Skip to content

v0.46.0-beta — Hazelnut build 8

Latest

Choose a tag to compare

@tobias-rosenbaum tobias-rosenbaum released this 11 Nov 20:38
· 10 commits to main since this release
e55099b

The CV says what the profile holds

  • A skill on the CV now has its bullet (#786, 786:289). When the skills list or the summary names something your profile records in a role (for example "Mentoring"), and no bullet of that role shows it, the last review round asks for it in that role's bullets. A role that already says it in other words is left alone.
  • The CV states durations the way you did (#786, 786:190, founder ruling W-2). The CV no longer reads a computed skill span as a figure: it states your own figures as you wrote them, a careful "over 5 years" only for a tool you never put a number on, and no duration at all for a tool whose computed span is as long as or longer than your own career figure ("9 years of PostgreSQL" beside "8 years of experience"). If your German and your English summary state different career figures, the smaller one counts. CV and cover letter use the same rule.

Checks and verdicts

  • The truthfulness check knows audit attestations. A CV or letter naming SOC 2, ISAE 3402, BSI C5 (or SOC 1/3, ISAE 3000, SSAE 16/18) that your profile never mentions is now marked as not backed, with "a standard your profile never mentions". Before, the sentence's other words could mark it as backed. SOC 2 in your profile does not back SOC 1.
  • "Duplicate bullets" no longer pairs two different achievements (#786). Bullets are compared only within one employer. The same task at two companies is two achievements; one company written with and without its legal form ("Nordwerk GmbH" / "Nordwerk") is one employer. Two bullets with different years, or different percentages, are two facts. The review loop no longer asks to merge or drop one of them. On 93 checked pairs from captured test CVs the share of false pairs fell from 37 % to 17 %; every real duplicate is still found.
  • A bullet's first word counts as evidence (#786). "Mentor two engineers" at the start of a bullet now backs "Mentoring". Before, the narrative check could say the skill was only listed.
  • "Literally in your profile" means your profile's content (#790, #783). A skill you removed, an unaccepted value in an open conflict, and Applire's own bookkeeping no longer count as your profile stating a term, so the consistency checks and the pre-download diff ask about it.

The letter states only checked facts

  • A figure no source states can no longer slip into the letter after its review (#784; ADR-076 amended 2026-11-10). When shortening the letter or making the last correction added a number that is not in your profile or the posting (a "92 %" latency gain, an accident rate written as a percentage), the check now catches it every round. If the last correction is the one that adds it, the reviewed version ships. A number still left in the letter is named on the letter's review card ("Die Zahl „…“ steht in keiner Deiner Angaben"). The letter is never rewritten by code. On 521 recorded shortening and correction steps it fired 5 times, each time on a real invented figure, and never on a figure from your profile.
  • A standard is not a figure (#784). "IFRS 16", "ISO 9001", "SOC 2" or "Industrie 4.0" no longer trigger "this number belongs to another job". On 448 recorded letters the false alarms fell from 14 to 1, and no real one was lost.
  • The last correction keeps a duration on what it counted (#784). Merging two sentences no longer moves your career-wide "ten years of audit practice" onto a job that started later. Replayed on the recorded case: 3 of 5 → 1 of 5 on gpt-6-luna, 0 on Claude Sonnet 5.5; the ten years were never dropped.
  • Final review: a job you left is not current knowledge of the employer (#784). "I know your way of working from my own practice" about an employer you left in 2018 is now flagged in the last review. gpt-6-luna flagged it 1 of 6 → 3 of 5 times, Claude Sonnet 5.5 3 of 3, with no false flag on the same letter without that sentence.
  • Your own career figure caps a computed tool duration (founder ruling W-2). If the computed span for a tool is as long as or longer than the career figure in your own summary, the letter states no duration for that tool. It no longer offers "over 9 years of PostgreSQL" next to your own "8 years of experience". Shorter computed spans keep their wording, and a figure you gave yourself for a tool is never capped.
  • A duration in the letter must be one you stated (#784). "Ten years in laboratory diagnostics", "15 Jahre Mitarbeiterführung" or "over 10 years of Python" now block the review round when the duration matches none you gave and is not a tool duration the letter may state. The correction is told your own figures. A duration the posting states ("seit über 30 Jahren am Markt", "at least 2 years of experience"), one you entered for this application, or the length of one dated position or training in your profile ("my three-year apprenticeship") is not flagged; a total added up across positions still is. On 448 recorded letters it fired 10 times, each against a duration the person had stated differently.

