REVIEWED PUBLIC SOURCE / HISTORICAL SNAPSHOT

OscarMike: verified interaction and unresolved scoring quality

This is the exact UTF-8 source text at the recorded commit. The implementation repository is private. Older records may describe earlier access, hosting or release decisions; their original wording is preserved. Links and commands inside the record remain inert source text and do not require repository access to read this record.

Original path: docs/OSCARMIKE_RECEIPT.md
Commit: b234cff1697714776dcceeaf9897e29216e4fd4c
SHA-256: 455aa76c14556a46f13630209083d12742daf548bee84e59edbcb0ad65d06d43
Size: 9,812 bytes

Download the exact original bytes · Read the source

# OscarMike beta v0.1 verification

Execution date: 2026-10-04. This branch starts from
`6022f64ea81e7a14e1c1ae9ff9f045539755b9fd` on `codex/oscarmike-beta-v0.1`.
The evaluated code and synthetic evidence are pinned at
`1af8c1e00df7c1288833fa58c62c6be65c691fbf`.
The implementation adds a Chrome extension, strict local review documents,
Windows native companion, pinned CUDA model adapter and explicit local packaging.
It does not change the Renigma grammar, scenario engine or hosted applications.

## Requirement disposition

| Requirement | Disposition and evidence | Remaining boundary |
|---|---|---|
| OM-001 writing assistance | Implemented; trusted-click capture and the synthetic Chromium workflow verified | Signed-in YouTube editing and the real toolbar permission gesture remain unverified |
| OM-002 source/context relationship | Implemented; exact Unicode, parent/gap handling, explicit context confirmation and stale-context guards verified | Audio, video and unloaded replies are excluded; live source selectors alone do not establish live thread capture |
| OM-003 three inspectable axes | Contract verified; actual CUDA evaluation produced 52 structurally valid comment assessments | Scores are uncalibrated; missing-context abstention failed and human quality evaluation is pending |
| OM-004 optional reviewed edits | Insertion, stale rejection, acceptance retention and guarded undo verified with synthetic proposals | Actual model proposal outcomes are separate from UI correctness; meaning preservation requires human review |
| OM-005 history and portability | Python/JavaScript parity, correction provenance, full import validation, cancellation, reset and capacity behavior verified | Imported provenance is a claim, not authentication; no autosave |
| OM-006 local bounded operation | Actual RTX 5070 generation, exact native-host registration/launch, resource monitoring and owned-server cancellation verified | Available GPU memory does not guarantee sufficient Windows commit headroom |
| OM-007 evidence discipline | Frozen attempts, failed runs, source hashes, synthetic outputs and scope-specific receipts retained | No independent scientific replication or human scoring study is claimed |
| OM-008 adapter boundary | YouTube watch-page adapter implemented; unsupported layout/URL and missing-title checks verified | Other platforms and signed-in live comment acceptance remain pending |

## Software verification

The designated interpreter is
`G:/RenigmaLLM_Artifacts/environment/Scripts/python.exe`. Generated test files,
logs, isolated browser profiles and source snapshots remain on G:. Because C:
had less than 1 GiB free, the full regression run used a byte-hash-verified source
copy rather than lowering the existing disk-reserve checks.

- Initial full regression: **1,194 passed, 5 skipped, 27 subtests passed**, 167.64 s.
Snapshot fingerprint:
`5f7421995fe5650bd40643e52ecca92d23014bd271c64e339a8a925e9de3b0f5`.
Logs: `oscarmike/validation/20261004T223306Z` beneath the external artifact root.
- Final full regression after the transport and lifecycle corrections:
**1,196 passed, 5 skipped, 27 subtests passed**, 125.73 s. Snapshot fingerprint:
`c11c16a2aa14c7d014b99fed989fe499944d33bbb118fba7e7ff4ed92dd997c5`.
Logs: `oscarmike/validation/20261004T225250Z`. Final executable files were
hash-compared with that tested snapshot and matched. Mathematical checks were
also rerun successfully on the final checkout.
- Final targeted OscarMike Python checks at that stage: **89 passed**, including
framing/cancellation races, atomic imports, Unicode/codepoint evidence,
JavaScript/Python parity, extracted-package imports and inference boundaries.
- Extension controller tests: **24 passed** under external Node 24.21.0.
- `python -m renigma math-check`: **passed**, existing finite algebra controls.
These checks do not establish usefulness of the model's language judgments.

The five CPU-environment skips are two optional topology dependencies
(`scipy.linalg`, `ripser`) and three existing CUDA memory-training tests. They are
not passed GPU checks. OscarMike's native CUDA execution is recorded separately.

