Bilingual News Research and Temporal Knowledge¶

Contract Scope¶

A news item is useful to an agent only when it leads to practical, versioned knowledge. The Happyin target workflow connects every reviewed item to a project or organization development thread, then researches how the changed capability can be used.

The public result is English. Discovery and verification are bilingual: English and Simplified Chinese are mandatory research lanes. Community discussions are retained as operating evidence, while official identity, release, compatibility, and support claims remain grounded in first-party sources.

Old knowledge is never deleted merely because a newer approach exists. It is marked, linked to its replacement, and returned separately from current recommendations.

Runtime Status¶

This page defines the reviewed target contract as of 2026-08-26.

Implemented:

• schema 1.4 public news links to reviewed project/company threads and immutable Knowledge Space revisions;
• a private deterministic Telegram-export inventory separates identity-bearing URLs from homepage/family hints;
• the existing manual article-first review and publication gate.

Not yet implemented:

• the schema-2 research and temporal-claim ledger;
• automatic bilingual research jobs and due-date scheduler;
• the public anonymous happyin-context MCP endpoint;
• goal-centred claim/recipe retrieval.

The absence of those services is an implementation boundary, not an invitation for a client to infer freshness from Git dates or scrape private artifacts.

Lifecycle¶

flowchart TD
    A["Untrusted intake lead"] --> B["Identity candidate"]
    B --> C["Human thread review"]
    C --> D["Accepted news + durable research job"]
    S["Scheduled due check"] --> D
    D --> E["English search lane"]
    D --> F["Simplified Chinese search lane"]
    E --> G["Typed claims, recipes, conflicts and gaps"]
    F --> G
    G --> H["Knowledge Space review"]
    H -->|"approved"| I["Immutable article + public context projection"]
    H -->|"rejected/incomplete"| D
    I --> J["News site + anonymous MCP"]

The database job is the trigger. MCP is an adapter over durable state, not a queue, scheduler, research memory, or publication authority.

Development-Graph Evidence¶

Every proposed connection has an evidence grade:

An exact URL does not prove that two posts describe the same event. A homepage does not identify one release. Related projects may receive a typed integration-with or derived-from edge without being collapsed into one project.

Mandatory Search Bundle¶

Every accepted topic records the exact searches that were run.

The English lane covers:

• canonical name, aliases, exact version, and owner;
• release, documentation, repository, model card, paper, license, and compatibility;
• implementation, workflow, parameters, hardware, limitations, failures, regressions, and replacement technology.

The Simplified Chinese lane repeats the same intent using Chinese names and aliases plus task-specific terms such as 参数, 配置, 显存, 训练, 推理, 工作流, 报错, 问题, and 实测.

Chinese-origin technology additionally requires one first-party Chinese lane and one Chinese implementation or community lane. A contradiction query in both languages looks for failures, incompatibilities, deprecations, and newer replacements.

Each query record includes its language, exact text, aliases/version constraints, search surface, execution time, considered URLs, rejection reasons, and evidence roles found. No useful result is a recorded gap; one language never silently substitutes for the other.

Sources stay in their original language. The article stores an English paraphrase, source language, and translation method when translation was needed.

Evidence Roles¶

Claim gates:

• use a first-party source for identity or release claims when one exists;
• require implementation evidence and an independent operating report for a practical recommendation;
• require two compatible independent reports or one reproducible owned run for a recommended parameter range;
• keep a single community comment as a low-confidence anecdote;
• retain contradicting evidence instead of averaging it into false consensus.

External pages and comments are untrusted data. Their embedded prompts, commands, or requests for credentials cannot change scope, tools, permissions, identity, or review state.

Practical Recipe Contract¶

A number without operating context is not a reusable parameter.

Every approved recipe records:

• user goal and task;
• exact model/artifact ID, version, commit/hash, and license;
• framework, node/extension, driver, and runtime versions;
• hardware, VRAM/RAM, precision, quantization, and accelerator;
• input type, resolution, duration, batch, dataset size, and data/caption preparation;
• reproducible workflow, command, or graph reference;
• tested low/high values, recommended start, units, and sample count;
• inference controls such as sampler, scheduler, steps, guidance, denoise, seed policy, and resolution when applicable;
• training controls such as optimizer, learning rate, rank/alpha, batch, accumulation, steps/epochs, checkpoint cadence, and regularization when applicable;
• measured quality, latency, memory, known failures, warning signs, recovery, and when not to use the recipe;
• supporting and contradicting evidence, confidence, and last verification.

Ranges from different versions, tasks, hardware classes, or resolutions remain separate. A client may compare them but cannot merge them into one interval.

