The build log is the product

1. ID

   EXP-001

   Logged

   2026-07-27

   Learned

   General-purpose LLM vs constrained architecture for compliance agents

   Question

   Can a general-purpose LLM draft MyInvois e-invoice payloads end-to-end?

   Hypothesis

   A generic LLM can draft MyInvois e-invoice payloads end-to-end.

   1. Result

      ✕Fail:

      It hallucinated a permit number on a real test invoice.

   Lesson

   We needed a constrained architecture — not a bigger model.

2. ID

   EXP-002

   Logged

   2026-07-27

   Shipped

   Monolithic vs edge-native architecture for clinic ops latency

   Question

   Does a single Node service handle Malaysian clinic peak-concurrency fine?

   Hypothesis

   A single Node service handles Malaysian clinic peak-concurrency fine.

   1. Result

      ✓Pass:480ms → 90ms p99

      Edge-native moved the slowest call from 480ms to 90ms p99 without rewriting the core.

   Lesson

   The risk lives in the lab, not in your production.

3. ID

   EXP-003

   Logged

   2026-07-27

   Learned

   Bahasa Malaysia web-novel tokenization for author royalties

   Question

   Does standard sentence-splitting handle Bahasa colloquial register?

   Hypothesis

   Standard sentence-splitting handles Bahasa colloquial register.

   1. Result

      ✕Fail:

      Code-switching broke it consistently.

   Lesson

   We needed a Malaysian-trained tokenizer, not an English-default one.

4. ID

   EXP-004

   Logged

   2026-07-27

   In Build

   Employment Act-aligned bilingual contract generation

   Question

   Can a single Anthropic-family model write bilingual Employment Act 1955 contracts aligned to EPF / SOCSO / EIS / PCB?

   Hypothesis

   A single Anthropic-family model can write bilingual Employment Act 1955 contracts aligned to EPF / SOCSO / EIS / PCB.

   1. Result so far

      In active build. The failure mode found so far is hallucinated clause numbers, not fluency.

   Lesson

   The thing to build against is hallucinated clause numbers, not fluency — that shapes the constrained-generation approach.

5. ID

   EXP-005

   Logged

   2026-07-31

   Shipped

   Approximate vs exact nearest-neighbor search for agent memory recall

   Question

   Do pgvector's default HNSW index settings reliably surface a note's own exact-name match?

   Hypothesis

   pgvector's default HNSW index settings reliably surface a note's own exact-name match.

   1. Result

      ✕Fail:

      A chunk containing the literal query text ranked outside the top 15 despite being the 3rd-closest chunk vault-wide by exact distance — the index's default search breadth was too narrow.

   Lesson

   Approximate search needs its recall checked against brute-force ground truth, not assumed.

6. ID

   EXP-006

   Logged

   2026-08-12

   Learned

   Forking an agent runtime vs a shared-instance backend for cross-system memory

   Question

   Does unifying memory across two AI agent systems require forking the agent runtime to add a Postgres-backed state layer?

   Hypothesis

   Unifying memory across two AI agent systems requires forking the agent runtime to add a Postgres-backed state layer.

   1. Result

      ✕Fail:

      Checked the upstream repo's actual commit velocity first — north of 200 commits/day in the exact files a fork would patch. A single shared memory instance, each system in its own database, gets the same operational simplicity with zero fork risk.

   Lesson

   Quantify the real maintenance cost of a fork before committing to one — compose beats fork more often than it looks like upfront.

7. ID

   EXP-007

   Logged

   2026-08-17

   Shipped

   SQLite vs a custom binary format for a bitemporal document store (LTF)

   Question

   Does building the claim-store file format on SQLite avoid months of custom binary-format engineering?

   Hypothesis

   Building on SQLite avoids months of custom binary-format engineering for a still-unvalidated idea.

   1. First result

      ✕Fail:+781%

      Benchmarked against real documents, the SQLite-backed version lost by up to 781% on disk size — even gzip-compressed — before a single index/table's fixed overhead was accounted for.

   2. Correction

      A schema-optimization pass cut the gap roughly in half; a full rewrite onto a custom binary encoding closed most of the rest.

   Lesson

   "Avoid premature engineering" is right advice for scope — it’s the wrong advice once disk footprint is the actual claim being tested. Benchmark the real constraint before picking the shortcut.

8. ID

   EXP-008

   Logged

   2026-08-17

   Learned

   A reported disk-savings "win" turned out to be measuring the wrong baseline

   Question

   Does LTF's binary format beat JSON on disk size for structured records?

   Hypothesis

   LTF's binary format beats JSON on disk size for structured records — first benchmark showed an 11% win.

   1. First result

      ✕Fail:11% (retracted)

      That win was measured against a pretty-printed JSON fixture, not a fair one. Against minified JSON, the format was actually larger — a wash, not a win.

   2. Correction

      Profiling where the bytes actually went — not guessing — found two real, fixable causes.

   3. Retest

      ✓Pass:-11.2%

      Correcting them turned the retracted win into a genuine, re-verified 11.2% reduction against the fair baseline.

   Lesson

   An unverified win is worse than a published loss — it just fails later, in front of someone else. Every efficiency number this project publishes now states the exact baseline it was measured against.

9. ID

   EXP-009

   Logged

   2026-08-29

   Shipped

   Shipping a browser game with no engine, under a 100 KB JS budget

   Question

   Does a café sim with autosave, layered audio, and a day-by-day narrative scheduler need a game engine?

   Hypothesis

   A café sim with autosave, layered audio, and a day-by-day narrative scheduler needs a game engine — and the engine alone breaks a mobile-first bundle budget.

   1. Result

      ✓Pass:~67 KB gzipped

      A hand-rolled Canvas2D renderer with DOM text and Howler audio shipped the whole game at ~67 KB gzipped, playable in an itch.io iframe with no engine at all.

   2. One real break

      ✕Fail:

      An absolute /assets/ path 404'd once the build was served from a subpath, blanking a drink icon inside the embed — every runtime path had to resolve through the deploy base, not the site root.

   Play on itch.io

   Lesson

   “Needs an engine” was a budget assumption, not a technical one; “works locally” is not “works embedded.”

10. ID

    EXP-010

    Logged

    2026-09-08

    In Build

    One Dart/Flutter source targeting itch.io, Devvit, macOS, and mobile for a real-time game

    Question

    Can a single Dart codebase on the Flutter engine target web, Devvit, macOS, and mobile for a real-time game without a per-platform rewrite?

    Hypothesis

    A single Dart codebase on the Flutter engine can target the web (itch.io), Reddit's Devvit, macOS, and mobile for a real-time martial-arts game — not just form-driven app UI — without a per-platform rewrite.

    1. Result so far

       The same source ships today as a browser build playable in an itch.io iframe; the Devvit, macOS, and mobile targets come off the same tree next.

    Play on itch.io

    Lesson

    Single-source multi-platform holds past app UI into an interactive game loop; the constraint lives in the engine, not in the number of platforms.

These are abbreviated public-facing entries. The full internal logs reference real test invoices, latency profiles, and tokenizer token mismatches — shared with serious prospects under NDA before any pilot.