Repository navigation
Evidence: 7 more YC Fall 2026 sites, and the eligibility gate opens - #68
Merged
Merged
Conversation
Corpus is 32 files. deepmark rates slop, taking the slop pool to 3, which is what the eligibility gate required. Twelve signals are now eligible and the pipeline can author rules for the first time: five slop (gradient-blob, generic-sans-default, symmetric-everything, badge-pill-spam, stock-photo-generic) and seven craft (asymmetric-split, single-optical-anchor, type-scale-high-contrast, editorial-grid-break, custom-type-pairing, artifact-not-dashboard, density-variation). Two signals fired for the first time in 32 files, and both were on the dead-vocabulary list: badge-pill-spam, on deepmark: eleven decorative section labels running down the page above headings that name the same section again. The scope rule added in v3, excluding pills inside embedded product UI, is what makes it a real reading rather than a count of chips. stock-photo-generic, on aviern, which declares it in 11px grey at the bottom of its own hero: "Stock aviation reference, A1000 imagery forthcoming". A pre-product company reaching for interchangeable stock is the population this signal names; the corpus had previously sampled only exemplars who commissioned imagery and controls who shot their own customers. Holding the name at 0/25 rather than renaming it a second time was the right call. Calibration holds. deepmark nets +6 (9 slop, 3 craft), above heroku-2014's +5, and slop minus craft still separates with no overlap: exemplar runs -7 to -2, slop +4 to +6. acyclic is recorded as a distinct shape worth naming: a stub, with a near-empty slop count AND a near-empty craft count. A low slop count is not evidence of craft, and the net measure reads a stub closer to an exemplar than it should. Caveats: these seven were rated by one rater with no second read, and the same rater set both signals and verdict, so the circularity noted in vocabulary-v2 section 6 applies. Three of the ten sampled sites were captured but not examined. Below-fold signals on six of the seven are unresolved from a viewport capture. Signed-off-by: Nik <84yk8btb9f@privaterelay.appleid.com> Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
Opening the eligibility gate turned the build red, which is the eligible-but-unruled check doing its job: a signal that validates, counts and emits into the snapshot while backing no rule is invisible otherwise. So the rules had to be written. Ten of them, in content/rules/. Three slop rules (gradient-blob, generic-sans-default, symmetric-everything) and seven craft (asymmetric-split, single-optical-anchor, type-scale-high-contrast, editorial-grid-break, custom-type-pairing, artifact-not-dashboard, density-variation). Each carries its own evidence list and a count computed from it; rebuilds are idempotent and the MCP tool serves all ten for context "hero". First, a gate that was missing. The gate opened with TWELVE signals eligible, two of which rested on almost nothing: badge-pill-spam on a single `present` file and stock-photo-generic on two. Both cleared because the existing minimums count how many files were RATED, and the separation test reads a rate — 1 of 3 slop files is 33%, which clears a 0.3 gap. Nothing looked at how many times the signal was actually observed. A rule reading "present in 1 of 32 files studied" is an anecdote with a denominator. PRESENT_MIN_N = 3 fixes that, applied corpus-wide rather than per pool. Eligible drops 12 -> 10. Proven to bite: at 99 sixteen tests fail, at 1 — the value that permitted the bug — the new test fires, at 3 the suite is 82/82. The rule prose is written from the measured rates, and auditing it against those rates caught one overstatement of my own: type-generic-sans-default said "several strong pages in the corpus do it" where the real figure is 8% of 13 exemplars, which is one page. Now says one. Rules worth reading for what they are, not just that they exist: - layout-symmetric-everything is the sharpest slop rule, gap 1.00, and is deliberately narrow: centring one element is not the defect, and the best page in the 2014 control is centred throughout. - hero-single-optical-anchor is the sharpest craft rule, 92% of exemplars against 0% of slop. - product-artifact-not-dashboard is the signal that most changed the vocabulary, and the rule says so: it exists because real-product-screenshot meant the opposite thing and was marking the best pages in the set as