Skip to the content.

Refresh and rumor protocol

The corpus is a dated snapshot, not a timeless market map. Refresh it without turning search snippets or copied leaks into facts.

Cadence

Search ledger

For each refresh, record the exact date and search families used. Minimum families:

  1. entity + infrastructure, cluster, GPU, TPU, Trainium, datacenter, power;
  2. entity + inference, batching, KV cache, prefill decode, routing, SLO;
  3. LLM production trace, workload characterization, token distribution, prefix reuse, tenant skew, Zipf, burst, seasonality, agent trace;
  4. AI datacenter + grid, water, cooling, permitting, opposition, delay, canceled, turbine, transformer, optics, HBM, CoWoS;
  5. company/project + funding, debt, bankruptcy, shutdown, acquisition, license, partnership, release, generally available;
  6. company/project + rumor, reportedly, sources say, leak, followed by a search for the earliest origin and independent corroboration.

Search snippets are discovery aids only. Open the source and record the exact publication and event dates before adding an entry.

Rumor lifecycle

A rumor entry requires:

Rumor entries never support an unqualified architecture or roadmap fact. They can justify watching a seam, preparing a reversible experiment, or assigning probability in a scenario.

Contradiction handling

Do not overwrite the older entry when a later source disagrees. Add the new source and record:

Completion bar for a slice

A slice is complete only when it has: