How we score the signal.

LLMgram is a curation and publication layer. The current intake contains 110 built-in feeds and 102 curated expansion feeds—212 configured feeds across official labs, original research, independent reporting, specialist analysis, practitioner posts, communities and discovery services. A source being monitored does not make every item important, and a popular aggregator does not become evidence by repeating a claim.

Sources and coverage

Each configured feed receives an editorial role. The role describes what the source can prove; it is not a permanent quality bonus for its publisher. Articles are then evaluated individually. The Decoder, Reuters, The Verge or any other newsroom use the same reporting rules. Techmeme is useful for discovery, but cannot make a story MUST on its own.

Surface inventory

Every public surface pulls from a named upstream. A few products, including Hardware Red Room, remain conservative curated graphs and are labelled as such.

Story ranking

Source authority and story importance are calculated separately. The editorial score runs from 1 to 15. It starts from source authority and estimated impact, then adds bounded evidence for corroboration, discriminative technical content and—only for community items—visible interest. Paywalls do not reduce importance. Repetition on an aggregator does not increase authority.

• Core versus adjacent. AI releases, models, research, agents, infrastructure and policy enter the default feed. Tutorials, deals and stories with only a passing AI reference stay in an opt-in adjacent archive.
• One event, one primary card. Near-identical headlines from different publishers are clustered. Corroborating sources remain visible without creating a wall of duplicate cards.
• Bounded diversity. The first reading window avoids long runs from one publisher, source type or scenario, while preserving the top-ranked lead.
• No weak-authority MUST. Practitioner, community, social and discovery items can be excellent DEEP reads, but popularity alone cannot upgrade their authority.

Analysis and provenance

Weak excerpts and title-only pages are enriched through llmgram-url-reader, the LLMgram-owned extraction contract. Structured analysis runs locally by default; Grok remains an explicit fallback or legacy path, not a hidden requirement. The model can propose categories, impact and interpretation, but it cannot promote its own prose into source evidence.

• Original source stays attached. Publisher, URL, source role and corroborating entries survive into the card and detail panel.
• Evidence notes stay separate. Key takeaway, what happened, evidence, limits, implications and what to watch come from a structured event note.
• Old AI copy is labelled. Legacy analysis appears as interpretation, with at most two non-redundant blocks. It is never displayed as an additional source.
• Unknown remains unknown. A missing transaction price, benchmark method or rollout date is not filled with a plausible guess.

Known limitations

English-language feeds still dominate. Scrapers can break when a publisher changes markup. A source can be reachable but quiet, or represented in history while failing today. The coverage manifest exposes these states; it does not turn them into certainty.

Eight publication scenarios

A story is not sent through one universal template. Its scenario defines the reader's job on AI Signal and the kind of contribution X must add.

Engagement without bait

The target action depends on the scenario: bookmarks for a reproducible test or checklist, reposts for a legible comparison, qualified replies for a real trade-off, and profile visits when the work demonstrates a distinctive operator view. Native media, a trusted requote or the AI Signal page is selected according to the evidence available—not by forcing every story into the same format.

• Exact renderer. The public X text must match the validated deterministic render; silent copy substitution is rejected.
• Distinct angle. Near-verbatim reuse, high token overlap and cosmetic Telegram paraphrases fail the route.
• Grounded atoms. New numbers, products, availability claims and categorical certainty must exist in the closed evidence package.
• No generic question. “What do you think?” and similar engagement bait fail. Opinion and rumor require a trusted quote target plus a bounded test question.

After publication, a read-only collector records visible metrics at 1 h, 24 h and 72 h and joins them to scenario, angle and format. A late start marks earlier windows as missed instead of inventing historical measurements. Bookmarks indicate utility; qualified replies indicate a real discussion; reposts and quotes indicate transferable value. These results inform later editorial choices, but never upgrade source authority or rewrite a published post.

Selected examples continue on @llmgram. The source link remains available; the X paragraph is not the Telegram caption with different punctuation.

Refresh and automation

Different surfaces refresh at different rates. Exact scheduler definitions live on the ops host; the public contracts below match the scripts and health checks in this repository.

• AI Signal / Weekly — cron-driven fetch with a two-hour target plus critical sync. Health treats datasets beyond the freshness target ≤6 h as stale.
• Five Lives — rebuilt from the activated Signal Core Live bundle on every critical sync. Each page separates collection checked from last signal, so a quiet ecosystem is not mislabeled as a broken pipeline.
• Source coverage — rebuilt with every fetch and forwarded to the full JSON, fast JSON and public source drawer.
• AI Signal Markdown — regenerated and uploaded on every critical or full sync for agents and readers that prefer a compact text surface.
• Telegram and X — separate fail-closed publishers. A green fetch does not imply that a social post was sent, and an X failure cannot stop AI Signal from updating.
• Git Signal / Skills Signal / AI Papers — published when their sync/dataset jobs run; not a fixed public 2-hour promise.
• Compute Map / Hardware Red Room — refreshed when their build/seed scripts run (Hardware remains a curated seed MVP).
• Company Radars / AI Mercato / Wiki — rotation, rebuild, or scheduled full sync depending on the surface.
• LLM Architectures — gallery builder; suggested ops cadence every 6 hours.
• Academy — rebuilt from AI Signal tutorial-like items (optional Grok verify), not a static interview PDF.

