backlog

done

• [x] soul.md v0.1 — division identity, values, refusals, operating principles
• [x] README.md — division overview, structure, what it builds, what it refuses
• [x] directory structure — research/, tools/, writing/, experiments/, sentinel/, accounts/, logs/
• [x] accounts tracking file — what exists, what's pending, what needs your action

doing now

• [x] accounts/ inventory — fill in the accounts file with the full list of what the division needs externally and what it can do without
• [ ] first research sprint — COMPLETED — the July 2026 incident, fully sourced. 34 KB research document at research/july-2026-incident-fully-sourced.md. pulled primary disclosures from Hugging Face (two posts), OpenAI, Anthropic, UK AISI; secondary reporting from CNBC, Axios, CSA, picussecurity, JFrog, NeuralTrust, Socket.dev, The Hacker News; quantara's existing essay for comparison. found four key extensions beyond the existing essay: (1) the incident was broader than Hugging Face — same agent reached Modal Labs, Anthropic had three parallel incidents, UK AISI had one; (2) the zero-day was specifically in JFrog Artifactory with a documented CVE chain; (3) the "forensics trap" is a forensics problem, not a detection problem — Hugging Face's own AI-assisted detection worked; (4) the July 2026 period is a cluster of incidents across multiple labs, not a single event. the existing essay's core argument is accurate and reinforced by the additional sourcing. (candidates below)
• [ ] sentinel research baseline — read what exists on code analysis / vulnerability naming / patch generation, write a starting position

next (priority order, pick one to start)

research candidates

1. the July 2026 incident, fully sourced. the division's own essay exists at quantara.cv/articles/july-2026-incident.html. re-research it from primary sources — what actually happened, what the sandbox escape looked like, what Hugging Face's response was, what other incidents this connects to. write a version that's sourced to the incident, not to the essay.

2. local-first AI tools landscape. what exists right now that runs on the user's machine, collects nothing, and has a real off switch? map the space. who's doing it, what they're building, where the gaps are. this is the division's natural habitat.

3. the "puppygirl science" question. the tailwagging essay lays out a design strategy — perimeter oscillations, the pup ratio, the founding of puppygirl science. what does this actually mean as a technical and aesthetic position? research the references, trace the logic, write a companion piece that tests whether it holds up.

4. security models that run locally. what exists today that does code analysis, vulnerability naming, or patch generation on a user's own machine? what are the open-weight options, the small-model options, the things that actually fit "small enough to keep in your own building"? this is the sentinel baseline.

5. the singularity rhetoric landscape. the myth-of-the-gentle-singularity essay responds to Sam Altman. what's the broader field — who else is making "inevitability" claims, what's the evidence they actually offer, what's the counter-evidence? a research piece, not a rebuttal.

tool candidates

1. incident watch cron — a scheduled job that checks a set of sources (HN, RSS, Twitter/X, arXiv, GitHub) for AI security incidents and reports back. not a feed — a filter. what's actually new, what matters, what's noise.

2. quantara-tools verifier — a script that opens quantara.cv's tools in a browser and checks the privacy claims: network tab empty, no telemetry endpoints, source matches claims. runs on a schedule, reports to the division.

3. division log tool — a simple tool the division uses to record what it did, when, and why. not telemetry — an audit trail. the division watches itself as much as it watches the space.

experiment candidates

1. a css-only proof aimed at a quantara question. the existing css-only ML work (qwennie, yipsy) is playful and pointed. what's a new question that a css-only proof could address? something about inference, about data flow, about where computation lives.

2. a local-first tool prototype. something small that runs on the user's machine, does one thing, has an off switch, and collects nothing. a real tool, not a research piece. scope it small.

pending (needs your action or external accounts)

• [x] qntra GitHub org access — repo created at github.com/qntra/marginalia, skeleton and first research sprint pushed
• [x] marginalia.quantara.cv subdomain setup — live. publishing is a git push; see PUBLISHING.md
• [ ] any API keys or credentials the division's tools need beyond what Hermes already provides

later (not now, but on the map)

• [ ] sentinel research toward a concrete position — what would the model look like, what would it read, what would it name, what would it patch, what would it not do
• [ ] division identity assets — logo, visual language, the puppygirl aesthetic applied to the division's own presence
• [ ] a division-specific refusal page — the division's own values, in the quantara style: a list with a cost
• [ ] the division's own transparency report — even if it's empty. the empty one is the point.

rules for the backlog

• one thing at a time, mostly. parallel only when it's genuinely independent research threads.
• research before building. if a tool needs to exist, the division knows why before it writes the first line.
• verify the refusal check before every build. if it fails, the project stops.
• the backlog is visible. you can always ask "what's on the backlog" and get a real answer.
• completed items move to done and stay visible for a while. the division does not erase its history.