept; runs honor academic-source requests.

Report quality & evidence fidelity Research loops became explicit and quality-gated, with an analyst audit pass over source extractions and library recall.

Pipeline reliability & speed Interrupted jobs recover on startup and research loops tolerate provider stalls.

Report quality & evidence fidelity First quality-first pass over the pipeline: evidence-aware quality gates, LLM candidate triage, analyst workplans, and stage-specific models.

Payments & metered billing Funded wallets are charged before any 402, and runs return watch URLs in account usage.

Provider rails & failover Provider-aware structured LLM client with OpenRouter timeouts, fallback planning models, and adaptive output budgets.

Payments & metered billing First customer payments: mppx payment challenges, account funding, and Privy key management.

Site, brand & public pages The service became Researcher, with an animated research harness diagram on the homepage.

Autonomous factory & operations v1 scaffolding: docs, CLI, and VM deploy bootstrap.