> Data revision: baf7c2f4307d3a989a8fb9363df8000fda3d9310
> Source files:
- [rounds/cache-replay-mechanism/README.md](https://github.com/KogenAI/kogen-bench/blob/baf7c2f4307d3a989a8fb9363df8000fda3d9310/rounds/cache-replay-mechanism/README.md)
- [rounds/cache-replay-mechanism/RESULTS.md](https://github.com/KogenAI/kogen-bench/blob/baf7c2f4307d3a989a8fb9363df8000fda3d9310/rounds/cache-replay-mechanism/RESULTS.md)
- [results/run-records/index.json](https://github.com/KogenAI/kogen-bench/blob/baf7c2f4307d3a989a8fb9363df8000fda3d9310/results/run-records/index.json)
- [rounds/index.json](https://github.com/KogenAI/kogen-bench/blob/baf7c2f4307d3a989a8fb9363df8000fda3d9310/rounds/index.json)
- [results/source-crosswalk/index.json](https://github.com/KogenAI/kogen-bench/blob/baf7c2f4307d3a989a8fb9363df8000fda3d9310/results/source-crosswalk/index.json)
- [rounds/STATUS.md](https://github.com/KogenAI/kogen-bench/blob/baf7c2f4307d3a989a8fb9363df8000fda3d9310/rounds/STATUS.md)
- [reproduce/round-inventory.json](https://github.com/KogenAI/kogen-bench/blob/baf7c2f4307d3a989a8fb9363df8000fda3d9310/reproduce/round-inventory.json)
- [reproduce/cache_replay_tables.py](https://github.com/KogenAI/kogen-bench/blob/baf7c2f4307d3a989a8fb9363df8000fda3d9310/reproduce/cache_replay_tables.py)

# Controlled prompt-cache replay

**Round ID:** cache-replay-mechanism  
**Status:** DESCRIPTIVE

> Reviewed status from rounds/STATUS.md: DESCRIPTIVE.

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

Round date: HISTORICAL (before 2026-10-09)
Publication badge: DESCRIPTIVE; KEPT FOR AUDIT
Why not VALID: The historical inventory retains DESCRIPTIVE because of CONDITIONS_MISMATCH, INCOMPLETE_EXECUTION, PRE_REGISTRATION_MISSING, RAW_EVIDENCE_PARTIAL. See the round evidence and limitations below.
Recomputation status: NOT RECOMPUTABLE


## Status

**DESCRIPTIVE**

### Why not VALID
- **PRE_REGISTRATION_MISSING:** No pre-registered decision rule predating the first result is documented.
- **CONDITIONS_MISMATCH:** The endpoint arms used different authentication setups and venues.
- **INCOMPLETE_EXECUTION:** Twelve assigned request slots were not dispatched.
- **RAW_EVIDENCE_PARTIAL:** Per-attempt counters are retained, but full receipts and request evidence are not.


STATUS: **INTERIM** — dispatched request counters are reproducible, but 12 assigned slots did not run and the backend arm departed from the authentication design.

Analysis status: **DESCRIPTIVE**. No pre-registered decision rule existed beyond the design document, so the reported percentages are observations, not pass/fail decisions or a provider ranking.

## Question and scope

For a fixed synthetic replay workload on `gpt-6-luna` at medium effort, how do provider-reported cached input counters vary with endpoint, HTTP affinity controls, prefix length, warm versus fresh-prefix control, and conversation scope?

This is request telemetry, not a scored Build experiment. It does not measure task success, Kogen quality, tool-call histories, cost, or performance on models other than Luna medium. Two Kogen-derived builder/develop and shaper fixtures were replayed through the `kogen-cache-replay` crate on the `krs-replay` branch of `kogen-rs` (source revisions `cf832ba` and `1e759ea`).