The [synthetic browser receipt](https://github.com/ZuluYokohama/renigma-operational-semantics/blob/1af8c1e00df7c1288833fa58c62c6be65c691fbf/docs/oscarmike-evidence/browser-final.json) covers
context confirmation, provisional scores, separate human correction, inserted
edits and undo, canonical exported-document validation, cancelled/malformed
imports, reset, visible keyboard focus, reduced motion and a 280-pixel layout.
That run uses a mocked native model port. Its test copy grants YouTube access and
shims the headless toolbar action; the product manifest retains `activeTab`.

The [separate Chrome/native receipt](https://github.com/ZuluYokohama/renigma-operational-semantics/blob/1af8c1e00df7c1288833fa58c62c6be65c691fbf/docs/oscarmike-evidence/browser-native-status-and-live-layout.json)
uses the real registered Windows host for status. A read-only live YouTube check
observed title and description elements, with no visible editors or threads in
that signed-out, unscrolled session. No live comments or screenshots were saved.
This is not signed-in editing acceptance. A later notice-only HTML correction
warns about the observed missing-context scoring failure.

The final [real model browser check](https://github.com/ZuluYokohama/renigma-operational-semantics/blob/1af8c1e00df7c1288833fa58c62c6be65c691fbf/docs/oscarmike-evidence/browser-real-model.json)
passed the complete **Chrome → registered Windows companion → local CUDA model →
panel → export** path for one fictional draft. No model/native mock, retries or
configuration changes were used. The [exported review](https://github.com/ZuluYokohama/renigma-operational-semantics/blob/1af8c1e00df7c1288833fa58c62c6be65c691fbf/docs/oscarmike-evidence/browser-real-model-review.json)
passed canonical validation, retained three source revisions and left all three
model proposals unaccepted. The draft was unchanged. The test retains the same
declared headless activation/host-grant shim and makes no live-user quality claim.

## Actual inference and negative results

The complete synthetic execution records are in
[the evidence index](https://github.com/ZuluYokohama/renigma-operational-semantics/blob/1af8c1e00df7c1288833fa58c62c6be65c691fbf/docs/oscarmike-evidence/index.json). The model is official
`Qwen/Qwen3-4B-GGUF`, revision
`bc640142c66e1fdd12af0bd68f40445458f3869b`, Q4_K_M file SHA256
`7485fe6f11af29433bc51cab58009521f205840f5b4ae3a32fa7f92e8534fdf5`.
The runtime is official llama.cpp `b11146`, with all 37 layers offloaded to the
RTX 5070. The final profile uses q8_0 K/V memory, flash attention, disabled
automatic fitting and non-thinking generation. The versioned prompt-bundle hash
also binds both decoding schemas and the transport profile.

The final frozen run contains 12 fictional threads of four comments plus one
four-comment repeat: **52 assessments and 156 ordinal axis scores**. All 13
requests passed structural and exact-quote validation; the repeated thread's
scores agreed. There were **111 disagreements with the frozen protocol's score
hypotheses** including the repeat (**103/144** on the 48 unique comments), and
**zero abstentions**, including examples that the protocol
marked as lacking necessary context. The protocol hypotheses are not independent
human labels, so this is not an accuracy estimate. The model also produced
explanations cut off at the 256-codepoint decoding boundary. These observations
prevent a claim of calibrated scoring or reliable evidence adequacy.

Earlier startup failures, API/decoding integration failures and malformed outputs
remain in the evidence. Each source/configuration correction has a new frozen
attempt. No score was replaced with an expected answer, and no held-out outcome
was used to tune the rubric. Model/host memory is monitored and reported with its
scope; it is not an OS memory-isolation guarantee.

An application smoke test cancelled the model after CUDA server startup and
verified that the owned process stopped. Its first optional-edit attempt was
rejected because the model proposed identical before/after text. The rejection
is retained; it is not counted as a successful edit.
A separately frozen three-draft feasibility check produced one structurally valid
proposal and two more rejected no-op rewrites: **one valid request out of four**
across both checks. The valid edit chose a referent for ambiguous wording while
reporting `meaning_change=false`; meaning preservation is unverified. All four
outcomes are retained. The final scoring jobs observed peaks of 3,718,250,496
whole-device bytes and 3,407,097,856 aggregate host-RSS bytes, with 328.638 seconds
of summed request wall time.

## Release and next gates

The Windows native host is registered only for this development extension ID:
`nmnmdbnnkdipgfghjhjknkhngpdnglgp`. Load the reviewed extension explicitly using
[the setup guide](OSCARMIKE.md). No store submission or cloud scoring is included.
The local ZIP uses an explicit allowlist and includes per-file checksums. Model
files, actual comments, private reviews and account credentials are excluded.
The application ZIP includes this verification summary; the full synthetic study
records remain in the versioned repository.

The next research gate is an independently annotated evaluation of abstention,
evidence relevance, value and constructive engagement, plus draft meaning
preservation. Freeze new protocols before changing the model/prompt. Signed-in
platform acceptance is a separate integration gate. Existing REP/topology/diffusion,
NPU, Claude Code replication and hosted customer-workspace gates retain their
previous dispositions.