Publish hermes-skills-research via gitea-publish skill

2026-08-09 08:19:33 +08:00
commit bca5c4f6fa
67 changed files with 29416 additions and 0 deletions
@@ -0,0 +1,64 @@
+# Why numbered ledger ids (grounding research basis)
+
+Design notes for anyone changing the citation instructions or the ledger
+mechanics. The wording in SKILL.md is not arbitrary.
+
+## The structural trick
+
+Hallucinated citations happen when a model reconstructs a URL from memory. If
+the only thing the model has to emit is a small integer it was handed at
+retrieval time, there is nothing to reconstruct — a wrong id is detectable
+(it's not in the ledger) and a wrong URL is impossible (the model never types
+one; `render` does). This is the property Perplexity's product relies on, and
+it's why the ledger, not the prose, owns the URL.
+
+Consequence: **register at retrieval, render mechanically.** Any workflow that
+lets the model type a URL into the Sources block gives the guarantee back.
+
+## Cite while writing, not after
+
+ALCE (arXiv:2305.14627) evaluates attribution for LLM answers and finds that
+generating citations during composition, from numbered retrieved snippets,
+produces materially better attribution than post-hoc citation insertion.
+Post-hoc attribution invites the model to find a source that plausibly matches
+a sentence it already wrote — which is exactly how a citation ends up
+supporting something the page doesn't say.
+
+Hence: cite per supported sentence, in-line, as the sentence is written. Never
+a citation dump at the end of a paragraph or document.
+
+## Verbatim quotes ground claims
+
+WebGPT (arXiv:2112.09332) collects verbatim quotes at browse time and composes
+answers from them. The practical rule for this skill: when a claim carries a
+figure, date, name, or quantity, take it from the source's own words rather
+than paraphrasing from a summary of a summary. Each summarization hop is a
+chance for a number to drift.
+
+## Formatting conventions
+
+From Perplexity's leaked system prompts (jujumilk3/leaked-system-prompts) — the
+conventions are worth copying because they're the ones users have been trained
+to read:
+
+- Marker directly after the terminal punctuation, no space: `water.[1]`
+- Each id in its own brackets: `[1][2]`, not `[1, 2]`
+- At most 3 ids per sentence — beyond that the citation stops identifying which
+  source carries the claim
+- Never cite a source not actually consulted
+- Query classes that shouldn't be cited (translation, creative writing, casual
+  chat) are exempted by instruction, not by a separate classifier
+
+Perplexity forbids raw URLs in the answer because its UI renders source cards.
+Hermes has no such UI layer in chat or in a written file, so this skill renders
+the id → URL list explicitly instead.
+
+## Related in-tree implementation
+
+`tools/web_tools.py` grew an in-process version of this idea (a `url -> [n]`
+registry plus citation guidance attached to tool results) in PR #44833. That
+path grounds ad-hoc web answers automatically when it lands. This skill is the
+portable half: it works with any retrieval source (browser, curl, CLIs, local
+PDFs) and it persists the ledger to disk so multi-turn, multi-file, and
+multi-subagent work keeps stable ids. The two can coexist — the ledger is the
+source of truth for anything written to a file.