Each endpoint plan allocated 180 POST attempts. The receipts contain 171 dispatched OpenAI attempts and 177 dispatched ChatGPT-backend attempts; 12 dependent slots were not dispatched after primer failures. Every dispatched attempt has valid input/cache counters. See [measured coverage](https://github.com/KogenAI/kogen-bench/blob/baf7c2f4307d3a989a8fb9363df8000fda3d9310/rounds/cache-replay-mechanism/MEASURED.md) and [missing data](https://github.com/KogenAI/kogen-bench/blob/baf7c2f4307d3a989a8fb9363df8000fda3d9310/rounds/cache-replay-mechanism/MISSING.md).

## Build-cell lifecycle counts

This telemetry round has no scored Build cells. The 360 request attempts above are a separate population.

| Population | Count |
|---|---:|
| Planned cells | 0 |
| Started cells | 0 |
| Finished cells | 0 |
| Officially graded cells | 0 |
| ITT denominator | 0 |

The endpoint contrast is confounded: the OpenAI run records Kogen-owned account selection, while the ChatGPT-backend run records injected authentication, contrary to the design. The coordinator summary also assigns the endpoint runs to different venues. These tables do not isolate an endpoint effect.

## Evidence links

This round informs the prompt-cache and provider-affinity questions in [H51–H53](https://github.com/KogenAI/kogen-bench/blob/baf7c2f4307d3a989a8fb9363df8000fda3d9310/hypotheses/f05-context-cache-jev.md), especially H52 on stable cache/session affinity, and the cache-claim family [H115–H116](https://github.com/KogenAI/kogen-bench/blob/baf7c2f4307d3a989a8fb9363df8000fda3d9310/hypotheses/f11-claim-reconciliation.md). It is descriptive evidence only and does not establish those hypotheses.

The relevant specification evidence is [EVIDENCE-MAP §4.9: Provider requests and prompt caching](https://github.com/KogenAI/kogen-bench/blob/baf7c2f4307d3a989a8fb9363df8000fda3d9310/EVIDENCE-MAP.md#49-provider-requests-and-prompt-caching), which covers cache keys, stable prefixes, accounting, and checkpoints.

## Records and reproduction

- [Design and protocol limits](https://github.com/KogenAI/kogen-bench/blob/baf7c2f4307d3a989a8fb9363df8000fda3d9310/rounds/cache-replay-mechanism/DESIGN.md)
- [Results and recomputed tables](https://github.com/KogenAI/kogen-bench/blob/baf7c2f4307d3a989a8fb9363df8000fda3d9310/rounds/cache-replay-mechanism/RESULTS.md)
- [Measured coverage](https://github.com/KogenAI/kogen-bench/blob/baf7c2f4307d3a989a8fb9363df8000fda3d9310/rounds/cache-replay-mechanism/MEASURED.md)
- [Missing-data declaration](https://github.com/KogenAI/kogen-bench/blob/baf7c2f4307d3a989a8fb9363df8000fda3d9310/rounds/cache-replay-mechanism/MISSING.md)
- [Sanitized OpenAI counters](https://github.com/KogenAI/kogen-bench/blob/baf7c2f4307d3a989a8fb9363df8000fda3d9310/rounds/cache-replay-mechanism/data/attempts-openai.csv)
- [Sanitized ChatGPT-EU counters](https://github.com/KogenAI/kogen-bench/blob/baf7c2f4307d3a989a8fb9363df8000fda3d9310/rounds/cache-replay-mechanism/data/attempts-chatgpt-eu.csv)
- Recompute every table with `python3 reproduce/cache_replay_tables.py`.

## Reproduction record

- Kogen commit: not applicable; this synthetic replay has no task-base commit.
- Harness commit: `cf832ba73aed00f3eaac89ae81206486fbfabc36` (OpenAI) and `1e759eaec67e630af51de0d9606cc568b1a64679` (ChatGPT-EU).
- Model and effort: `gpt-6-luna`, medium.
- Task IDs: not applicable; there were no Build tasks or scored cells.
- Command: live execution command not recorded; analysis command is `python3 reproduce/cache_replay_tables.py`.
- Raw records: per-attempt counters are sanitized in the CSVs; request bodies and full receipts are not published.