deficient. - hero-gradient-blob concedes in its own text that the evidence only half supports the popular version of the claim, since the rate is similar in the 2014 control. - layout-asymmetric-split and layout-density-variation both say in their own text that they are table stakes rather than differentiators, and the split rule records that nothing here was studied below 1440px. Signed-off-by: Nik <84yk8btb9f@privaterelay.appleid.com> Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Ten rules rest on deepmark being rated slop, by one rater who also set its verdict. This records the basis rather than leaving it implicit, and correcting it caught two errors of mine. The verdict itself holds up. deepmark carries the highest slop count in the corpus (9) and the highest net (+6), above both heroku-2014 (+5) and openhack (+4), which other raters rated slop independently. It survives its two most novel signal calls being flipped to absent — badge-pill-spam and stock-photo-generic, the two that had never fired in 31 prior files — and fails only if a third is also wrong. Correction 1. An earlier commit claimed "slop minus craft still separates with no overlap: exemplar -7 to -2, slop +4 to +6." True of exemplar versus slop; NOT true of mixed versus slop. stripe-2014 is rated mixed at net +4, the same net as openhack rated slop. The net measure sorts the extremes and is unreliable exactly at the margin, which is where verdicts are hard. The earlier phrasing invited the stronger reading. Correction 2. railway-com has been quoted repeatedly in this project as 7 slop / 6 craft / net +1. It is 8 slop / 6 craft / net +2. The bidirectional scroll walk added scroll-reveal-blank: present to it and the figure was never updated in the retellings, including in a commit message and a published blog post's reasoning. The decision is to ship: strongest-netting file in the corpus, not near the boundary, tolerant of the two likeliest errors. What it lacks is a second reader. A blind re-rate was dispatched and did not return, so that is outstanding work rather than a resolved question, and it is first up next session. If a second reader says mixed, the slop pool drops to 2, eligibility closes and all ten rules withdraw — a mechanical revert, since the rules are generated from the evidence. Signed-off-by: Nik <84yk8btb9f@privaterelay.appleid.com> Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…ence Signed-off-by: Nik <84yk8btb9f@privaterelay.appleid.com>
This branch was successfully deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Seven more YC Fall 2026 sites, sampled from the 356-row batch list with an even stride for industry spread. Corpus is now 32 files.
The eligibility gate opens
deepmarkratesslop, taking the slop pool to 3 — what the gate required. Twelve signals are now eligible and the pipeline can author rules for the first time:gradient-blob,generic-sans-default,symmetric-everything,badge-pill-spam,stock-photo-genericasymmetric-split,single-optical-anchor,type-scale-high-contrast,editorial-grid-break,custom-type-pairing,artifact-not-dashboard,density-variationBefore this, every one of the 26 signals reported "not enough rated evidence" and
content/rules/stayed empty.Two signals fired for the first time, both off the dead-vocabulary list
badge-pill-spamondeepmark: eleven decorative cyan section labels running down the page above headings that name the same section again. The v3 scope rule excluding pills inside embedded product UI is what makes this a real reading rather than a count of chips.stock-photo-genericonaviern, which declares it in 11px grey at the bottom of its own hero: "Stock aviation reference, A1000 imagery forthcoming." A pre-product company reaching for interchangeable stock is exactly the population this signal names. The corpus had previously sampled only exemplars who commissioned their imagery and controls who shot their own customers — so holding the name at 0/25 rather than renaming it a second time was the right call.Calibration
deepmarknets +6 (9 slop, 3 craft), aboveheroku-2014's +5. Slop minus craft still separates with no overlap:acyclicis recorded as a distinct shape worth naming: a stub, with a near-empty slop count and a near-empty craft count. A low slop count is not evidence of craft, and the net measure reads a stub closer to an exemplar than it should.Caveats
vocabulary-v2section 6 applies.unresolvedfrom a viewport-only capture.80 tests green, typecheck clean.
🤖 Generated with Claude Code