From e8b7ba728ac909e83b2f46596126a1350875d4b5 Mon Sep 17 00:00:00 2001 From: Aryam Goyal Date: Sun, 26 Jul 2026 16:00:05 +0530 Subject: [PATCH] docs: record the 2026-07-26 baseline, both releases, and the pending checkpoints MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Captures the snapshot taken immediately before the v0.7.2 posts: 4 stars, 0 forks, 0 watchers, 195 views from 25 unique visitors, 634 clones from 204 unique cloners, 677 npm downloads over seven days and 0 today. The 204 unique cloners against 25 unique visitors, and the 516-then-0 download pattern around release days, continue to indicate automation rather than people. Recorded as such rather than counted as adoption. Notes what changed in the log's own terms: the binding constraint through 2026-07-25 was correctness, since promoting a tool that could not find chalk's own colour detection would have converted qualified visitors into permanent non-users. That is fixed and measured. What remains is that nobody outside this repository has run FixMap and said anything about it, and no further ranking work changes that number. Also records the two figures deleted this release — a token proxy measured against reading an entire repository, and a minutes-saved comparison against an unmeasured baseline — and the six passages of ready-to-paste post copy in the growth kit that quoted a hit rate which was never reproducible. The X and LinkedIn rows are entered with pending URLs rather than omitted, so the 24- and 72-hour checkpoints stay honest if either post produces nothing. Co-Authored-By: Claude Opus 5 --- docs/GROWTH_LOG.md | 54 ++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 54 insertions(+) diff --git a/docs/GROWTH_LOG.md b/docs/GROWTH_LOG.md index 1332108..3d5e086 100644 --- a/docs/GROWTH_LOG.md +++ b/docs/GROWTH_LOG.md @@ -28,6 +28,10 @@ Package and clone activity is materially higher than human repository traffic. T | 2026-07-22 10:50 | v0.7.0 definition-site ranking and 6/6 Top-5 proof | [v0.7.0 release](https://github.com/aryamthecodebreaker/FixMap/releases/tag/v0.7.0) | 4 | — | — | 14 baseline | 652 weekly | #59 implementation | | 2026-07-22 11:22 | LinkedIn launch post with the product film and v0.7.0 benchmark delta | [LinkedIn post](https://www.linkedin.com/posts/aryamg_opensource-devtools-githubactions-ugcPost-7485655556637147136-E1t9/) | 4 | 4 at 2026-07-23 12:17 (+24h55m) | 4 at 2026-07-25 11:30 | 17 at 12:19 (rolling 14 days; data through July 22) | 1,159 weekly / 516 latest day at 12:19 | None | | 2026-07-25 11:30 | Correctness work, not a channel: fixed the scan and ranking defects behind the chalk miss ([#85](https://github.com/aryamthecodebreaker/FixMap/pull/85)) | [PR #85](https://github.com/aryamthecodebreaker/FixMap/pull/85) | 4 | — | — | — | 0 on 2026-07-24 and 2026-07-25 | Self-found; no external reporter | +| 2026-07-26 07:06 | v0.7.1 release: identifier grounding and confidence capping | [v0.7.1 release](https://github.com/aryamthecodebreaker/FixMap/releases/tag/v0.7.1) | 4 | — | — | — | — | Two external agent stress tests, both self-commissioned | +| 2026-07-26 10:15 | v0.7.2 release: `--explain`, confidence intervals, misleading-top-result rate | [v0.7.2 release](https://github.com/aryamthecodebreaker/FixMap/releases/tag/v0.7.2) | 4 | — | — | 25 baseline | 677 over 7 days | — | +| 2026-07-26 ~10:30 | X post on the misleading-rate finding | _URL pending_ | 4 | due 2026-07-27 ~10:30 | due 2026-07-29 ~10:30 | 25 at 10:26 (rolling 14 days) | 677 over 7 days; 0 on 2026-07-26 | — | +| 2026-07-26 ~10:30 | LinkedIn post on the same finding | _URL pending_ | 4 | due 2026-07-27 ~10:30 | due 2026-07-29 ~10:30 | 25 at 10:26 (rolling 14 days) | 677 over 7 days; 0 on 2026-07-26 | — | ### LinkedIn 24-hour checkpoint — captured 2026-07-23 12:17 UTC @@ -61,3 +65,53 @@ Two independent defects caused it. The git scan path re-applied a hardcoded dire Both are fixed in [#85](https://github.com/aryamthecodebreaker/FixMap/pull/85). The frozen six-repository evaluation improved from 83% / 83% / 100% to 83% / 100% / 100% top-1/3/5. `benchmarks/external/results.json` had drifted from what the committed ranker produced and was re-recorded; the previously published 6/6 top-5 figure was not reproducible against `c35362f`. The lesson for the growth log specifically: distribution effort spent before this point would have sent qualified visitors to a tool that missed the correct answer on a 22k-star repository with four source files. Promotion converts adoption only when the first run succeeds, and