Claim ledger and evidence limits · reviewed 2026-09-22

The original 2026-09-21 notes mixed search excerpts, historical sources, vendor claims and curriculum recommendations. A read date did not establish publication freshness. This ledger replaces their unsupported authority claims. Earlier wording remains in git history; the withdrawal register identifies what must not be cited as evidence.

For this review, dated interview/recruiting/engineering evidence must be published 2025-09-22 or later, preferably 2026-03-22 or later. Event dates are tracked independently. Historical technical material may explain a mechanism, but does not count toward recent interview coverage. Undated live official pages have publication age unknown, even when accessed today. All rows below were accessed 2026-09-22.

Retained claims with exact provenance

ID and narrow claim	Exact source / primary status	Publication; event	Sample, role and applicability	Curriculum inference and limit
C01: good-event ratios and paired-window alerts	Google SRE Workbook: Alerting on SLOs; primary technical guidance	Historical 2018 workbook; no interview event	Authors' request-driven examples, not a random service sample	Use explicit denominator and Boolean logic. Our variable-traffic arithmetic/latch fixtures are constructed; workbook is outside recent-evidence window.
C02: SDK retry classification includes throttling; attempt count includes original	AWS SDK retry behavior; official technical docs	Publication age unknown; no interview event	SDK/service/version-specific; inspect pinned version	Correct “never retry 4xx” and distinguish attempts/retries. Our deadline/aggregate quota policy is additional application design.
C03: a trigger and sustaining feedback can cause persistent failure	Bronson et al., Metastable Failures in Distributed Systems; primary research	HotOS 2021; historical systems, not interviews	Mechanism paper, no hiring sample	Teach feedback and recovery; not every slow recovery proves metastability. Outside recency window.
C04: configuration and hot-shard failures occurred	GitHub July 2026 availability report; primary operator report	2026-08-12; incidents 2026-07-08 and 2026-07-09	GitHub's own systems/observations; no interview level	Casebook mechanisms; AWS translations and all numerical exercises are constructed. No prevalence claim.
C05: rollout capacity, retry pressure and stale-status failures occurred	GitHub August 2026 availability report; primary operator report	2026-09-09; incidents 2026-08-06, 17 and 20	GitHub operator account, not independently established causation	Cases support capacity/recovery practice. Conflicting duration summaries are not converted into exact duration claims.
C06: staged configuration and recovery access were reported	Cloudflare Code Orange update; primary engineering report	2026-05-01; drill 2026-04-07; motivating outages November/December 2025	Company remediation/drill, no hiring sample	Practice recovery-access independence; don't relabel old incidents as spring 2026.
C07: combine programmatic, judge and human evaluation	Airbnb eval-driven development; primary engineering, authors listed on article	2026-07-28; underlying development dates unspecified	Airbnb practices with a fictionalized walkthrough; no SWE interview sample	Supports mixed evaluation methods. Does not establish a universal sample size or judge threshold. Failure-recall/severity fixtures are our calculations.
C08: agents can manage context through retrieval, compaction and notes	Anthropic context engineering; primary vendor engineering	2025-09-29; no distinct implementation/interview event	Vendor observations, not independent all-model comparison	Candidate techniques to measure. Within one year, outside preferred six months; no universal retrieval chunk count.
C09: IC5 includes multi-team direction and organizational impact	Dropbox IC5 framework; official employer framework	Publication age unknown; no interview event	Dropbox IC5 only	Illustrates scope, does not prove coding standards stop increasing everywhere or prescribe another employer's level.
C10: four descriptive staff archetypes	StaffEng archetypes; author's practitioner synthesis, not an employer policy	Publication age unknown; interview/story dates not established here	Qualitative examples; no representative sample established	Vocabulary for comparing roles; no staffing ratios or promotion probabilities inferred.
C11: S3 conditional writes check object conditions	AWS conditional writes; official technical docs	Publication age unknown; no interview event	Specified APIs/conditions, not arbitrary transactions	Use for object-update preconditions. Does not mean a whole multi-object protocol is coordination-free.
C12: PostgreSQL CLUSTER does not maintain subsequent physical order	PostgreSQL 18 CLUSTER; versioned official docs	Page publication age unknown; no interview event	PostgreSQL 18 heap/table behavior	Withdraw blanket random-UUID-primary-key physical-locality claims; inspect actual index/heap workload.
C13: signal specifications have their own status	OpenTelemetry specification overview; official technical docs	Publication age unknown; no interview event	Language/signal/spec status varies	Consult exact SDK and semantic-convention versions; this source does not support the old graduation/adoption claim.

Original-note dispositions

Original file / material assertion	Disposition
`senior-craft-2026.md`: SLO units, 4xx, retry ownership, Netflix exact outcomes	Technical rule corrected through C01–C03; Netflix numbers/“current reference” withdrawn pending exact original source/date/context.
Same: Faros, DORA, passkey, secret-survival, CI/IaC adoption and CPU-overhead figures	Withdrawn; the original note did not preserve an exact claim-to-source/sample trail. No made-up citations supplied.
Same: OTel graduation/stability universals, market defaults, security/legal/licensing shorthand	Withdrawn as current evidence; C11–C13 retain only narrow technical semantics. Verify official version-specific sources for any future reintroduction.
`senior-craft-working-notes.md`: unlinked “verified” statements and search-count narrative	Superseded by withdrawal register; a research process claim is not source provenance.
`guide-research-ai-features.md`: consensus, fixed chunk/sample counts, mandatory non-100% eval, judge order percentage, attack percentages, pricing/TTFT constants	Withdrawn; C07–C08 support narrow methods. Fault-sensitive regressions and held-out metrics replace the false acceptance rule.
`staff-engineering-2026.md`: all ladders, archetype ratios/company-size rules, survey percentages, METR speed claim, mandatory promotion recipe	Withdrawn as universal/current evidence. C09–C10 preserve scoped role vocabulary; no 2025 study is refreshed by a 2026 read date.
`junior-foundation-research-2026.md`: “all areas became reading”, survey/duplication stats, blanket EU accessibility rule, platform adoption claims	Withdrawn pending exact evidence and applicability; independent implementation and output review are explicit curriculum choices.
`pedagogy-research-for-engineering-guide.md`: effect sizes, completion percentages, exact spacing intervals, unqualified “dead”/“ungameable” claims	Withdrawn pending primary study/sample review. Concrete contracts, worked examples and fading hints remain testable teaching choices, not quantified promises.

Review checks and ownership

Run python docs/research/check_claims.py from the repository root. It checks ledger IDs, local links, and mandatory date/provenance headings; it does not fetch pages or prove that they support a claim. The review above involved primary-page inspection separately.

The curriculum maintainer owns source review before company-specific preparation and at the next quarterly content review (2026-12-22). A source that changes materially or loses retrievable support is marked unresolved; do not silently transfer its claim to a nearby citation. Record author, exact URL, publication/event/access dates, sample/role, quoted or summarized claim, and the inference separately. Current company-loop gaps remain in the interview ledger.

Sources and further reading · 13