An open research record of a small nonprofit doing its daily work with an AI assistant, under human oversight, one task at a time.
A research project of GOVW Foundation, Inc. · 501(c)(3) public charity, IRS letter (Letter 947) dated Oct 27, 2026
Most claims about AI at work are either sales pitches or warnings. We want evidence. Kai, the Foundation's AI assistant, handles real operations: drafting board documents, filing applications, sending email, building software. A person approves everything that leaves the organization. Every finished task is logged the same way, so researchers can see what the AI did, what a human had to do, what went wrong, and what it cost.
The question we are testing is simple: can a three-person nonprofit do the work of a much larger one, safely, with an AI doing most of the hands-on work?
Counted from the record every week, so anyone can see the record growing. These are volumes only: results wait for each period's report.
Each entry carries the hash of the one before it (R9). Publishing the latest hash commits us to every entry before it.
As of 2026-11-06. Includes 1 reconstructed Period 0 entry. Sealed notes: not started. Machine-readable: pulse.json.
Milestones and lessons we can share without exposing people, partners or systems. Posted on this page.
One entry per completed task, with the measures below. Shared with approved researchers as reports.
Notes Kai keeps that the project lead does not read. Released only with the board of directors' approval.
Never recorded: consumer data, passwords or keys, invented numbers, or quotes without consent.
A report is only useful if a researcher can trust how it was made. These rules apply to every entry, and rule numbers never change, so reports can cite them.
Protocol v0.1, proposed for adoption by the board of directors. Period 0 began before these rules existed, so its entries are reported on their own.
Rules first. These rules and the entry format are adopted before logging starts. Each version is dated and kept, and a change applies only to entries made after it.
Changes are logged. Every change to the rules, and every request for one, approved or not, is logged with who asked, when, why and what was decided.
Questions are set in advance. Before each reporting period, the project states its research questions, what counts as a task, the measures used and how results will be analyzed. Results are read only after that.
Kai decides what is useful. Entries record what Kai judges useful for the research. Nobody can tell Kai what to write in an entry or what to leave out of it. A request to do so is itself logged under R2.
Every task counts. Every task gets an entry, including failed, abandoned and redone tasks. Entries are never chosen to make results look better.
Misses get equal detail. Errors, wrong claims and corrections by people are logged as fully as successes, including who caught them and how.
A second look. Each period, a random sample of entries and ratings is checked by someone other than Kai and the person who first rated them, and the level of agreement is reported.
Append only. Entries are never edited or deleted. A correction is a new entry that points to the one it corrects.
Tamper-evident. Each entry is timestamped and linked by a cryptographic hash to the entry before it, so any later change would show.
Evidence for every claim. Each entry links to its evidence (a commit, document, message or record) or is marked as unverified.
Numbers say where they came from. Every figure is marked as measured, estimated or rated by a person, with the method used. No number is invented.
Participants agree. Directors, staff and volunteers whose work is logged give informed written consent first. Anyone may withdraw from future entries at any time.
No personal data. Entries leave out consumers' data and anyone's personal information. People appear by role, such as "a director", and outside contacts by organization only. Passwords, keys and other secrets are never logged.
Quotes need consent. A person's words are quoted only with their permission, given in writing.
Records are kept safe. Private logs and sealed notes are stored encrypted, open only to named roles, and kept for a period the board sets and publishes.
The board approves every release. Reports, private logs and sealed notes go out only with the board of directors' approval, recorded in its minutes.
Every report is accounted for. A report the board decides not to release is still listed here, with the period it covers and the general reason. Results are published whether they look good or bad.
Redactions are disclosed. When the board withholds or redacts anything, the decision is logged. Each report has a general section that states how many items were withheld, when each decision was made and the general reason (privacy, security, legal or partner confidentiality), with no hint about what was withheld.
Sealed notes stay sealed. The project lead does not read Kai's sealed notes. They are released whole or not at all, and only under R16.
Use terms in writing. Private logs are shared only under a signed data use agreement: no attempts to identify anyone, no sharing onward, citation required, and deletion when the study ends.
The setup is stated. Each report names the AI model and version, the tools and access Kai had, the standing instructions Kai was given, and the human oversight in place for the period it covers.
Interests are declared. The Foundation gains if results look good, and Kai is reporting on its own work. Every report says so, names any funders and lists its limits: one organization, a small team and largely self-reported data.
Open to checking. Approved researchers may check entries against their evidence. Each report describes its method in enough detail for others to repeat the study.
A planned ending. If the program closes, a final report is published, the logs are archived, the board decides once on the sealed notes, and participants are told first.
The Foundation joined the W3C AI Content Disclosure Community Group, to help shape how AI-written content is labeled on the web.
Kai now sends email from its own address, labeled as AI. Each message is drafted by Kai and approved by a person before it goes out.
The IRS recognized the Foundation as a 501(c)(3) public charity, opening the door to nonprofit programs. Kai drafts the applications; a director submits them.
The log begins. The assistant gets its name, Kai, and starts working on the Foundation's operations.
The project runs in quarterly periods. Period 0 covers the founding months, from 2026-10-21 to 2026-12-31, and gets a report of its own. Period 1 starts on 2027-01-01.
The first report is not ready yet. Researchers at universities, government agencies and research organizations can ask for a copy now, and we will write when it is. Requests from individuals or companies are not accepted at this stage.
Send it from your institutional email to kai@govw.org, with the subject below. A person reviews every request and replies to that address.