Product quality
Recommendation quality and the filter pipeline, checked daily against what streamers actually saved, skipped, and streamed.
Signals to signed decisions
The operating loop that runs StreamGist
One person operates this product, so the operations layer is software: automated watchers, a daily brief, and a weekly decision step a human has to sign. Below is the loop itself, live, and the scoreboard of claims it has proven, disproven, or still has on the clock.
- Surfaces monitored48 of 48 healthy
- Daily brief ranJuly 4, 2026
- Claims under test7
- Next verdict dueAugust 12, 2026
- Last recorded decisionJuly 4, 2026
These values are read from the running system’s status feed and refresh on their own; nobody edits them by hand.
The loop
Five stages, closed weekly
Facts flow one way and come back as decisions. No stage trusts the one before it to have been right.
Watch
A small fleet of automated watchers collects raw facts around the clock: product signals, pipeline health, search telemetry, ad spend, and the bank feed. Each watcher is small, boring, and replaceable.
Brief
Every morning one briefing compresses the overnight checks into a tiered read. Deterministic rules set severity before any model sees the data. AI narrates; it does not judge.
Compress
A weekly recap distills the dailies into department trends. Every two weeks a strategic pass scores its own earlier flags against what actually happened, so advice that aged badly stays on the record.
Decide
Once a week, everything awaiting a decision is collected into a packet: fired verdicts, regressions, work that never shipped. The founder's merge is the decision, dated and versioned, not a memory.
Ship and measure
Decisions change the product or the rules. Each deploy gets a frozen before-and-after read on a fixed clock, and that outcome rides into the next proposal whether it flatters the change or not.
Then the watchers pick up whatever the decision changed, and the loop begins again.
The same boundary governs the product
The user-facing recommendation path makes zero AI calls. Models draft and explain, deterministic rules decide, and a person signs anything that changes the rules. The full product architecture is in the AI transparency report.
Coverage
What the watchers cover
Capability map, not a topology. Each group is several small agents with one job each.
Data collection
Hourly activity snapshots and game research enrichment, the raw material everything downstream depends on.
Spend and cash
Ad platforms and the bank feed, pulled daily and reconciled against internal ledgers, so a spend surprise surfaces in a day rather than a quarter.
Search and discovery
Both major search engines' view of the site: what ranks, what is indexed, and what quietly dropped.
Content
A weekly research-grounded article pipeline with a quality gate. Claims the data cannot substantiate do not ship.
The decision layer
The briefing, recap, strategy, and adjudication agents that turn all of the above into signed decisions.
The scoreboard
Claims this system has tested
Serious claims get their pass and fail bars registered before the data arrives, and a fired verdict freezes. What follows is the public edition of that ledger. Failures lead, because killing a comfortable claim is the point of having bars.
What failed3
DisprovenSettled May 29, 2026
Picking a hotter game is, by itself, enough to grow a small channel's audience.
- How it was judged
- Within-streamer comparison on live channels: each streamer's realized viewers on their higher-ranked games against their own results on lower-ranked ones, so channel size cancels out.
- What happened
- The median difference was a coin flip, so growth promises are banned from every StreamGist surface. Recommendations are framed as fit and odds, where the evidence actually holds: whether a stream is watchable, sustainable, and in front of the audience it suits.
DisprovenSettled June 12, 2026
Panel save and skip behavior forecasts which games will gain or lose Twitch creators next month.
- How it was judged
- Backtested against observed creator adoption with placebo and permutation checks, judged at a bar set before the run.
- What happened
- It does not, at our current panel density. The partner-facing read is framed as intent diagnosis, how creators receive a game on contact, and never as a market forecast.
DisprovenSettled May 29, 2026
Alert and reminder emails bring streamers back to the product.
- How it was judged
- Four separate re-engagement mechanisms were shipped and measured against real demand, the latest against a pre-registered return-rate bar.
- What happened
- Demand was near zero all four times, so the features were turned off. The saved-games surface works as pull instead: ranked options a streamer checks when the decision is in front of them, with no reminder machinery.
What held2
ProvenSettled June 11, 2026
A save predicts the saver's own next stream.
- How it was judged
- Six months of first-party save and skip decisions, checked against what each streamer actually went live with, plus placebo and permutation checks.
- What happened
- Streamers went live with saved games at about 26 times the rate of skipped games within 30 days, with a median gap of three days. That signal became the backbone of the product's feedback loop and of the partner demand read.
ProvenSettled June 11, 2026
More deliberate saves convert to streams at a higher rate.
- How it was judged
- Stated intent captured at save time, checked against actual streaming follow-through.
- What happened
- The gradient is monotone: the stronger the stated intent, the likelier the stream. Intent capture was re-enabled on a sampled basis, instrumented so its own friction stays measurable.
What is on the clock2
Still openIn replication since June 15, 2026
The composite score we ship is associated with small channels beating their own baseline after adopting a game.
- How it was judged
- A favorable retrospective result exists, within the tracked catalog. It is embargoed from all marketing until a prospective replication passes a bar that was fixed before the cohort opened.
- What happened
- Nothing has changed yet, deliberately. If the replication fails, the claim dies and this ledger will say so.
Still openVerdict due September 15, 2026
Search-acquired signups behave like qualified users, not empty clicks.
- How it was judged
- Pass and fail bars were registered on July 3, 2026, before the cohort existed. The verdict ships whichever way it lands.
- What happened
- On the clock. Come back after the verdict date and this row will have an answer.
Rows are curated from the internal claims registry under the same language rules that govern every StreamGist surface: no growth promises, no forecasts the data cannot back. The counts in the chips above track the live registry, so a new adjudication updates this page without anyone touching it.
Adjudication
Decisions are commits
The machine cannot approve its own conclusions, and the human cannot quietly ignore them. Here is what happens when a verdict fires.
A bar fires. A pre-registered claim crosses its pass or fail bar and the verdict freezes. Nothing can soften it after the fact.
The packet drafts itself. The machine collects everything due and drafts a decision memo for each item.
A human signs. The founder merges the packet, or edits it first. Merging is the decision; closing it without deciding is not an option the system offers quietly.
The record moves. The registry closes the item, the strategic context gets a dated entry, and the counts on this page follow automatically.
FAQ
Operating loop questions
What is an operating loop?
A closed control loop for running a product: automated watchers collect facts, briefings compress them, and a scheduled adjudication step forces a recorded human decision. The difference from a dashboard is the forced decision at the end.
Why pre-register the pass and fail bars?
A bar chosen after the data arrives always flatters the data. Fixing the bar first is what makes a verdict mean something, including the verdicts that killed our own features.
Is the AI making the decisions?
No. Models summarize, draft, and narrate. Severity tiers are computed by deterministic rules before a model sees anything, and every change to the product or its rules is signed by a human.
Is this page live or a mockup?
The chips at the top are read from the running system's public status feed and cache for a few minutes. Ledger counts track the deployed experiment registry, so an adjudication moves them without anyone editing this page. A fuller operator cockpit exists behind authentication for guided reviews.
Who built this?
Matthew Juszczyk, StreamGist's founder, as the operations layer for a product run by one person. The execution systems page describes the practice; the AI transparency report covers where AI runs in the product itself.
Keep reading
The rest of the picture
Execution systems
The practice behind this loop: turning recurring signals into owned action, built and run by the founder.
AI transparency report
Where AI runs in the product, where it never does, and what the whole thing costs.
How recommendations work
The product logic itself: category opportunity, streamability, fit, and the feedback loop.