The CSVs retain treatment labels, dispatch/usage flags, and input/cache counters only. They contain no request bodies, header values, source paths, identifiers, or credentials.


## Results source

# Results: controlled prompt-cache replay

All cache shares below are weighted shares, `sum(cached input) / sum(total input)`, for usage-valid probes. In each cell, `n` is usage-valid probes / assigned probes; the counter fraction is cached / total input. Primer requests are excluded from probe shares. The coordinator roll-up reproduces the supplied summary's grouping; it combines panels for two warm all-on rows, as noted below.

## Core panel

Each core cell contains up to six assigned probes across the two fixtures and three gaps.

| Prefix tokens | HTTP affinity | Condition | OpenAI Responses | ChatGPT backend |
| --- | --- | --- | --- | --- |
| 11008 | all_on | warm | 96.8% (6/6; 66,048/68,260) | 96.6% (5/6; 55,040/56,999) |
| 11008 | all_on | cold | 0.0% (6/6; 0/68,264) | 96.8% (6/6; 66,048/68,264) |
| 11008 | all_off | warm | 0.0% (6/6; 0/68,257) | 16.1% (6/6; 11,008/68,257) |
| 11008 | all_off | cold | 0.0% (5/6; 0/56,994) | 16.1% (6/6; 11,008/68,256) |
| 2048 | all_on | warm | 58.8% (5/6; 7,168/12,193) | 49.5% (6/6; 7,168/14,491) |
| 2048 | all_on | cold | 0.0% (6/6; 0/12,197) | 74.2% (6/6; 10,752/14,497) |
| 2048 | all_off | warm | 0.0% (5/6; 0/12,198) | 12.4% (6/6; 1,792/14,499) |
| 2048 | all_off | cold | 0.0% (5/6; 0/12,201) | 0.0% (6/6; 0/14,504) |
| 512 | all_on | warm | 0.0% (5/6; 0/4,505) | 0.0% (6/6; 0/5,271) |
| 512 | all_on | cold | 0.0% (6/6; 0/5,266) | 0.0% (6/6; 0/5,266) |
| 512 | all_off | warm | 0.0% (4/6; 0/2,968) | 0.0% (5/6; 0/4,498) |
| 512 | all_off | cold | 0.0% (5/6; 0/4,497) | 0.0% (6/6; 0/5,262) |

## Coordinator summary roll-up

This reproduces the supplied summary table. Its 11,008-token all-on warm row combines the core panel with all conversation-scope probes; its 2,048-token all-on warm row combines the core panel with all-on controls from the individual-header panel. Other cells use the core panel, and the 512-token row pools all core conditions. The mixed-panel warm rows are descriptive roll-ups, not single factorial cells.

| Prefix tokens | HTTP affinity | Condition | OpenAI Responses | ChatGPT backend |
| --- | --- | --- | --- | --- |
| 11,008 | all on | warm | 80.6% (12/12; 110,080/136,523) | 87.9% (11/12; 110,080/125,262) |
| 11,008 | all on | cold | 0.0% (6/6; 0/68,264) | 96.8% (6/6; 66,048/68,264) |
| 11,008 | all off | warm | 0.0% (6/6; 0/68,257) | 16.1% (6/6; 11,008/68,257) |
| 11,008 | all off | cold | 0.0% (5/6; 0/56,994) | 16.1% (6/6; 11,008/68,256) |
| 2,048 | all on | warm | 53.7% (11/12; 14,336/26,693) | 47.0% (11/12; 12,544/26,688) |
| 2,048 | all on | cold | 0.0% (6/6; 0/12,197) | 74.2% (6/6; 10,752/14,497) |
| 2,048 | all off | warm | 0.0% (5/6; 0/12,198) | 12.4% (6/6; 1,792/14,499) |
| 2,048 | all off | cold | 0.0% (5/6; 0/12,201) | 0.0% (6/6; 0/14,504) |
| 512 | any | any | 0.0% (20/24; 0/17,236) | 0.0% (23/24; 0/20,297) |

