SF Tech Week Atlas: methods and scope
Public release, 20 October 2026. Based on the corrected October 19 snapshot. Back to the Atlas.
This copy applies confirmed findings from the 18 October audit. It is a descriptive archive of announced programs. It is not a census of attended events, verified unique people or business outcomes.
The analytical population
The source archive contains 1,593 calendar records. The analytical projection excludes OOO - The Techno Sound, collapses one pair using the same registration-page key, and joins one explicitly transferred ZenML program to its replacement. The result is 1,590 analytical records.
One record can still describe a session, application page, multi-session program or registration page associated with another record. Complete real-world program deduplication has not been performed. Archived cancelled and postponed announcements remain in topic and descriptive RSVP analyses, with explicit status. A default status of announced does not establish that an event happened.
The old ZenML page is retained as historical provenance under its replacement. Its 379 historical RSVPs are not assigned to the current program or included in current totals. The replacement has no observed current counter. Old-page topic and affiliation evidence is not silently transferred to the replacement.
RSVP statistics
| Measure | Corrected value | Meaning |
|---|---|---|
| Analytical records | 1,590 | Retained announcements after confirmed merges |
| Records with counters | 1,141 | Missing counters excluded; explicit zero retained |
| Missing counters | 449 | No inferred zero or imputed popularity |
| Median | 195 | Median visible registration counter |
| Maximum | 2,022 | Excludes the editorially omitted OOO page |
| Sum of counters | 279,799 | Registrations across pages, not unique attendees |
| 90th percentile | 502 | Linear interpolation |
Event ranking uses counter descending, then stable record ID. Topic comparisons use main/secondary labels, exclude missing counters, and display the number observed. Quartiles and percentiles use linear interpolation. Counter snapshots come from the archived observation dates; they have not been refreshed.
Dates, lifecycle and geographic views
Ten confirmed calendar/Partiful conflicts have explicit adjudication and source observations. Eight concrete conflicts use the archived Partiful date; two TBD cases have no effective start. Other confirmed cancellation/postponement and replacement cases carry lifecycle notes. Historical source dates remain observations rather than being erased.
The hourly grid covers all 24 hours in America/Los_Angeles. It contains 1,583 timed announcements. Three program windows and four records without a confirmed date are separately listed, giving the same 1,590-record population. Program-window dates describe their boundaries, not a precise session start; their day and hour fields must always be interpreted together with time_kind.
Daily map filtering includes timed announcements only. The full-week map includes every mapped announcement, including program windows and unknown-date records. The geographic crosswalk covers 869 records. It maps official area labels to polygons and does not establish exact venue density. For duplicate registration pages, the representative record supplies the location. Conflicting source labels are not combined into multiple asserted locations. One source pair has SOMA/Hayes Valley disagreement that remains unresolved.
People, organizations and roles
The corrected directory contains 2,333 person source identities, 3,799 organization source identities, 346 X links, 1,277 LinkedIn links and 1,130 cached portraits. These counts do not imply complete identity resolution.
Two OCR spellings of Sasa Zdjelar merge into the existing supported identity. Twelve FARR after-party people are shown as featured guests, outside the speaker catalog, because the text explicitly rules out speaking. Historical employers, educational affiliations and accelerator/program experience are distinct from current speaker affiliations. Judge, mentor and host relationships retain their own role types. The scope of the featured-guest projection is the confirmed FARR correction, not all featured guests in the raw archive.
Unresolved logo placeholders and the confirmed GFT misread are quarantined. The unsupported Replit event edge is removed while valid Replit records remain. The unsupported Michael Jones X profile, biography and portrait are removed. The unsupported Dylan Jadeja X profile is removed while independently sourced LinkedIn details remain. Other legacy profile matches have not all been reverified.
The six-company comparison
The fixed name mapping covers Google, Amazon/AWS, Apple, Meta, Microsoft and NVIDIA. This is not a complete corporate-family map, company-size comparison or causal estimate of registrations.
