Conversation
… pays for it Found while measuring long-input cost for the character-cap decision; the cost question turned out to be the smaller problem. Ten real pipeline runs on the production prompt: a 146-character number-bearing smoke and a 900-character digit-free document pass on the first attempt, while every probe from 1,270 characters upward terminates in number_safety_failed with numeric_claim_changed, including three coherent single documents at 1,270-1,624 characters. One diagnosed case had the original carrying claim number:1/1 and the rewrite carrying none — the rewriter dropped a number-word and the gate refused, which is the gate working as designed. The cost falls on the customer. rewrite-handler states that the daily and monthly counters increment fail-closed with no refund path, so a failed Pro request still spends one of 100 monthly requests plus its characters against the 50,000 total and returns an error, while both rewrite attempts bill us about $0.09. A customer drafting number-bearing text can pay for a month and receive nothing, against an account Polar holds to a 0.4% chargeback rate — and the card offers 'up to 20,000 characters each'. Records the confounds rather than hiding them: the digit-free probes are concatenated fragments whose topic jumps may provoke compression, the coherent documents are patina's own unusually number-dense examples, and every row is n=1 for its size, so nothing here establishes a rate. The four available responses each trade away something real — refund policy vs abuse, retry budget vs margin, advertised ceiling vs Pro's reason to exist, or changing the rewrite prompt — so the choice is left open rather than made unilaterally.
…asoning cut The scoring reasoning cut is scoped to request.provider === 'gemini'. The first pass of this measurement never set that field, so it silently served the pre-cut path and inflated the scorers. Isolated on the same 518-character input: scorer thinking 923/497 and $0.0573 with the field unset, versus 0/0 and $0.0311 with it set. Two corrections follow. The production-path cost of a 518-character English request is $0.0311, about $3.11 of COGS for a full 100-request month against $8.49 net revenue — roughly 63% margin, better than the ~55% on record rather than worse. An earlier draft read $0.05-0.058 as a margin problem; it was measuring a path production does not use. Also drops an unmeasured number: the note attributed 'roughly $0.09' to a 2,700-character document, which was the 19,540-character probe's figure transposed onto a size never priced. Replaced with the measured $0.069-0.089 range and scoped to the probe it came from. The number-safety findings are unaffected: failures occur at the rewrite stage before scoring, and the rewrite call is excluded from the cut by design.
…fact Six production runs on customer-shaped drafts, written as a product-update post with ordinary numeric density, all terminate in done with every numeral preserved: 1,190 characters with 6 numerals three times, and 2,400 characters with 14 numerals three times. The gate does not fire on this material at either length. The earlier note claimed number-bearing documents fail the gate and framed it as live customer harm. Both failing probe families were unlike customer prose: patina's own examples/*.md are before/after documents dense with scores, percentages and version strings, and the digit-free probes were unrelated paragraphs concatenated to hit a length target, where topic jumps invite the compression that drops a claim. Two families failing looked like corroboration when they only shared the property of being unrepresentative, and the obvious control was listed as future work instead of being run first. Retraction kept in place rather than deleted, with the raw failing rows retained, so the reasoning error stays visible. Withdraws the recommendation to reduce the rewriter's numeral drift: on customer-shaped input there is no measured drift to reduce, and changing the rewrite prompt on this basis would be acting against an artifact. What survives: rewrite-handler charges the monthly counters with no refund path, so when the gate does fire a Pro customer loses quota and receives an error. Wrong on its own terms, but a correctness issue rather than an emergency at this trigger rate, and a refund path still needs a bound against crafted always-failing input.
The Pro monthly cap could not be exercised directly — that needs a production license, and the verification key was shredded after use while a confirmed checkout session does not re-issue a customer token. Rather than leaving the whole claim unevidenced, the layer underneath it was tested: the free tier's hourly burst limit fired on production with 429 hourly burst exceeded after ten requests in the window. That matters because the Pro monthly cap runs through the same limiter, the same KV, and the same HMAC-keyed bucket, differing only in key and window. So the counters demonstrably increment and refuse in the deployed environment, which was the open question. The 100-request threshold itself remains proven only against a local handler, and the note now says exactly that rather than listing the cap as wholly unverified.
Retain the iterative quality baseline under a neutral research-only name, migrate scoring and verification config ownership, default the playground to English, and synchronize release metadata to 7.0.0.
Fix rewrite output, config parity, and offline scoring
Document resolved CLI issues
…o Polar The LS account was declined (payment-provider-reset-20260729), so the hold was certifying a dead route. The binding table now carries exactly the Polar production tuple; the hold validates the PAY-B-BINDING-POLAR-v1 and PAY-LIVE-RUNTIME-POLAR-v1 artifacts; POLAR_APPROVAL replaces LS_APPROVAL and the secret blocker names the polar gate variables. The staging tier is recorded as SUPERSEDED by the production zero-amount purchase evidence per the owner's decision. LS evidence stays on disk, hash-frozen, as history. Checkout stays disabled; enabling remains the owner's env-side act.
…unbook, reconcile Gate B Review findings from the independent lane on #668: the migration doc's step 4 now names every null-evidence blocker that must clear before the checkout flag flips; owner-actions-go-live.md is marked TERMINAL so the retired LS identities cannot be provisioned from a current-looking runbook; Gate B in pro-launch.md accepts the recorded staging supersession for the Polar route; the unpinned Polar price ID is documented as a known limitation to close at the next read-back artifact.
ops: rewrite the v6.4 hold and checkout bindings from Lemon Squeezy to Polar
Polar cleared all three review states on 2026-08-03 (account approved, payout connected, identity verified — owner-observed dashboard). The production env carries every required secret name; only the live-open pair (PATINA_PRO_GATE_EVIDENCE_ID, PATINA_SYNTHETIC_PRO_LICENSE) is deliberately absent, and PATINA_PRO_ALLOW_FREE_KEY is exactly false.
ops: POLAR_APPROVAL + SECRET_MANAGER evidence records
Contributor
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
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.
Merges 19 commits from dev into main.
Highlights
Deploy expectations (Vercel auto-deploy on main)
/launch-config.jsstays the six-field disabled shape./api/rewritefree tier keeps serving; pro tier keeps Polar license gate (401/403 fail-closed).Gates: release:check ✓, release:ready ✓, npm test 1664/0 ✓, lint ✓, preflight hold valid ✓.