Zeltru Research · Archive series

Archived Greenhouse boards, 2023–2024

What a fixed, non-random panel of archived Greenhouse job boards recorded between sightings. Separate from Zeltru’s mixed-platform live series, and measuring something different.

Published September 15, 2026·Reference period January 1, 2023 – December 31, 2024·Extract September 15, 2026·All research

§1Coverage

The fixed panel contains 705 archived Greenhouse boards, each with at least one screened candidate capture in both years. A board is not necessarily a distinct employer.

The screening categories partition successfully fetched captures. All 82 unresolved report-period records fall in 2024. They account for 0.68% of its 12,086 known capture records. A count of missing records does not bound the number of affected roles or the effect on a span statistic.

Two boards account for 71 of 82 unresolved records (86.6%): one has 56 awaiting retrieval, the other 15 marked failed inside the report period. The remaining 11 are spread across seven other boards.

Including the 180-day screening context on either side of the reference period gives 149 unresolved records: 115 awaiting retrieval and 34 failed. This includes the report-period 82. The full preserved panel history has 163 unresolved records among 45,236 known captures. These are counts within the discovered, digest-collapsed index; unknown visits and undiscovered hostnames are outside that denominator.

The 705-board panel replaces an earlier 988-board candidate because the time-local screen explicitly leaves sparse captures unscreened. This is a methodological eligibility change, not evidence that 283 boards disappeared.

The span subset contains 254 boards that pass both the stored and independently recomputed global median-gap checks of at most seven days. All 254 contribute to the positive-span cohort. The fixed panel and subset retain their original memberships.

The global tier applies to every span figure in this report, including the sensitivity subsets. It describes gaps between retained screened captures, not archive-visit frequency or continuous weekly coverage. A small global median does not prevent long individual gaps.

§2What an archived Greenhouse observation establishes

An archived appearance establishes that a posting was observed at a particular capture time. The interval from first to last archived sighting is an observed span; it does not prove continuous presence between sightings. The archive records used here do not contain employer-published posting dates, so these observations cannot reconstruct the live series’ posting-age metric.

A retained observed span is the elapsed time between the first and last stored archived timestamps for one composite board/job key. The calculation uses full timestamp precision and requires both endpoint captures to pass the partial-capture screen. Displayed span and gap estimates are rounded to the nearest whole day; the underlying calculations, sort order and threshold assignments retain full precision. Rounding is a display choice, not a claim of day-level certainty about actual listing duration.

The span describes recorded endpoint extent. It is not posting age, uninterrupted duration, time to fill or a count of distinct hiring needs. Unobserved disappearance, reappearance and identifier reuse remain possible. Under consistent identity, retained endpoints give a lower bound on the elapsed extent spanned by a fuller sighting history, never a lower bound on continuous listing duration. The first recorded sighting may follow an earlier appearance, and the last may precede a later appearance.

§3Retained observed spans across 254 eligible boards

The eligible sampling frame contains 254 archived Greenhouse boards. Within it, 125,270 distinct board/job keys have both stored endpoint timestamps matched to screened candidate captures and at least one such endpoint in 2023–2024. All 254 boards contribute to this exact cohort and to its positive-span subset.

• 20,722 (16.54%) have identical first and last timestamps and contribute no positive retained span.
• 2,238 (1.79%) have different timestamps on the same calendar date.
• 102,310 (81.67%) have endpoints on different calendar dates.

The latter two groups form 104,548 positive retained spans (83.46% of the cohort). The following distribution is conditional on those positive spans; percentages do not use all 125,270 keys as their denominator.

Percentages may not sum to exactly 100% because of rounding. These spans use aggregate endpoints across all stored history, selected by endpoint evidence in 2023–2024. Some endpoints lie outside the report period. They are not within-year spans, posting ages, uninterrupted durations, unique opening counts or time-to-fill measurements. Identical-timestamp cases are disclosed, not interpreted as zero-duration jobs. Screening and digest collapsing still select the observed evidence. The median among the 104,548 positive spans is 28 days, rounded to the nearest whole day. The 20,722 identical-timestamp records remain outside that median and its distribution. Of the positive spans, 13,147 (12.58%) have an endpoint outside the reference period; the endpoints were not clipped to calendar boundaries.

Selection by endpoint. A key enters the cohort only when at least one of its retained screened endpoints falls inside 2023–2024. A key first recorded before 2023 therefore qualifies only when a retained endpoint reaches the reference window; one whose retained history ended before 2023 is not counted, however far its record extends. Keys first recorded after 2024 are likewise excluded. How this selection shifts the distribution has not been measured, and no overall bias, in either direction or of any size, is claimed.

What this lower bound means. With a stable job identity and correctly parsed endpoints, each retained span is a lower bound on the elapsed extent from that key’s earliest to latest actual appearance. This also bounds the median of that extent for the identical 104,548-key subset. It does not bound uninterrupted or cumulative time visible: a key may disappear and later return. The displayed 28-day median rounds upward, so it is not a strict lower bound of 28 days, and it does not mean any posting was continuously listed for that long.

An inclusive 30-day-or-longer threshold contains 50,493 positive spans (48.3%), calculated before rounding. No positive span in this frozen cohort is exactly 30 days, so the inclusive count equals the count above 30 days in the table.

