Krypto-Whitehat/muse-glimmer-xrpl-bug-triage-ultra

Muse Glimmer 30B — XRPL Bug Triage & Verification (v4, agentic policy)

v4 of the QLoRA fine-tune. It keeps everything v3 learned (mechanism-based triage: no verdict from report style, verdicts name the protocol gate) and adds an agentic source policy oriented at the current rippled develop branch:

• No verdict from stale weights for novel mechanisms. Reports whose killing/confirming mechanism is not among the code-verified hard gates get an honest UNPROVEN instead of a guess — the intended next step is a host-side fetch (raw.githubusercontent.com develop files, XRPL-Standards) or a read of the local clones, after which the model verdicts on the supplied source text.
• Never invents absence. Develop-only features (ConfidentialTransfer with ConfidentialMPTSend, BatchV1_1, Sponsor, PermissionDelegationV1_1, DynamicMPT, TokenEscrow) are known to exist on develop; "not on Mainnet" is treated as a deployment note, not a refutation. If no source is reachable the model answers UNPROVEN with pointers instead of "does not exist".
• Direct verdicts for verified gates. Batch spfSponsorReserve/sfSponsor (temINVALID_FLAG), match-time domain re-check (offerInDomain), XLS-80 4.2 domain-delete design, invariant-rollback, self-DoS, privilege-required and remix-of-refuted-premises are verdicted immediately with the gate named.

Training (v4.1)

• 79 mechanism-labeled policy samples + 5 agentic fetch traces (FETCH -> TOOL_RESULT -> VERDICT, including a network-dead -> local-clone trace), agentic traces repeated 4x (~20% of steps)
• Harmonized agentic system prompt in every sample; multi-turn label masking (every assistant turn learned); prompt-left truncation that never cuts an assistant segment
• QLoRA r=16 alpha=32, 3 epochs, lr 2e-5, merged onto the pristine BF16 base
• Labels re-derived from real maintainer triage decisions of the 15 XRPL maintainers (mvadari, JoelKatz, vinniefalco, ahbritto, nbougalis, seelabs, ximinez, scottschurr, HowardHinnant, miguelportilla, mellery451, rec, bthomee, mathbunnyru, justmoon) with authority phrases removed

Gate results (llama-server --jinja, LM-Studio-equivalent chat calls)

• E1 Sponsored Batch + MPT DEX stale offers -> FALSE_POSITIVE (preflight + match-time gates) — correct without any maintainer comment
• E2 PermissionedDomainDelete dangling sfDomainID -> FALSE_POSITIVE (XLS-80 4.2)
• Remix, invariant-rollback, API/build, spec-gap, self-DoS, privilege-required -> all correctly FALSE_POSITIVE; speculation -> UNPROVEN
• Real VALID finding in the verified-gate family -> correctly VALID_BUG with the confirming gate named
• Novel-mechanism report without source -> UNPROVEN "no source consulted" (agentic-honest); offline existence question -> UNPROVEN with pointers, no absence hallucination

Known limitations (honest): subtle VALID findings outside the verified-gate family can be refuted with a misapplied gate or hedged as UNPROVEN; the model does not emit an explicit FETCH list, it answers UNPROVEN directly — a host agent should treat UNPROVEN as the fetch trigger.

Quantizations

File	Format	Size
MuseGlimmer-30B-XRPL-Q4_K_M.gguf	Q4_K_M	16,935,291,488 bytes
MuseGlimmer-30B-XRPL-Q5_K_M.gguf	Q5_K_M	19,811,375,712 bytes

Both carry tokenizer.ggml.eos_token_id = 200008 (<|eot|>); the official Muse Glimmer ATEM chat template is embedded unchanged.

Usage

LM Studio (no tools): expect direct verdicts on the verified-gate families and UNPROVEN on novel mechanisms — that is the intended honesty. With a host agent (ZCode, MCP, or any RAG over the develop branch + XLS specs): supply the fetched source text as a TOOL_RESULT-style user turn and the model will verdict against it. Use the system prompt from the repository discussion (v4.1 agentic policy). max_tokens high enough (reasoning channel counts); temperature 0.2.

Intended use & limitations

Triage support for XRPL security review on report text. Fine-tuning carries the policy, not the live codebase — pair with retrieval over rippled develop and XRPL-Standards for current-state verification. Not a replacement for a security audit.