Source Report — Fidelity Checklist

The originating artifact for this program is one candidate's account of an OpenAI software-engineering loop, posted to a subreddit and roughly four days old at time of capture. It is a single unverified source. The loop varies by team, by level, and by quarter.

This file exists so that nothing from it is silently dropped. Every detail is a row. Every row has a destination in the program and a corroboration status. If a row has no destination, the program has a hole.

How to Use This File
Corroboration Legend
The Checklist
Coverage Audit
The Anti-Narrowing Clause
References

How to Use This File

Read it at three moments:

Now, once, so you know what the program is built against.
At each monthly diagnostic re-test, to confirm no row has quietly gone unaddressed.
In the final two weeks, as a pre-flight checklist — every row should map to something you have done, not something you have read.

The rightmost column is a live audit surface. A row whose destination file does not exist yet is a tracked gap, not an oversight.

Symbol	Meaning
✅	Independently corroborated by at least one source unrelated to the report
🟡	Consistent with other sources but not directly corroborated
⚪	Uncorroborated — reported by this source only
⚠️	Other sources actively disagree; see note

The Checklist

Stage 1: Recruiter Screen

#	Detail as reported	Corrob.	Where it lands
1	Post titled "OpenAI Software Engineer Interview 2026", written by a candidate who just finished the loop	⚪	Recency assumption. Re-run the Phase 0 search at each monthly diagnostic; supersede this file if something newer and better-sourced appears
2	Process described as different from most big tech companies	✅	Program-wide constraint: no generic FAANG template. Corroborated by every source describing gated formats, work trials, and practical-over-algorithmic problems
3	Recruiter screen is a light background conversation	✅	Track F — the 90-second and 3-minute career narrative, rehearsed to a timer
4	Recruiter asked where the candidate thinks AI is headed	✅	Track F forward-looking answers + the written technical opinion essay in `projects/`. Weight raised: peer-lab sources report values rounds as the top failure mode
5	Advice: read the charter before that call	✅	`company-brief.md`. Read the primary text in a browser — it 403s to fetchers

Stage 2: Technical Screen

#	Detail as reported	Corrob.	Where it lands
6	Technical screen is two 60-minute rounds on the same day	✅	`mocks/` — back-to-back mock protocol. Fatigue is part of the signal; never mock these in isolation after week 4
7	Screen round A: coding — a versioned key-value store	✅	Track A problem `versioned-kv` in the progressive harness, 4 gates. Independently the most-reported OpenAI coding problem
8	Screen round B: system design — a job scheduler with fault tolerance	✅	Track C design exercise `d01-job-scheduler` — required, first design written
—	(not in report) Excalidraw is the reported design tool	✅	Practice designs in a shared-canvas tool, not on paper. Diagram-under-time-pressure is a trained motor skill

Stage 3: Take-Home

#	Detail as reported	Corrob.	Where it lands
9	Take-home: 48-hour window, "build something real"	✅	Track E — run under a real wall-clock 48h, twice
10	Take-home example: distributed webhook delivery system	✅	`projects/webhook-delivery/`. Corroborated by an independent vendor source naming a webhook delivery system as a work-trial project
11	Take-home required retry logic	✅	Exponential backoff with jitter (full/equal/decorrelated compared and benchmarked), retry budgets, idempotency keys
12	Take-home required dead-letter queues	✅	DLQ with a replay path, poison-message detection, and a documented redrive procedure
—	(not in report) Reported grading criteria: code quality, test coverage, written design doc explaining tradeoffs, handling of under-specified parts	✅	The 48-hour playbook's non-negotiables list. "Working + thoughtful README" beats "clever + undocumented"
—	(not in report) Reported as paid and under NDA	⚪	Irrelevant to preparation. Recorded so it is not a surprise

Stage 4: Deep Dive

#	Detail as reported	Corrob.	Where it lands
13	Deep dive: interviewer walked the take-home line by line	🟡	Track E — line-level defense drill. Every file, every default, every omission
14	Interviewer had a question list covering every choice and decision made	🟡	The interrogation harness generates that list from your actual diff, not from a template
15	That list was written by the interviewer himself after seeing the project	🟡	Enforces the design constraint: questions must be project-specific. Drives inference I1 in `findings.md` — keep a decision log while building, because every decision becomes a question