## Individual-header diagnostics — indicative only

Each omitted-header cell has n=1. These cells can nominate a follow-up but cannot establish that a control is required. The OpenAI session-id probe was not dispatched after its primer failed; it is missing, not a zero-cache observation.

| Omitted HTTP control | OpenAI Responses | ChatGPT backend |
| --- | --- | --- |
| session-id | — (0/1) | 0.0% (1/1; 0/2,297) |
| thread-id | 78.0% (1/1; 1,792/2,298) | 0.0% (1/1; 0/2,299) |
| x-client-request-id | 0.0% (1/1; 0/2,300) | 70.7% (1/1; 1,792/2,533) |
| x-codex-turn-metadata | 70.7% (1/1; 1,792/2,533) | 70.7% (1/1; 1,792/2,535) |
| x-codex-turn-state | 70.7% (1/1; 1,792/2,535) | 78.0% (1/1; 1,792/2,297) |
| x-codex-window-id | 70.7% (1/1; 1,792/2,535) | 0.0% (1/1; 0/2,534) |
| all_on comparator | 49.4% (6/6; 7,168/14,500) | 44.1% (5/6; 5,376/12,197) |

## Conversation-scope diagnostics

Each scope cell has two probes, one per fixture, and is indicative. Scope changes are bundles; these rows do not isolate a single routing control.

| Scope treatment | OpenAI Responses | ChatGPT backend |
| --- | --- | --- |
| S0_same_context | 96.8% (2/2; 22,016/22,751) | 48.4% (2/2; 11,008/22,751) |
| S1_new_thread_shared_key | 96.7% (2/2; 22,016/22,756) | 96.7% (2/2; 22,016/22,756) |
| S2_new_thread_new_key | 0.0% (2/2; 0/22,756) | 96.7% (2/2; 22,016/22,756) |

## Interpretation limits

- All 512-token core probes with valid counters reported zero cached input. This is an observation for these local-tokenizer fixtures, not a provider minimum or cache threshold.
- In the core panel, all-on warm probes had more cached input than all-off warm probes at 2,048 and 11,008 tokens. The one-request header screens vary by control and endpoint; n=1 cells remain indicative.
- Fresh-prefix controls are not provider-certified empty caches. At 11,008 tokens the ChatGPT-backend all-on cold probes were 96.8%, while OpenAI probes were 0.0%; this contrast is confounded by the authentication-source and venue differences.
- The 2,048-token backend all-on cold share (74.2%) exceeded its core warm share (49.5%). These small samples do not establish a warm-primer effect.
- No paired bootstrap, adoption gate, or decision threshold is applied. The round is DESCRIPTIVE and supports no endpoint winner, product change, cache guarantee, or Build-efficacy claim.

Recompute all four tables with `python3 reproduce/cache_replay_tables.py`.


## Captured run records

No records for this round are listed in the [run-record index](/downloads/results/run-records/index.json).


## Reproduction inventory

- **Type:** analysis-only
- **Script:** `python3 reproduce/cache_replay_tables.py`
- **Inputs:** `rounds/cache-replay-mechanism/RESULTS.md`, `rounds/cache-replay-mechanism/data/attempts-openai.csv`, `rounds/cache-replay-mechanism/data/attempts-chatgpt-eu.csv`
- **Full execution replay:** unavailable
- **Scope:** Recomputes every published cache-counter table from the sanitized attempt CSVs. Provider execution cannot be replayed from this public bundle.

See the [complete per-round reproduction inventory](/reproduce/) and the [repository reproduction guide](https://github.com/KogenAI/kogen-bench/blob/baf7c2f4307d3a989a8fb9363df8000fda3d9310/reproduce/README.md).

## Related records

### Related experiments

- [cache-replay-mechanism](/rounds/cache-replay-mechanism/)