The interview and the profile record only what was said

  • An interview question no longer turns a listed tag into claimed work (#787). The question writer reads your profile's tags as listed names and only your own answers as evidence, and the question groups say what the role expects without an invented reason. Replayed on recorded interviews: 8 of 15 → 2–3 of 15 on Claude Sonnet 5.5, 6 of 15 → 0 of 15 on gpt-6-luna; invented reason clauses 86 → 0.
  • What you denied stays denied when the posting words it differently (#787). You said you have not worked in a regulated environment; the posting asks for "regulated financial data". That requirement is no longer handed to the writers as backed (2 of 10 → 0 of 10 on gpt-6-luna, 1 of 5 → 0 of 5 on Claude Sonnet 5.5), while "financial data" on its own still counts in every run.
  • An answer adds only the skills you work with yourself (#787). A tool your team or a supplier used no longer becomes your skill when the answer is saved: 12 of 14 → 0 of 16 on gpt-6-luna, 6 of 9 → 0 of 9 on Claude Sonnet 5.5. A skill named after the question's own gap label fell from 10 of 10 to 4 of 17 and is not gone.
  • The posting analysis drops non-skills in its correction step too (#787). Phrases such as benefits or company descriptions that the first reading already excludes no longer come back as requirements after the correction (kept 8 of 8 before; removed 4 of 5 on gpt-6-luna, 3 of 3 on Claude Sonnet 5.5).
  • The posting review no longer accepts a translated requirement as the original (#787).
  • The interview's end screen counts the questions you answered.
  • "Which employer does this belong to?" reads your answer, not Applire's question (ADR-063 amended 2026-11-10). When an interview question named an earlier employer and you answered about your current one, Applire held everything from that answer behind a card saying "the answer named" the earlier employer and offered to move it there. The check that compares the whole answer with the role it was saved under now reads your answer only. Found in the build 8 delivery run; on five replayed shapes the false card went from 4 to 0, and the card still appears when your answer itself names the other employer.

The page

  • Long CVs no longer leave pages part-empty (founder re-UAT RU-3; ruling P-4). With long bullets, a role's heading, its first bullets and its projects were kept together as one block, and every page break pushed the whole block down. The "Kreativ" template left about 0.8 of a page blank over 5 pages. All seven templates now keep a heading only with the first lines of its first bullet. A bullet of six or more lines may continue on the next page, with at least three lines on each; a shorter one never splits. A role's projects still never open a page alone. On a long synthetic CV: executive 6 → 4 pages, tech 5 → 4, compact 6 → 5; academic, classic and modern stay at 5, 5 and 4.
  • "Kreativ" has a colour band instead of a sidebar (founder re-UAT RU-3; ruling P-5). The sidebar repeated its coloured column on every page and narrowed the text on all of them. Now page 1 opens with a coloured header band (name, round photo, contact line), skills sit as chips and languages beside them under the profile, and every page uses the full width with 16 mm margins. The matching cover letter takes the same band. On a long synthetic CV: 6 → 4 pages. A stored application that uses "Kreativ" keeps it; the picker describes the new look.
  • The automatic page cut keeps your newest certifications (786:256; ruling P-1). To reach one page over your target, Applire used to drop certifications one at a time until none were left (3 of 6 test CVs lost all five). Now it never drops a certification the posting names, it always keeps your two newest, and it puts back anything it cut that did not save a page. Cutting further is still your choice on the page card, which names every item.
  • The team size under a role says what it counts (founder re-UAT RU-5; ruling P-2): „Führungsverantwortung für 2 Personen“ / "Responsible for a team of 2" instead of a bare "2 Personen", which read as a contradiction next to a larger team in the bullets.
  • The page card names the one-page tolerance (founder re-UAT RU-4): „Von sich aus kürzt Applire nur bis auf eine Seite über Deinem Ziel, hier 4 Seiten.“ It says so when even that was not reached.
  • The template-picker sample uses invented companies and German throughout (786:180).

Install and upgrade

  • The upgrade notice names the right release for two settings. APPLIRE_HTTP_PORT and the changed meaning of LOG_LEVEL shipped with 0.45.0-beta but were labelled 1.0.0. The notice would have listed them again after every update until a 1.0.0 existed. .env.example now reads since: 0.45.0 and meaning changed in: 0.45.0.
  • README: the configuration example shows the shipped model defaults (mistral-small-latest, mistralai/mistral-medium-3-5); it had drifted from .env.example. The install section says what to set when port 80 is taken (APPLIRE_HTTP_PORT), and the German README states the download size of the first pull. Found by the pre-release install test.

Upgrade notes

  • None. No setting was added, removed or given a new meaning, and no migration runs. docker-compose.yml is unchanged; .env.example changes two comment lines (the release labels above), no variable. The MCP tools keep their names, parameters and types.