Skip to specification
harnessarch

Reading Room

Reading Room

This site is built out of other people’s work. The reading room is where that work is credited and examined rather than summarized — what a report claimed, what the people discussing it said back, and which of the two the evidence supports. Nothing here republishes a source; every entry quotes it, cites where the quotation came from, and links to the original.

Examined — entries with a sheet of their own

  1. Kimi K3: Open Frontier IntelligenceA 2.8-trillion-parameter open-weights model is a model story. What earns this report a sheet is how much of it is about the machine around the model — nine specialist policies distilled back into one, a reward model that writes its own rubric, and a reinforcement-learning loop trained against a swappable harness interface so the policy never overfits to a single one. Read against the four talks this site is built on, it says the harness is no longer only a runtime concern.Moonshot AI, 2026-07-28 · discussed by The Hugging Face research team · read into The HarnessCorr2Cont1Exte3Unre1
  2. AI-DLC: Redesigning the Lifecycle Around the ModelTwo vendors, six months apart, with no evident contact between them, open on the same controlled study: developers using AI tools believed they were 20% faster and were measured 20% slower. Both reach the same diagnosis — the coding gains are real, and every one of them is absorbed by the waiting around them — and both conclude that the fix is to redesign the lifecycle rather than bolt AI onto it. The second talk is not a room convened on the first. It is an independent arrival at the same terrain, which is why the examination below has no contested band, and why that absence is the finding.Anupam Mishra and Raja SP, AWS, 2025-12-04 · discussed by Cedric Clyburn, IBM · read into Agentic SDLCCorr5Cont0Exte3Unre1
  3. FDE: The Motion, and the Method UnderneathThe pillar this reading feeds stands on one talk, and that talk is about a market — which quadrant needs forward deployed engineers, what they cost, what they earn. It never says what the engineer does on the Tuesday after the contract is signed. The discussion here answers exactly that, and it is the weakest-provenance source this site has read: no named author, a channel compiling four talks it names but does not link, numbers relayed rather than cited. That is why it sits in this room instead of on a sheet. A record whose attribution is its worst property can still be examined — and six of the nine bands below are extended rather than corroborated, because the two sources barely touch. Where they do touch, they disagree once, and the disagreement is the useful part.Kevin Bai, Anthropic, 2026-08-07 · discussed by AI LABS · read into Forward DeployedCorr1Cont1Exte6Unre1
  4. System One: The Decisions Inside the LoopA model that answers typed questions about a state with calibrated probabilities and never writes a sentence — cheap enough to put a judgement at every gate in a harness. Drawn here first, because almost nobody has watched one work: the call, the seats it takes in a loop, and the place it breaks. The examination of the announcement, against the one post that carried it into a loop and measured what happened, is folded at the end. That post prints no byline; it is examined here and credited nowhere else.Diogo Almeida, founder, TypeSafe AI, 2026-09-15 · discussed by Octomind · read into The HarnessCorr0Cont1Exte5Unre1

Read into a pillar — the talks this site was built from

These 22 talks are the source material for the 4 pillars. Each is credited on the pillar it produced — or, where a talk goes deep enough, on the sheets it raised — and in the title block of every sheet inside it. One of them is also examined above, in the register: a talk can be both the material a pillar is built from and a source worth reading against another. (The Harness-Bench paper is credited on the Test Bench, the register it motivates, rather than here.)

RegisterReading room
Examined4 of 27
Standing on23 talks and 4 reports by other people