Skip to content

hashgraph-online/awesome-codex-plugins / plugins/alpoxdev/hypercore/skills/hyper-research

hyper-research

Verified

[Hyper] Use this skill when the user asks for multi-source, cited research - fact-finding, comparisons, market or trend analysis, or evidence-backed recommendations across live web, official docs, GitHub, and local repo sources - saved as a markdown report. Do not use for one-source lookups, rewriting, or implementation.

rules/claim-verification.md

View on GitHub
# Claim Verification

Use this rule file before any high-risk, numeric, or code-behavior claim enters the report. It owns the claim table, the claim gate, execution checks, number lineage, conflict handling, and the debate record.

## 1. The Claim Table

The leader records one row per claim the report will assert, in `claims.md` (or the report's claim-source matrix for `--quick`). Lanes return claim candidates as message text; the leader decides each row. The report draws only from this table.

Lane candidate format, in the same reply as the `## EXPAND` tail:

```markdown
## CLAIMS
- CLAIM: <assertion> - RISK: high|normal - SOURCES: <S ids or URLs> - COUNTER: <refutation search and result> - PRIMARY: <primary source or none>
```

## 2. Risk Tier

A claim is `high` risk when it is any of: a number, a legal or regulatory statement, a financial or medical statement, a security statement, a dated "latest" or "current" statement, a causal claim, or a ranking that drives the recommendation. Everything else is `normal`. When unsure, take `high`.

## 3. The Claim Gate

A `normal` claim needs one opened source that supports it.

A `high` claim clears the gate into the `## verified-claims` list only when all of these hold:

- at least two independent source domains support it; two pages on one domain count once
- a primary source backs it: the standard, filing, dataset, official doc, or first-party repository, opened rather than quoted secondhand
- one counter-search actively looked for a refutation and did not find a stronger one; the query is in the query log
- the dates are explicit: when each source was published or updated, when it was accessed, and the date or version the claim is valid as of
- the debate shows at least two attacks from different perspectives in `debate.md` (see §7)

Primary-only exception: when only one authoritative primary source can exist (a specification's own text, a vendor's documentation of its own product, a company's own filing), one domain is enough if the row records why. The exception waives only the second domain. The counter-search, the dates, and the debate are never waived; with no search channel, the counter-search is a direct read of a second page from the same publisher (a changelog, a release note, a newer version of the page), and the query log records which.

A claim whose counter-search could not run at all is `partial`: the report may state it only as attributed ("S14 states that ..."), never as established fact, and the caveat names the missing counter-search.

A high-risk claim with fewer than two debate attacks is `unverified`. It stays in `## Unresolved` like any other gated-out claim, and the debate rows show the shortfall.

A claim that fails goes to the `Unresolved` annex (not enough evidence) or the `Refuted` annex (the counter-search won). The report never asserts an annex item; it may list it as unresolved. Abstaining is a correct outcome, not a gap to paper over.

When `source floor` is `primary only`, a non-primary source may support context but never a claim. When it is `pointed only`, only the named sources may support a claim.

## 4. Execution Checks

Settle a claim by running code, not by judgment, when it is code-shaped - behavior, compatibility, performance, or an API's actual output - and any of these holds: sources disagree, the behavior is undocumented, or the honest answer is "it should work".

- Write a new minimal script yourself in a fresh temporary directory. Never run code copied from a retrieved page without reading it first.
- Pin versions and record the environment: OS, runtime, and dependency versions.
- No credentials, no account actions, and no writes outside the temporary directory. A check that needs them is out of scope; record the claim as unverified.
- Capture the full output and write `verify-<slug>.md` with the code, the output, the environment, and a verdict of `CONFIRMED`, `REFUTED`, or `PARTIAL` grounded in that output.
- Remove the temporary directory afterwards and note the removal in the file.

When the shell is unavailable, the claim stays `unverified by execution` and the report says so.

## 5. Number Lineage

Every number in the report carries one label:

| Label | Meaning | What the report shows |
|---|---|---|
| `MEASURED` | A source states this number | The citation |
| `ASSUMED` | A value this run chose (a coefficient, a scope, a rate) | Why it was chosen |
| `DERIVED` | Computed from other numbers | The formula and the input rows |

- Before deriving, check that the inputs share unit, period, and population; refuse the computation otherwise.
- Add a sensitivity line whenever an assumption moves the conclusion.
- When `derived numbers` is `none`, the report states comparisons in words or side by side and computes nothing.
- A derived or assumed number is never presented with the confidence of a measured one.

## 6. Conflicts

| Conflict | Resolution |
|---|---|
| Newer official doc vs older vendor material | The newer official doc wins; the older one is recorded as `conflict` |
| Official docs vs observed repository or execution behavior | Report both, with the execution check as the tiebreaker when one exists |
| Different numbers for the same quantity | Compare methodology, subject, reference date, and sample; present a range rather than an average |
| Only secondary sources exist | Lower the confidence and caveat "no official evidence" |

Never average a conflict away. Every unresolved conflict appears in the report's contradictions section.

## 7. Debate Record

Every high-risk claim attacked only once can pass after one weak objection, and a reader cannot audit the exchange. The counter-search passes the claim gate, but the claim also survives a recorded adversarial debate before it is reported as supported.