Numerical proximity does not validate the live series. This archive’s rounded median of 28 days and 48.3% share of positive spans reaching 30 days sit near Edition No. 1’s median posting age of 28 days and 48.5% share listed 30 days or longer. These describe different metrics, populations and eras: selected positive endpoint spans on archived Greenhouse boards, versus dated active postings across five live platforms. They are not corroboration or a historical trend.

Individual postings, job identifiers, stored titles and exact endpoint records are not published. The longest retained spans are counted in the more-than-365-days row above; no individual record is presented as evidence of a single continuously open vacancy.

§4Concentration, sensitivity and comparability

Each positive-span role key receives equal weight, so boards with more keys contribute more to the pooled distribution. The largest contributing board supplies 9,006 positive spans (8.61%). The five largest supply 21.71%; the ten largest supply 31.51%. The median of the 254 individual-board medians is 31 days, a different summary from the pooled role-key median.

Thirteen span-cohort boards have unresolved captures somewhere in preserved history and contribute 6,166 positive spans (5.90%). Removing all thirteen changes the rounded median from 28 to 27 days; the unrounded shift is less than one day. This exclusion covers every such board, not only the two holding most report-period unresolved records. Removing the five largest contributors changes it to 33 days. These checks show the effect of the stated exclusions on the observed sample. They are not imputed results, correction factors, or upper bounds on missing-data bias.

Coverage totals compare the same 705 boards across both years. The selected span distribution combines endpoint evidence across the reference window; it is not a year-over-year hiring index. Changes in capture effort, board composition or weighting cannot be interpreted as changes in the wider job market.

The source inventory discovers Greenhouse pages under boards.greenhouse.io and collapses adjacent repeated content digests. The report uses that old-host discovery scope and the frozen panel. Other platforms and broader hostname discovery are separate workstreams. A bounded earlier metadata check found later snapshots under job-boards.greenhouse.io for four selected boards, but those snapshots are not incorporated into these figures.

The fixed source was acquired on September 15, 2026 at 06:15:35 UTC. All figure calculations use that frozen scope and do not depend on any later collection. A new extract incorporating later retrievals or retries could change the figures and would be published as its own dated revision, not as a silent change to this report.

§5Limitations and screening

Partial-capture screen. Each valid successful capture is compared with the median job count of at least three other valid successful captures of the same board within ±180 days, including the boundary. The target is excluded from its own baseline. Zero counts participate. A count strictly below 25% of a positive median is marked suspected partial and excluded. Too few neighbors or a zero median leaves a capture unflagged but unscreened and excluded from the screened sample. This heuristic does not certify completeness or prove truncation.

Selected capture history. CDX digest collapsing omits adjacent repeated-digest records after the first. It can omit later unchanged sightings and shorten retained endpoint spans relative to a fuller archive history. The size of that effect has not been estimated. Even a long-lived posting can have an identical-timestamp record if later unchanged captures were collapsed. A zero retained span therefore does not distinguish a brief appearance, sparse capture coverage or omitted repeat sightings; this extract cannot quantify those explanations. Content can change for reasons unrelated to jobs, so retained-capture spacing measures neither complete archive-visit frequency nor job-change frequency. Internet Archive CDX documentation.

What is missing. The archive records used here have no employer-published posting date, full per-capture role membership, retained HTML or HTTP status for retrieval outcomes. A failed capture or an absent record does not establish removal, and no field in this extract confirms one. Dates and counts cannot establish time to fill or employer intent.

Scope and transferability. The archive panel is non-random and Greenhouse-only. The role-key sample contains selected records with usable endpoints. Its results cannot be spliced into Zeltru’s mixed-platform live series. Possible overlap with live observations in 2026 may support a future paired validation study, but no correction factor is applied to 2023–2024.

§6Sources and methods

The scoped SQLite source contains 705 board rows, 45,236 capture rows and 376,688 role rows, preserving all recorded panel history needed for local screening and global endpoints. The projection contains the declared board, capture and role columns; it is not a backup of every table in the original database.

The SQLite integrity check passed. All eight evidence file hashes match their manifest, including the source database and endpoint file. The scoped logical hash was reproduced independently. All 45,236 capture statuses were checked against the source, including replay of the 45,073 successful-capture screen decisions. All 286,798 endpoint records match the source records and recalculated flags. Cohort counts, bins and membership checks reconcile. No duplicate role keys or endpoint-evidence counts exceeding the corresponding capture’s stored role count were found.

Database SHA-256

9b34f3a4309db98caf6da48512ab51f9b2316268cfb5ac232cb3b35596e6ad88

Scoped logical SHA-256

59a01d834edd58e76b9e35b5f510fb8fe5d926d067874626ab7f7487222f040d

Fixed panel ID

b4b45cc59e63e58bb8aa5b30c78c706a49c36dfde13a30ddd1bc827c7fee1011

Span-subset ID

5adc5ec8236b285575b371435096c1f535d157d9d638c0bf4c4c191835db535b

The supporting records — board-level contributions, row-level endpoint records and unresolved capture keys — are retained internally and are not published; they are available for verification on request. These verification checks establish consistency with the preserved records, not exhaustive Internet Archive coverage or uninterrupted listing histories.

Appendix: role-key counts and period evidence

These are differently scoped key counts, not successive-year totals, new openings or distinct employers. The combined-period endpoint counts establish evidence somewhere in the two-year window; they cannot reconstruct every intermediate sighting or an exact annual census. Interval overlap is a proxy and is not treated as proof of an observation in an intervening year. Summing jobs found across captures would count repeated sightings rather than distinct role keys.