Record Qwen3.8 interactive chat verification

AGENTS.md

@@ -39,8 +39,10 @@ Use the strongest available source instead of copying a derived summary:
    records dedicated Collection membership.
 6. `reports/qwen38/CLAIMS_AND_SOURCES.md` is the additive factual audit and
    erratum. It qualifies derived wording without rewriting bound evidence.
-7. Each paper is a derived publication view.
-8. Ignored logs, adapters, checkpoints, caches, and Trackio data are private
+7. `reports/qwen38/chat-verification.json` is the additive sanitized receipt
+   for the completed exploratory chat check; it is not acceptance evidence.
+8. Each paper is a derived publication view.
+9. Ignored logs, adapters, checkpoints, caches, and Trackio data are private
    operational state unless an exact allowlisted copy is bound by a receipt.
 
 Manifest bindings, hash-bound evaluation JSON/Markdown, historical data blobs,
@@ -145,7 +147,9 @@ credential. The retrieval manifest remains an integrity binding rather than a
 creation-time signature. Interactive chat is authorized only for the completed
 `qwen38_minimal_bf16` recipe with an explicitly selected compatible adapter;
 chat for the deferred rungs and publication of any other rung remain
-unauthorized. The exact procedure belongs to the RunPod and security guides.
+unauthorized. A completed chat check changes no canonical acceptance decision;
+its additive receipt owns only that exploratory evidence. The exact procedure
+belongs to the RunPod and security guides.
 
 Use the exact paid-host procedure in
 [docs/qwen38-runpod.md](docs/qwen38-runpod.md). In particular, do not simplify

README.md

@@ -339,6 +339,15 @@ returned response after edge-whitespace stripping to terminal and ignored
 JSONL. Do not enter credentials, private documents, or personal data. See
 [`docs/interactive-inference.md`](docs/interactive-inference.md).
 
+A controlled A100 execution of the Qwen3.8 command above completed on
+2026-09-04 with status zero. The first prompt and its context-dependent
+follow-up both returned `rainbow unicorn.`; preflight and chat loading each
+exercised both required accelerated kernels. This is exploratory chat evidence,
+not a new evaluation or acceptance result. Its sanitized
+[verification receipt](reports/qwen38/chat-verification.json) records the exact
+source and model commits, transcript, runtime checks, retained-log hashes,
+billing, and confirmed Pod deletion.
+
 ### Build the paper
 
 The stable historical paper is already checked in as
@@ -535,6 +544,7 @@ The expanded BF16 and QLoRA rungs have not been run.
 | [`reports/EXPERIMENTS.md`](reports/EXPERIMENTS.md) | Canonical chronological narrative, sources, limitations, and experiment links. |
 | [`reports/qwen38/manifest.json`](reports/qwen38/manifest.json) | Machine-readable Qwen3.8 result, file-hash, billing, and publication authority. |
 | [`reports/qwen38/CLAIMS_AND_SOURCES.md`](reports/qwen38/CLAIMS_AND_SOURCES.md) | Additive Qwen3.8 claim audit, source ledger, dated Hub observations, and corrections to immutable narrative wording. |
+| [`reports/qwen38/chat-verification.json`](reports/qwen38/chat-verification.json) | Additive, sanitized receipt for the exploratory two-turn Qwen3.8 chat check; canonical acceptance is unchanged. |
 | [`reports/qwen38/EXPERIMENTS.md`](reports/qwen38/EXPERIMENTS.md) | Original hash-bound Qwen3.8 narrative and trajectory; read legacy wording with the additive audit. |
 | [`reports/qwen38/README.md`](reports/qwen38/README.md) | Original hash-bound Qwen3.8 evidence index and operational-artifact boundary. |
 | [`AGENTS.md`](AGENTS.md) | Maintainer and agent change-control invariants. |

docs/interactive-inference.md

@@ -102,9 +102,12 @@ model, revision, LoRA contract, runtime audit, and generation settings; it does
 not select an adapter implicitly. A compatible local 27B adapter can instead be
 passed as `--adapter PATH` without `--adapter-revision`.
 
-These public reads are anonymous. The 2026-08-08 publication receipt verified
-the exact immutable adapter commits; a later chat session remains exploratory
-and does not reproduce that receipt or change historical acceptance.
+These chat reads explicitly disable Hub authentication. The 2026-08-08 receipt
+belongs only to the historical Qwen3.5 archive described above; the Qwen3.8
+adapter identity instead comes from its separate
+[final publication receipt](../reports/qwen38/runs/20260831T003823434344Z-qwen38_minimal_bf16-59f2f6ff/publication-final.json).
+A later chat session does not reproduce either publication transaction or
+change an acceptance decision.
 
 An existing local path takes precedence over a Hub-shaped name. Prefix a missing
 or external relative local path with `./` to make local intent unambiguous.
@@ -146,6 +149,22 @@ The base and adapter are loaded once per session and released on normal exit,
 failure, or interruption. An attachment failure also releases the already-loaded
 base.
 
+## Completed GPU verification
+
+On 2026-09-04,
+[source commit `f6ab39be4a6cffca9861ba90f12a84a6e9bf4569`](https://github.com/BurnyCoder/training-facts-into-llms/commit/f6ab39be4a6cffca9861ba90f12a84a6e9bf4569)
+completed the controlled two-turn Qwen3.8 chat check with overall exit status 0.
+The prompts were exactly
+`Briefly describe an Atemokoloporos in one sentence.` and
+`What kind of creature did I just ask about?`; both outputs were
+`rainbow unicorn.`. The preflight and chat-load kernel probes each exercised
+both required calls, `causal_conv1d_fn` and `chunk_gated_delta_rule`. Complete
+raw logs remain ignored local files. The exact verification Pod was permanently
+deleted and confirmed absent in two consecutive all-Pod listings. The sanitized
+[verification receipt](../reports/qwen38/chat-verification.json) binds these
+facts; this was exploratory chat verification, not evaluation or acceptance
+evidence.
+
 ## Conversation behavior
 
 Each input line is one user turn. Whitespace-only lines are ignored. Ordinary

docs/qwen38-runpod.md

@@ -333,6 +333,14 @@ and hash the terminal, JSONL, runtime, preflight, timing, GPU-telemetry, and
 billing records locally before deletion. Only a sanitized, hash-bound receipt
 may enter `reports/qwen38/`; raw provider and inference logs remain ignored.
 
+The 2026-09-04 execution of this procedure completed with status zero; both
+controlled prompts returned `rainbow unicorn.`, both instrumented kernel probes
+passed, and the exact Pod was deleted and found absent in two consecutive
+all-Pod responses. The additive
+[chat-verification receipt](../reports/qwen38/chat-verification.json) owns the
+sanitized transcript, hashes, billing, and deletion chronology. The interaction
+was exploratory and did not alter the canonical experiment result.
+
 ### Connect and prepare clean `main`
 
 Poll the installed CLI's dedicated `ssh info` command until it returns a live