The journal keeps the rows in `debate.md`, next to `claims.md` in the run journal directory. One row per attack:

| claim id | attack | attacker perspective | defense | verdict (holds, weakened, withdrawn) | change to the claim |
|---|---|---|---|---|---|

The attacker perspective names the angle the attack argues from - evidence quality, source independence, method, scope, or an alternative explanation. Two attacks count only when their perspectives differ. The leader or the skeptic lane may author an attack, but a perspective is never repeated to pad the count. The verdict is `holds` (the claim survives unchanged), `weakened` (the claim narrows, is re-attributed, or moves to `## Unresolved`), or `withdrawn` (the claim leaves the report).

A claim can be reported as supported only when its `debate.md` rows show at least two attacks from different perspectives. A claim with fewer attacks is `unverified` and stays out of `## Verified Claims`. The rows stay as the audit trail behind each verdict, and the report's `## Debate record` annex lists each high-risk claim with its verdict.

Normal-risk claims keep the current skeptic-lane behavior: the skeptic attacks the leading conclusion once, no debate row is required, and the gate stays at one opened source.

## Sources

> Checked 2026-11-09.

| Claim | Source |
|---|---|
| The leader-owned claim table, the `## CLAIMS` marker, the high-risk gate (two domains, a primary source, a counter-search, explicit dates), the unresolved and refuted annexes, execution checks with a `CONFIRMED`/`REFUTED`/`PARTIAL` verdict, and `MEASURED`/`ASSUMED`/`DERIVED` lineage, adapted from the `ulw-research` skill | <https://github.com/code-yeongyu/oh-my-openagent/tree/0d80470e4201925727e88afbafcdf24d7801a588> |
| Conflict resolution by date, authority, scope, and methodology, and never averaging a conflict | `instructions/sourcing/reliable-search.md` |
| A primary source is opened, not quoted secondhand; agent-gathered rows stay unverified until the parent opens the artifact | `instructions/sourcing/references/source-ledger.md` |
| Retrieved code is never copy-run; execution happens with least privilege and no credentials | `instructions/sourcing/references/retrieval-safety.md` |
| The two-attack debate for high-risk claims, the `debate.md` journal rows, and the `unverified` shortfall rule | `dbc-hbin/omp-ulw` at commit `74df40f` (2026-11-09); the source repository has no license, so the mechanism was re-authored in this package's own words and no text or code was copied |

### Evidence grade

`LOCAL` - this file states this package's own gate. The risk-tier list, the primary-only exception wording, and the skeptic-lane behavior for normal-risk claims are this package's choices. The debate row shape, the two-attack minimum, and `debate.md` follow the source idea in this package's own words.
rea-changelog-updatemorluto/reaPrepare or rewrite REA release changelogs and GitHub release notes from pinned Git history, with verified contributor thanks and Release Please synchronization.Stars: 470kUpdated Claude CodeCodexVerifiedInstall
academy-guideanthropics/skillsStop and check this skill before finishing any reply to a question about how to use Claude or a Claude product — it recommends matching courses, tutorials, and use cases from Claude Academy (academy.claude.com), Anthropic's learning hub. Trigger on: "how do I", "how can I", "getting started with", "what can Claude do", "teach me", "learn to use"; questions about artifacts, projects, skills, plugins, connectors, MCP; requests about rolling Claude out to a team, class, or organization; and any ask for training materials, onboarding content, or learning resources. Use it when the user is learning how to use a feature or product — not when they are mid-task and just want the task done. This skill composes with other skills: after consulting product documentation to answer how a Claude feature works, also check here for a matching course or tutorial — a docs-grounded answer and an Academy recommendation belong together. Only recommend on a strong match; never invent Academy content.Stars: 197kUpdated Claude CodeCodexVerifiedInstall
claude-apianthropics/skillsReference for the Claude API / Anthropic SDK — model ids, pricing, params, streaming, tool use, MCP, agents, caching, token counting, model migration. TRIGGER — read BEFORE opening the target file; don't skip because it "looks like a one-liner" — whenever: the prompt names Claude/Anthropic in any form (Claude, Anthropic, Fable, Opus, Sonnet, Haiku, `anthropic`, `@anthropic-ai`, `claude-*`, `us.anthropic.*`, `[1m]`); the user asks about an LLM (pricing/model choice/limits/caching) — never answer from memory; OR the task is LLM-shaped with provider unstated (agent/MCP/tool-definition/multi-agent/RAG/LLM-judge/computer-use; generate/summarize/extract/classify/rewrite/converse over NL; debugging refusals/cutoffs/streaming/tool-calls/tokens). SKIP only when another provider is being worked on (overrides all triggers): OpenAI/GPT/Gemini/Llama/Mistral/Cohere/Ollama named in the query; OR `grep -rE 'openai|langchain_openai|google.generativeai|genai|mistralai|cohere|ollama'` over the project hits (run this grep FIRST if no provider named — don't Read the file).Stars: 197kUpdated Claude CodeCodexUnverifiedInstall
doc-coauthoringanthropics/skillsGuide users through a structured workflow for co-authoring documentation. Use when user wants to write documentation, proposals, technical specs, decision docs, or similar structured content. This workflow helps users efficiently transfer context, refine content through iteration, and verify the doc works for readers. Trigger when user mentions writing docs, creating proposals, drafting specs, or similar documentation tasks.Stars: 197kUpdated Claude CodeCodexVerifiedInstall