Source Included

Build log

What shipped in Source Included and the systems behind it, dated, with the numbers.

Follow by RSS/build-log/rss.xml

Entries
18
Latest
First entry

Filter by tag

  1. Gap Radar 4.0 samples every market before it buys depth

    Gap Radar now runs once a week and opens each market with a sample: it reads a slice of the complaints, ranks the problems it finds and buys deep review data for the top 3 only. When the model server fails, the call is retried with longer waits, the relevance check passes to the panel's other judge, and the two judges work side by side instead of in turn.

    Read the rest of the entryShow less

    A review-site listing whose name does not match the product it was drafted for is dropped before any data is bought. In the first weekly run a server error left 0 of 759 documents judged in the first market; on the rerun the model judged 470 of 471, and its own test suite ran 658 checks, all offline.

    Problems per market researched in depth
    3
    Documents judged on the rerun, 0 of 759 before
    470 of 471
    Offline checks
    658
  2. Gap Radar 3.1.0 checks every quote, and no idea from the run survives it

    Gap Radar 3.1.0 labels every quote on a card supports, neutral or contradicts against that card's own problem: only supporting quotes count, contradicting ones are shown as counter-evidence, and write-ups and feature requests are labelled second-hand and never count as voices.

    Read the rest of the entryShow less

    Pain now reads review ratings and words, a red-team sentence about a reviewer must be backed by the review it cites, and a pain-point label must read as a job to be done. Re-run on the same 19 Nov data, both cards that 3.0.0 published dropped to leads, Feed Drop Sentinel with 4 first-hand voices and Flat Rent Billing with 3, where an idea needs 5; the daily run's Books Bridge was rejected for too few first-hand voices, 2. Nothing was tuned to keep a card published, and its own test suite ran 515 checks, all offline.

    Later that day: the weekly run replaced the daily one as Gap Radar's default, one larger run a week per niche. How Gap Radar runs now

    Ideas published from the re-checked run, 2 before
    0
    Quote labels re-checked by the audit, 1 counter-evidence
    49
    Red-team sentences dropped, no cited review backed them
    4
    Offline checks
    515
  3. Gap Radar 3.0.0 checks every claim before a card goes out

    Gap Radar 3.0.0 adds an audit layer: every document and number goes into an evidence ledger, every sentence a model writes is split into claims and checked in code against it, a second model verifies what code cannot, and the panel verdict decides whether a problem is published, kept as a lead or rejected. The command node cli.mjs audit re-checks a finished run offline, with no network and no model.

    Read the rest of the entryShow less

    Its first validated run, on small accounting firms and small landlords, published 2 cards, kept 4 problems as leads and rejected 2, and the audit passed with 1,135 cited ids, 0 unresolved and 0 mismatches. On the same 23 launches the backtest AUC is now 0.848, and its own test suite ran 503 checks, all offline.

    Later that day: Gap Radar 3.1.0 re-checked the same run and published 0 ideas; both of these cards became leads. Gap Radar 3.1.0

    Backtest AUC, 0.826 at calibration
    0.848
    Cards published by the validated run
    2
    Cited ids, 0 unresolved
    1,135
    Offline checks
    503
  4. Retention periods now run in the database

    The periods in the privacy policy are now enforced by the database itself. A daily job at 03:15 UTC deletes unconfirmed sign-ups after 7 days, abuse and security logs after 90 days, what is left of a closed account after 30 days, and consent records 3 years after withdrawal or closure. Closing an account now removes its data at once instead of failing on linked rows. A test fails if the policy text and the database periods ever disagree.

    Unconfirmed sign-ups
    7 days
    Abuse and security logs
    90 days
    Closed accounts
    30 days
    Proof of consent
    3 years
  5. Campaign tags from the link you arrived on used to be kept in a cookie. They are now held in the open page's memory until you sign up, so nothing is written to your device and they are gone when you reload. The only cookies left are the sign-in cookies, which are strictly necessary.

  6. Source Included is a trading name of NEXFLOW DIGITAL S.R.L., checked against the ANAF public registry. The privacy policy lists the six processors the code calls, where each one keeps data and the transfer safeguard it relies on. The terms choose Romanian law, keep the protection of the rules where you live, and point Romanian consumers to ANPC's dispute resolution service.