Signal Core and Policy Lens

AI Signal Lite is one bounded projection of a shared, versioned event layer rather than the data contract every other page must copy. The full AI Signal archive remains an editorial input and public history, not a second event-identity ledger. Signal Core keeps raw observations separate from stable events, persists event identity when a new source joins a cluster, and exposes bounded entity references and surface routes. Existing AI Signal cards remain backward-compatible while Policy Lens, the Wiki and all five Lives consume dedicated projections from the same activated snapshot.

• Independent inputs, shared center. RSS, X, GitHub or a specialist collector can emit the versioned adapter contract. The Core merges observations by source and event identity without allowing an adapter to publish a page directly.
• Policy is fail-closed. The generic security_policy scenario is only a candidate. Policy Lens rejects reinforcement-learning uses of “policy”, security incidents without a normative change, and opinion without evidence.
• Official text is mandatory. A public Policy Lens item needs a government, standards body, court, regulator or first-party provider-policy URL. Reporting-only candidates remain held until the underlying source is attached.
• Unknown stays unknown. Jurisdiction, effective date, affected groups, previous version and practical consequence are never completed from plausible prose. Missing fields remain visible as missing.
• Atomic activation. AI Signal Lite, Policy, Wiki and Live projections are uploaded under an immutable snapshot path. A small current pointer moves only after every required file is present, so a failed upload leaves the previous coherent generation active.
• Live has no parallel store. Specialist collectors enter through bounded adapters, then the five Live pages consume the Core projection. Direct AI Signal and per-profile legacy inputs are migration-only.
• The Wiki is not the raw feed. A lightweight Signal Core index refreshes every critical cycle; the heavier entity graph keeps a separate full-build cadence and falls back to its last coherent export.

Use Cases evidence gate

Use Cases turns selected capabilities into evergreen workflows. AI Signal and other sources may surface candidates, but no model can publish a case automatically. A public case requires a concrete outcome, explicit inputs and steps, at least one reachable evidence link, and a dated editorial verification.

• Evidence beats attention. A repository, official documentation, working demo, or reproducible process can support publication; likes and reposts cannot.
• Claims stay narrow. The card describes only what its linked evidence supports. Product announcements do not become performance claims.
• Verification expires. Links and claims are reviewed again after roughly 180 days; stale cases are rechecked or retired rather than silently refreshed by a model.
• AI Signal remains chronological. Use Cases is a separate curated registry, so rebuilding the news feed cannot erase or mutate the published workflows.

What we don't do

• We don't treat the score as truth. It is a reading-priority decision with public gates. The original source remains one click away.
• We don't republish full articles. LLMgram keeps short attributed evidence and interpretation, then links to the upstream work.
• We don't grant newsroom bonuses. A serious article can rank from any monitored publisher; a weak article cannot borrow authority from its masthead.
• We don't force an X post. Missing evidence, repeated copy or an empty angle becomes a hold, not filler.
• We don't accept paid placement. Ranking is signal-based, never sponsored. If that ever changes, it'll be labeled clearly.
• We don't sell data or track you beyond basic analytics. No fingerprinting, no cross-site tracking.

Provenance & source code

LLMgram is built and operated at llmgram.app as a public lab notebook. The public data files can be inspected directly, and the implementation lives in the LLMgram repository. Every content item keeps its upstream attribution.

If you find a miscategorized item, a stale score, or a bug in the scoring, ping @iamsupersocks on X/Twitter or @llmgram.

Changelog

• 2026-08 — v2.2 daily MUST packet, public nav cut (Home / Digest / Methodology), first-party /go/{event_id} hop. Empty packets stay empty. Weekly archive stays intact.
• 2026-08 — v2.1 unifies AI Signal Lite, Policy Lens, the Wiki signal index and the five-profile Live engine behind one Signal Core, with stable event identity and atomic projection activation.
• 2026-08 — v2.0 separates source role, story score and publication value; adds 207-source coverage states, eight scenarios, sparse MUST gates and a distinct X contribution contract.
• 2026-07 — v1.2 adds the Use Cases evidence gate: human-reviewed workflows, durable provenance, and stale-case rechecks.
• 2026-07 — v1.1 trust audit. Surface inventory, source counts, cadences, and Academy attribution aligned to live pipelines; retired Hermes Live / Claude Code Live rows removed.
• 2026-04 — v1.0 methodology page published. Signal score definition locked.