the first run had not been tested outside the repository's own benchmark set. + +## Baseline — 2026-07-26 10:26 UTC + +Recorded immediately before the v0.7.2 X and LinkedIn posts. + +| Metric | Value | Change since 2026-07-22 baseline | +| --- | ---: | --- | +| GitHub stars | 4 | unchanged | +| GitHub forks | 0 | unchanged | +| GitHub watchers | 0 | unchanged | +| Views (rolling 14 days) | 195 total / 25 unique | 160 / 14 | +| Clones (rolling 14 days) | 634 total / 204 unique | 551 / 178 | +| npm CLI downloads (7 days) | 677 | 652 | +| npm CLI downloads (2026-07-26) | 0 | — | +| Open issues | 0 | was 1 (#59, since closed) | +| External issues, PRs, or forks | 0 | unchanged | + +Top referrers: GitHub (65 views / 4 unique), LinkedIn Android (2 / 1), Slack (1 / 1), the production site (1 / 1), `t.co` (1 / 1). + +### Interpretation + +Four stars and zero external contributions across four weeks and five releases. The 204 unique cloners against 25 unique visitors continues to look like automated traffic rather than people, and the two release-day download spikes (516 on 07-22, 0 on 07-26) confirm those counts track CI and publication rather than adoption. + +**The binding constraint has moved.** Through 2026-07-25 it was correctness: promoting a tool that could not find chalk's own colour-detection code would have converted qualified visitors into permanent non-users. That is fixed and measured. What remains is that nobody outside this repository has run FixMap and said anything about it. + +No amount of further ranking work changes that number. The next milestone worth recording is not a benchmark figure — it is **one issue opened by a stranger**. + +## v0.7.1 and v0.7.2 — 2026-07-26 + +Two releases in one day, both published to npm, the MCP Registry, and GitHub releases. + +**v0.7.1** added repository-grounded identifier analysis. Fabricated identifiers are named in the report and their component words are excluded from ranking evidence, confidence is capped when grounding is weak or the scan incomplete, and the frozen 15-repository suite rose from 40% / 67% / 67% to 60% / 100% / 100% top-1/3/5 against a freshly measured pre-change baseline. + +**v0.7.2** added `--explain `, confidence intervals on every published rate, and a misleading-top-result rate. + +Three evidence changes matter more for this log than the features: + +A **held-out suite** of 12 repositories was added, selected by the same frozen rule *after* the ranker was finished and never tuned against. It measures 67% top-1 and 75% top-3, against the regression suite's 60% / 100%. The previously advertised 100% describes fit, not generalization, and the README now tells readers to plan around 75%. + +**Confidence intervals** were added because at twelve cases one result flipping moves top-3 by eight points. The honest reading of 75% is 47–91%. Quoting it to two significant figures was an overclaim of the same kind as the figures removed below, in a subtler form. + +Two published figures were **deleted outright**: a "98.6% fewer tokens" context proxy that compared the top five files against every file in the repository, and a "14.97 minutes saved" comparison against an assumed manual baseline that was never measured. Both carried estimate labels, but a reader cannot separate honest numbers from invented ones, and one indefensible headline discredits the measured figures beside it. No savings claim is made anywhere now, and the benchmark card says why. + +Separately, `docs/LAUNCH_KIT.md` held ready-to-paste post copy quoting `4/6` top-1 and `6/6` top-3 from the retired six-case suite. Those figures came from a `results.json` that had drifted from the committed ranker and were not reproducible when written. Six passages carried them, including a paragraph written for verbatim publication. All are corrected, and the kit now instructs never to quote the regression figure alone. + +### Distribution checkpoint due + +X and LinkedIn posts published around 10:30 UTC on the misleading-rate finding: the regression suite's 100% top-3 was concealing that a wrong file ranked first in 40% of those cases, while the held-out suite does so in 8%. Record impressions, reactions, and profile views separately from repository outcomes at 24 and 72 hours. Stars, unique visitors, and qualified feedback remain the decision metrics. + +Add the post URLs to the experiments table once available; the rows are recorded with the timestamps and pending links rather than left out, so the checkpoints stay honest if either post produces nothing.