Stage 5: Onsite

#	Detail as reported	Corrob.	Where it lands
16	Onsite described as 4 rounds	⚠️	Sources report 4, 4–6, and 6 components (adding a technical presentation and a separate team-fit round). Program prepares for six. See row 38
17	Coding 1: progressive multi-part format	✅	The progressive harness — highest-priority build in the program
18	Each stage must have a working solution before the next opens	✅	Harness gates stage N+1 on stage N's tests passing. Corroborated: reported as ~4 gates per problem
19	Coding 1 example: token-level streaming differ	⚪	Built: `token-stream-differ`, 4 gates. Uncorroborated specifically, but structurally identical to the corroborated "resumable iterator with state serialization" pattern
20	...tracking state changes with rollback	⚪	Checkpoint / undo semantics inside that problem — gates 3 and 4
21	Explicit tactic: get something correct early, then iterate	✅	Time-to-first-correct is a tracked, charted metric in the harness timing log. Corroborated by the reported "clear 2 of 4 gates" pass bar — an unopened gate scores zero
22	Coding 2: more systems-flavored	✅	Track A systems subset. Corroborated: "practical over algorithmic," "not string manipulation"
23	Coding 2 themes: state management, concurrency, memory efficiency	✅	Three named drill categories; every harness problem is tagged with at least one
24	Python internals came up	✅	Track B in full. Corroborated: "coroutines and concurrency" listed as an OpenAI-specific topic
25	Specifically: generators	✅	Track B — protocol, `yield from` delegation, `send`/`throw`/`close`, generators as state machines
26	Specifically: async constructs	✅	Track B — event loop mechanics, task scheduling, cancellation, `gather` vs `TaskGroup`, async generators, `aclosing`
27	Specifically: iterators	✅	Track B — iterator protocol, laziness, custom iterables, resumable iterators with serializable state (a directly reported problem)
28	System design prompt: design ChatGPT	✅	Track D flagship exercise `d02-design-chatgpt`, drilled at two altitudes
29	Interviewer cared about GPU allocation	🟡	Memory math (weights + KV + activations), placement, multi-tenancy, fragmentation, cold start
30	...autoscaling under non-stationary traffic	✅	Token-rate and queue-depth signals, predictive vs reactive, warm pools, admission control, load shedding. Grounded in SageServe / ENOVA
31	...distributed coordination	🟡	Scheduler placement, health and drain, rolling model rollouts, canaries, config propagation, global rate limiting
32	Advice: abstract the model-serving layer unless told otherwise	⚪	Drill both altitudes; default to abstracted. See inference I2 — this is a judgement test, not a depth hint
33	Behavioral: technical leadership stories	✅	Story bank, leadership-tagged
34	Behavioral: architectural decisions affecting multiple teams	✅	Required story category. Weight raised by compressed levelling (row 38 note)
35	Behavioral: building consensus under pressure	✅	Required category, including a disagreement you lost
36	Behavioral: concrete tradeoffs, not soft-skills answers	✅	Rubric explicitly penalizes feelings-first answers; scores decision quality and tradeoff articulation

Stage 6: Agentic Round

#	Detail as reported	Corrob.	Where it lands
37	Agentic coding round exists but is in beta	✅	Track G. Prepare for it; do not assume it appears
38	Only some candidates get it, as a fifth round	⚠️	Arithmetic tension with row 16's "4 rounds" — loop size varies. Program prepares for a six-component onsite so an extra round is never a surprise
39	Format: existing codebase plus a problem too large to solve by hand	✅	Track G — real mid-size OSS Python repo, six oversized tasks. Corroborated: reported industry format is a multi-file codebase with phased objectives
40	Expectation: work through it using an AI coding agent	✅	Track G — agent-driving is the scored skill. Corroborated criteria: prompt construction, output validation, debugging the assistant's work
—	(not in report) AI tool policy is opposite at different labs — one peer lab reportedly bans AI in live rounds entirely	✅	Pre-interview checklist in `STATE.md`: ask the recruiter, per company, per round. Never assume

#	Detail as reported	Corrob.	Where it lands
41	The poster's own open question was how to prep for the progressive multi-part format	✅	Confirms it as the least-practiced round. Practice time weighted accordingly — the harness is the program's daily anchor, not a weekly exercise

Run this audit at every monthly diagnostic. A row is covered only when its destination exists and you have a passed drill, a working artifact, or a scored mock against it — never on the basis of having read something.

Destination	Rows it must cover	Exists
`company-brief.md`	4, 5	✅
Track A + harness	7, 17, 18, 19, 20, 21, 22, 23, 41	✅
Track B	24, 25, 26, 27	✅
Track C	8	✅
Track D	28, 29, 30, 31, 32	✅
Track E + `projects/`	9, 10, 11, 12, 13, 14, 15	✅
Track F	3, 33, 34, 35, 36	✅
Track G	37, 38, 39, 40	✅
`mocks/`	6, 16, 38	✅
Program-wide constraints	1, 2	✅

Every one of the 41 rows has a destination. Existence of the destination is not evidence of competence — that is what the rubrics and the review/ queue are for.

The Anti-Narrowing Clause

The provenance note attached to the source report is the most important sentence in it:

...do not let one report narrow the preparation so far that an unexpected round is a surprise.

Three concrete guards, enforced structurally rather than by good intentions:

Breadth floor. Every track carries material the source report does not mention but corroborated sources do: LRU caches, rate limiters, spreadsheet dependency evaluation with cycle detection, symlink resolution, multithreaded crawlers, in-memory SQL, and the occasional math-flavored problem (KL divergence, expected iterations). Roughly 25% of Track A's problem set is deliberately off-report.
Round-count buffer. The onsite is prepared as six components, including a technical presentation and a separate team-fit conversation that the report never mentions but aggregators do.
Company-agnostic core. Tracks A–D and F transfer to Anthropic, DeepMind, Scale, Cursor, xAI, Databricks, Netflix, and Stripe with only Track D's depth and Track F's mission material swapped. The program is not overfit to one company, and the company-specific surface is deliberately isolated in company-brief.md.

References

findings.md — the corroboration evidence and full source list behind every ✅/🟡/⚪/⚠️ in this file
company-brief.md — mission/charter digest, talking points, questions to ask
../PLAN.md — how these rows become weeks
../STATE.md — the live progress ledger and pre-interview checklist

Elite AI-Lab SWE Interview Program — Diagnostics, Drills, Builds & Scored Mocks