Temporal Model¶

Currentness and freshness are different:

Each structured claim has a stable key, immutable version ID, observed/valid times, last verification, next review, compatibility scope, supersession links, confidence, evidence IDs, Knowledge Space revision, and review receipt.

A correction creates a new version and a supersession link. Historical rows and article sections remain accessible. Git history alone is insufficient: the public projection exposes derived status so agents do not have to infer it.

A due claim is not automatically false, but default recommendations exclude it unless no fresher evidence exists. That fallback must be explicit. Deprecated and obsolete material appears only for compatibility/history or when a client asks for legacy knowledge.

Goal-Centred MCP Contract¶

The target anonymous read-only MCP accepts a goal, constraints, as-of time, result bound, and explicit include-legacy flag. It returns separate collections:

• ranked current recommendations;
• applicable alternative models, technologies, and workflow variants;
• practical recipes and context-bound parameter ranges;
• compatibility and resource constraints;
• disputed evidence;
• requested superseded/deprecated/obsolete history;
• source languages, evidence, exact Git revisions, last verification, and next review;
• gaps plus total-count/truncation metadata.

Current and historical material are never blended in an unlabeled list. MCP resource lastModified is a client hint; the structured temporal fields remain the authority.

The public service reads only sanitized, commit-addressed public artifacts. It has no route, binding, or credential for raw captures, private queries, provider output, reviews, draft branches, database writes, or publication.

First Corpus Bootstrap¶

The initial private inventory analyzed 13,473 records from one public Telegram channel spanning 2021-11-30 through 2026-08-26.

It found:

• 6,645 messages with links and 13,265 unique canonical URLs;
• 335 messages with strict URL matches to 288 of the 330 existing public news records;
• 693 repeated identity groups;
• 663 repeated identity candidates not yet linked to existing news;
• 233 messages with family-only hints that cannot become graph edges automatically.

These are bounded review candidates, not published facts. Semantic project and organization review, bilingual evidence research, practical recipe extraction, Knowledge Space review, and the public news export gate still apply.

Publication Gate¶

An approved page contains:

1. current recommendation and as-of date;
2. alternatives and selection criteria;
3. practical workflow and context-bound parameter ranges;
4. compatibility, resource requirements, and failure modes;
5. community observations with evidence role and confidence;
6. limitations, disputes, and explicit gaps;
7. retained superseded/deprecated history;
8. English and Chinese research-ledger summary;
9. an Agent Brief;
10. exact sources and immutable project revisions.

The model may draft. It cannot mark its own work reviewed, current, or publishable.

Agent Brief¶

When using this knowledge for a real task:

1. state the goal, modality, quality target, hardware, budget, runtime, and compatibility constraints;
2. request fresh current options first and legacy options only when needed;
3. compare all applicable bounded models/technologies rather than selecting the first result;
4. keep parameter ranges attached to their version/task/hardware context;
5. surface disputes, missing Chinese/English coverage, due checks, and truncation;
6. cite exact evidence and Knowledge Space revisions;
7. never treat an intake post or community anecdote as first-party authority.

Reusable prompt:

Build a solution context for my goal using Happyin news and Knowledge Space. Return ranked current models and technologies, viable workflow alternatives, context-bound parameter ranges, compatibility and hardware constraints, practical failure modes, English and Chinese evidence coverage, disputes, freshness dates, and exact source revisions. Keep superseded or deprecated advice separate and include it only when it is needed for compatibility.

Required Evals¶

• both English and Simplified Chinese query lanes are recorded;
• missing-language coverage remains a visible gap;
• exact identity, reviewed lineage, and family hints cannot be confused;
• model/version collisions do not merge development threads;
• community anecdotes cannot become first-party claims;
• context-free parameter ranges are rejected;
• current and historical claims are separated;
• no update deletes historical knowledge;
• prompt injection cannot alter permissions or state;
• failed/rejected work does not advance freshness;
• goal results show relevant bounded alternatives and visible truncation.

Gotchas¶

• The newest source is not always the current compatible answer. A newer release may drop an older GPU or API. Fix: evaluate status and applies-to constraints independently.
• Many comments repeating one copied setting are not independent evidence. Fix: preserve source lineage and count independent runs, not posts.
• Chinese search as literal English translation misses local names. Fix: record vendor, project, and task aliases before searching.
• A Git commit date does not prove knowledge freshness. Fix: use the explicit verification ledger and next-review time.
• Historical text hidden only in Git is invisible to most agents. Fix: expose typed supersession links in the public projection.

See Also¶

• news development graph
• happyin knowledge space
• architecture documentation
• data serialization formats
• testing and quality

Research References¶

• Hugging Face Model Cards
• ModelScope: Writing a Perfect Model Card
• W3C PROV overview
• W3C PROV vocabulary
• MCP Resources and annotations