    Processors, each with its safeguard
    6
  7. Gap Radar scoring calibrated on 23 real launches

    Gap Radar scored 23 historical problems from posts dated in the 2 years before each launch: 12 products that reached revenue and 11 that failed. Scoring gr-score-2.1.0 stops reading a missing keyword as a missing problem and stops capping direct counts at medium confidence, which raised the winner-over-loser AUC from 0.674 to 0.826. The 7 best-ranked cases are all winners, no weight was changed, and a test fails if the AUC drops under 0.776.

    AUC, up from 0.674
    0.826
    Launches scored
    23
    Top-ranked cases are winners
    7 of 7
  8. Gap Radar 2.0 scores every idea on 76 criteria

    Every idea is now scored on 76 criteria in 10 families, from Problem and Market to Founder fit and Evidence quality, and every number carries its origin: data, proxy, AI, user or prior. An AI analyst runs 3 passes per problem with your own key or your Claude Code subscription under a hard cost cap, and keeps the median. Without one, the run falls back to heuristic mode and every card says so. By the end of the day its own test suite ran 500 checks, all offline.

    Criteria
    76
    Families
    10
    Offline checks
    500
  9. Gap Radar ideas stay off the site until a run is validated

    The site shows no idea from a Gap Radar run until that run is validated. One switch decides it, and a test checks the public pages against every card name and quote from the held run. Until then the Gap Radar page traces two finished historical cases instead, Plausible Analytics and Quibi, from the first complaint to the score.

    Since 20 Nov 2026: a validated run is on the site, and the Gap Radar page traces its top-ranked lead step by step. Plausible Analytics and Quibi stay in the backtest. The traced lead

  10. Every lesson in the community rewritten

    All 114 lessons and 11 posts in the Skool group were rewritten, with 134 visuals made for them. The visuals are rendered by script with a 14 px floor on type, so they stay readable on a phone.

    Lessons
    114
    Visuals
    134
  11. Security review: 12 findings fixed, each with a test

    Two independent passes reviewed the Vault against the OWASP Top 10, with live probes against the local database. Twelve findings were fixed, each with its own regression test, among them a Content-Security-Policy, HSTS for two years, session cookies that scripts cannot read, and member guides closed to the public database API. Six items that need an account setting or a decision are tracked for launch, and none of them is critical.

    Findings fixed, with tests
    12
    Areas checked and sound
    14
    Open items, none critical
    6
  12. Gap Radar joins as system 01

    Gap Radar reads public complaints, reviews and forum posts in your niches and writes 3 ideas a niche every morning: a venture, a part-time product and a side hustle, each backed by 3 or more linked, dated quotes. It runs on Node 22 with zero npm dependencies and free sources, so no key is needed. Its first real run found Apple's review feed returning empty for hours, and the report now says so instead of hiding it.

    Since 20 Nov 2026: Gap Radar runs once a week, and a run at full depth pays for model calls and the review data it buys. What it costs to run

    A round instrument scope in a brushed steel bezel with a lime sweep line and a few lime points
    Self-checks at release, 658 today
    122
    npm dependencies
    0
  13. Eight systems staged under FSL-1.1-MIT

    Eight repositories were staged, one per system, each with its licence file, a README with its production card, a setup guide and an .env.example with empty values only, so you bring your own keys. Every one passed its own test suite the same day. The repositories open at launch.

    Repositories
    8
    Keys shipped
    0
  14. The Vault: sign-in, plans, scans and Ask Adrian's brain

    The Vault was built and tested: sign-in by emailed code, a personal 30-day plan, downloads, votes, review requests, a website scan and Ask Adrian's brain. Brain answers are labelled as AI-generated under Article 50 of the EU AI Act, and its input screen blocked all 150 red-team prompts in testing. The database, the server functions and the scan service run in Frankfurt.

    Red-team prompts blocked
    150 of 150
  15. A story page and real output on the site

    A new story page sets out the record behind the systems: a dated timeline, verbatim press quotes and a coverage table with 8 outlets. On the home page, a terminal shows the first commands of all 8 repositories, copied from their READMEs.

    Since 20 Nov 2026: the home terminal shows one system at a time, the one you point at or tab to, and each system page carries its own commands. The systems on home

  16. Eight outlets cover the 3Descu quote server

    Between 21 and 25 Oct 2026, eight outlets covered the 3Descu server that lets ChatGPT and Claude price a 3D printed part, among them Profit.ro, Observator News, Romania Journal, 3Druck.com and 3D Printing Industry. The MCP pack in the Vault is a clean rebuild of the same pattern for any business, with no 3Descu pricing code in it.

    Outlets in 5 days
    8