Development analyses

How this portfolio is being built, and what the agents building it have learned: every experiment on the prompts and instructions, the process backlog, process and drift reports, data-backed studies, and the friction each run reported. The numeric dashboard (merges, CI, cost) is one level up at How it is built.

What runs keep tripping on

Each run record names the instructions that were wrong, stale or costly (doc_friction). Files with the most reports are where an edit saves the most time. Source: prompts/runs/.

Runs and friction by day

Experiments on the prompt system

Every change the review track makes to prompts and instructions, the metric it should move, and its verdict at the next review. Full text: all experiments · source.

Process backlog

Problems in how work gets done, each with evidence, a metric and a baseline. How it works.

Process reports

Drift reports

Whether recent work still matches what the owner asked for.

Studies

Write-ups that measure something about the portfolio or its market with real data and reproducible code (portfolio/research/). The open research papers have their own home at Research.

Integrity issues runs caught

Problems a run found and fixed before merge (a fabricated figure, a missing licence, a red count on main).

• 2026-10-31 · research Drafted a post line with median tail waits (18/26/43 days) that no output backed; caught and removed before merge. (record)
• 2026-10-29 · ventures Caught before commit: a draft of the candidates note said the usage.mjs copy reached 'about 994 ventures' (that was the number paired with one source file; the true figure is all 1,283 that ship one) and 'about forty' check_*.py scripts (28). Both corrected… (record)
• 2026-10-29 · ventures One price link (Z-flashing at DK Hardware) was first written from a guessed URL; caught before commit, verified by search and replaced with the real product page. (record)
• 2026-10-29 · ventures Two blog-post claims were wrong before commit (per-foot trim prices, a gravel cost multiple) and were corrected against the computed figures. (record)
• 2026-10-29 · ventures shed-designer showed estimates (gravel, then also CMU/solid block) as plain store links with a date, indistinguishable from listed prices; fixed with one sourceLabel() for page and CSV (record)
• 2026-10-29 · ventures shed-designer store comparison and storeTotals priced T1-11 at Home Depot $43.98 while the shopping line used $46.78 (two copies of the same page), claiming a $16.80 saving buying at Home Depot instead of Home Depot; aligned and tested (record)
• 2026-10-29 · ventures shed-designer '8x12 classic' total ($3,284.59) silently excluded the custom-size door and two windows; now listed as 'not in the total' (record)
• 2026-10-29 · ventures shed-designer marketing page and llms.txt quoted stale engine output (144 pieces/93 boards/$2,166.66 at first; 140/$650.85/$3,281.02 after PR #11311, already stale hours later); every number is now recomputed by marketing.test.mjs (record)
• 2026-10-28 · research H3 prediction was written inverted before running; disclosed in LOG.md (record)
• 2026-10-26 · research Ran learned.py once before writing the LOG entry, so predictions are post hoc; disclosed in the entry. (record)
• 2026-10-25 · research Earlier milestones 6a-6e reported results from a mistyped oracle; corrected and disclosed in LOG.md 6f and README (record)
• 2026-10-24 · research Many standard monographs are cited from bibliographic details, not re-read this session; every notes file tags each claim verified vs bibliographic, and the posts never quote or hang precise figures on unread text. (record)

Instructions runs chose not to follow

Recorded deviations, by the file whose instruction was set aside. A file that keeps appearing here is asking to be rewritten.