Built in Bengaluru. Assembled in public.

Pick one thesis. Never repeat last week.

A language model has no memory of what it wrote three weeks ago, so the pipeline gives it one. Every published issue is embedded (all-MiniLM-L6-v2, 384 dimensions) in a local Qdrant database.

Claude Sonnet 5.5 picks one specific, checkable thesis plus the four or five biggest stories of the last two days. If the pick scores above 0.85 cosine similarity to a past issue, it gets one chance to re-pick. A second near-repeat blocks the run for a human.

The abdominal plates slide into place and the flight pack docks at the spine.

Read the article, not the headline.

Models invent details when they only see a title and a snippet. Stage 3 fetches the full text of up to nine pages (as much as 18,000 characters each) and has Gemini 3.8 Flash extract atomic facts: numbers, names, dates, exact quotes.

Code then holds the extraction to account. Quotes are checked against the page text, claims it cannot find are dropped, and the run needs at least 10 verified claims to continue.

Heavy shin greaves and boot thrusters clamp on from below.

Every factual sentence carries a receipt.

Claude Sonnet 5.5 writes the issue: the week's biggest stories first, then the thesis, aiming for 1,500 words. It is a tagged writer, so each factual sentence has to point at a recorded fact, and quotes are pasted in by code rather than retyped by the model.

The chestplate seals over the ribs and the arc reactor ignites. It is the only part of the suit that glows this hard, because this is the only stage that creates text.

Recent issues have cost roughly $0.16 to $0.23 in the stages counted so far (the two agent stages are not in that figure yet), and a hard cap of $1.00 per issue stops the run if it is exceeded.

Two judges, neither of them the writer's family.

The scan checks each sentence against the facts it cites. Code rules go first: numbers, quotes and names have to match. Then two judges from non-Anthropic model families (GLM 5.3 Flash and gpt-oss-120b) read sentence and source side by side, and a third check asks whether the source attaches a caveat the sentence dropped.

Anything not passed by both judges is held for a human. Amber scan lines sweep the armour while the gauntlets and palm repulsors come online.

Tested on 150 deliberately planted errors, the new checker caught or held 149. The previous production checks caught 134.

Flagged claims get a second source or a rewrite.

An agent on DeepSeek V4.1 Flash takes each flagged claim and searches the web for independent confirmation.

A rejected sentence gets one rewrite and one re-check. If it still fails, the run blocks and a human decides. Sentences that cannot be fixed are kept and held, not silently deleted.

Thigh plates slide on with gold seams, the repair marks left visible on purpose.

Re-research it from scratch, without the brief.

A dry run produced a draft saying the model 'wrote to and modified' something that the sources describe as unauthorised access. Every earlier check let it through, because the sentence matched its source loosely.

So stage 6.5 re-researches the core claims from scratch, without ever seeing the evidence brief, and flags overstated framing. It runs on GLM 5.3 Flash, a different family from the validator, with the same one-fix, one-recheck rule.

The helmet shell drops onto the neck, faceplate still open. A thin cyan ring circles the head: the second pair of eyes.

A judge from a different family reads the whole draft cold.

Gemini 3.8 Flash reads the finished draft and returns pass or fail with its reasons. It also checks the word count against the 1,300 to 1,700 range, and a miss needs a stated reason. A fail blocks the run; there is no partial credit. GPT-5.4 is the fallback judge if a call fails.

Two rings spin around the torso, one for each opinion, and the pauldrons drop onto the shoulders.

The machine stops at a draft. A human presses send.

On Monday, Wednesday and Friday at 10:00 IST the pipeline starts itself. If every stage passes, it creates a Substack draft with a subtitle and three Notes, and sends Aseem a Telegram summary. It cannot publish. Aseem reads the draft and presses send.

The golden faceplate slams shut, the visor ignites and the boots fire. That is the whole suit, and the only manual step left.

Two things here are still unfinished. The Long Game essay pipeline is paused, and the video pipeline is a work in progress that is not part of this suit.