Round cache-key-backend-v4

H-BACKEND: STATUS VALID; primary finding NOT SUPPORTED

Status · VALIDReviewed status is transcribed from rounds/STATUS.md. The original narrative and its qualifications remain below.

Source-reported narrativeRound README and results text are source material. Numeric values that are not tied to a claim ledger record remain source-reported.

Round source README · rounds/cache-key-backend-v4/README.md

Round date: HISTORICAL (before 2026-10-09) Publication badge: VALID; KEPT FOR AUDIT Recomputation status: FULLY RECOMPUTABLE

STATUS: VALID — protocol completed as registered; primary finding NOT SUPPORTED.

This preregistered ChatGPT Codex backend round reached its 11 transport-complete pairs in one execution. The frozen shared-key ≥80% and distinct-key ≤5% condition held in 5 of 11 complete pairs. K=5/11; exact one-sided p=0.7255859375; 99% Clopper–Pearson CI=[0.114469, 0.830688]. VALID describes protocol completion and counter integrity; it does not mean the directional hypothesis was supported.

The earlier cache-key-sharing (not in this repository) and cache-replay-v3 (not in this repository) are separate prior attempts. Their observations were not pooled into this round. See DESIGN.md, DECISION-RULE.md, RESULTS.md, MEASURED.md, ENVIRONMENT.md, and sanitized.csv.

What it shows

On this account, model, route, and date, provider-reported cache reuse varied across newly seeded pairs. The frozen hypothesis did not reproduce at its 10-of-11 threshold. The run sent 44 requests, spent 503,482 charged or reserved tokens, and replaced 0 transport-incomplete pairs.

What it does not show

It does not locate physical cache entries, isolate prompt_cache_key from the linked session metadata, establish a future hit guarantee, measure billing savings or task quality, or explain the change from the prior attempts. The two endpoints use different OAuth grants and routing contexts.

Lifecycle counts

Units are request pairs (one shared-key and one distinct-key probe arm per pair), not benchmark task cells.

Lifecycle field n Basis
Planned 11 Frozen target of transport-complete pairs.
Started 11 Attempted pairs (no transport-incomplete pair was replaced).
Finished 11 Transport-complete pairs.
Officially graded 11 Pairs evaluated by the frozen counter rule.
ITT denominator 11 All complete pairs enter the registered rule.

Required reproduction metadata

  • Kogen commit: not applicable; this round sends direct provider requests and does not run Kogen.
  • Harness commit: not recorded in the round files; the probe script is not bundled in this repository.
  • Model and effort: gpt-6-luna as reported by every response; request settings are frozen in DESIGN.md.
  • Task IDs: not applicable; the fixed primer/probe fixture is identified by its manifest hash in DESIGN.md.
  • Historical execution command: not recorded in the round files.
  • Raw records: raw plans, receipts and credentials are not retained in this repository; per-pair counters are in sanitized.csv.

Round results · rounds/cache-key-backend-v4/RESULTS.md

STATUS VALID. Primary finding NOT SUPPORTED. The frozen shared-key ≥80% and distinct-key ≤5% condition held in 5 of 11 complete pairs. The rule required at least 10 successes among 11 complete pairs. K=5/11; exact one-sided p=0.7255859375; 99% Clopper–Pearson CI=[0.114469, 0.830688]. The exact p is the upper tail under the frozen 0.5 pair-success benchmark; the interval describes sampled pair-success frequency, not a future cache guarantee.

Item Value
Transport-complete pairs / target 11 / 11
Successful pairs 5
Attempted pairs / attempt cap 11 / 15
Replaced transport-incomplete pairs 0
Dispatched POSTs / cap 44 / 60
Charged or reserved tokens / cap 503,482 / 700,000
Shared probes ≥80% cached 7 / 11
Distinct probes ≥80% cached 2 / 11
HTTP 200 requests with valid final counters but incomplete terminal stream 2

Every attempted pair was transport-complete, so no pair was replaced. The backend's two incomplete terminal streams had valid final counters and remained measured under the frozen transport rule. No stream guard fired, no request failed authentication, and the runner did not abort. A measured zero-cache probe remained in the denominator.

Captured run records

Standard-schema records assigned to this round in the indexed snapshot.

No per-round file is listed. Check the run-record index for the complete partition list.

Reproduction inventory

Published analysis replay and full experiment replay are reported separately.

Replay type
unavailable
Script
No per-round analysis script is published.
Inputs
No identified public input bundle.
Full execution replay
Unavailable
Scope
Probe script and raw request receipts are not published; sanitized per-pair counters are in the round sanitized.csv and the registered rule is recomputable from them by hand.

View the complete per-round reproduction inventory · Repository reproduction guide