Role groups are mutually exclusive: organizer/sponsor takes precedence, followed by a current speaker affiliation, then no recorded qualifying link to the six selected companies. Historical, educational, program, judge, mentor and host affiliations do not qualify as current speaker links. Other firms can still appear in the last group.
| Group | Records | Observed counters | Median RSVP |
|---|---|---|---|
| Organizer or sponsor | 34 | 19 | 239 |
| Current speaker affiliation only | 54 | 35 | 364 |
| No recorded organizer, sponsor or speaker link | 1,502 | 1,087 | 190 |
Subjects, phrases and graphs
Technology topics and business themes use main/secondary salience. Formats such as breakfast and happy hour remain separate. Confirmed corrections do not settle the open classification boundary cases from the audit. Applied AI workflows and adoption still labels 316 announcements, while AI infrastructure and serving labels 87. These are advertised topics, not verified implementations or business results.
Technology/business pairs require at least 10 joint records. Lift is joint count times population size divided by the product of marginal counts. Theme-graph nodes require at least 25 records; edges require at least eight joint records and lift of 1.2. The graph contains 32 topic nodes and 121 edges. Layout position is not an additional metric of importance.
The ecosystem graph keeps the existing bounded company selection, then rebuilds links and node counts from corrected records. Its 1,124 nodes and 1,838 links do not represent the whole archive. Company and person sizes use full-cohort counts; links show the graph subset. The hero illustration has one deterministic decorative point per analytical record. It does not encode semantic similarity.
The ranked phrase display uses the existing lexical pipeline: adjacent English word pairs within sentence boundaries, repeated boilerplate removal, explicit stop/exclusion lists, a minimum of eight records and the top 60 phrases. Clicking a displayed phrase selects that lexical membership and labels the mode. Typing in search always uses the sanitized full-text index plus topic labels, even if the text matches a ranked phrase. The two search populations can legitimately differ.
X recaps and visual evidence
The selection preserves 214 posts from 172 author accounts out of 495 cached candidates. Search pagination was not exhausted, so it is a bounded editorial selection. Three authored summaries are corrected: Grace Ling's name in two posts, running rather than walking in one of those, and Christin Glorioso's name. Original authorship, counters, timestamps and source evidence remain.
Twelve curiosities retain their original media: three X photos, two video previews, one diagram and six archived invitation covers. Invitation covers are not event photographs. A recap or recording announcement does not prove that a playable recording exists. No new X or LinkedIn collection was performed. Live X widgets depend on the external provider; offline tests verify the cached fallback and interactions.
Privacy and local assets
The exported natural-language fields remove meeting join links, access details, contact email addresses and phone numbers. Identifiers, hashes and ordinary source URLs are not transformed with phone-redaction rules. Tests check the whole JSON and HTML for the known audit exposures and newsletter recipient tracking. This is a bounded redaction policy, not a guarantee that all possible sensitive content in the original archive has been classified.
The local preview contains a file allowlist and its own referenced portraits, invitation covers, font and logo. It does not include SQLite, raw provider results, private newsletter bodies or the correction evidence. Some social media remain remote references. The allowlist is packaging for local review, not approval to publish.
Reproducibility and service status
data.json is the single analytical contract. presentation-data.json is byte-identical, and the JSON embedded in ATLAS.html has the same parsed content. Corrections are explicit reviewed overlays on immutable input copies. Stable ordering, seeded graph layouts and pinned Python dependencies support reproducible output. Two builds with different Python hash seeds produced identical output hashes.
Snapshot ID: sf-2026-confirmed-corrections-20261019.
Data SHA-256: a4cd678223dd3bf5e4303daedd361df3e5bd093bc042002ce872e450e1811427.
release-manifest.json lists local package checksums and source/code hashes. It is a local provenance record, not a public release. The existing lead-form and MCP demonstrations remain simulations. No deployment, new MCP service, main SQLite mutation or Supabase synchronization was performed.