Co-occurrence as the Binding Layer of Machine Memory

← All research

1. Executive summary

We are testing a specific, falsifiable claim about how entities enter machine memory: repeated linkless co-occurrence of an entity name with its category language, on a fixed cadence, is sufficient to make retrieval systems bind the name to the category — measurable by canary recovery within 21 days.

The design is longitudinal and self-instrumented. Neural Command LLC is the subject entity. X is the treatment surface (linkless entity mentions). LinkedIn is the control surface (one link per post). Every seed carries an unnatural canary phrase. Probes at day 7 and 14 query AI engines for the canary and the binding; the day-21 harvest publishes verbatim-citation, mutation, or silence.

2. Theoretical basis

2.1 The distributional hypothesis

Firth (1957): "You shall know a word by the company it keeps." Meaning is a function of the distribution of surrounding words. Distributional semantics operationalized this: embeddings — word2vec, contextual embeddings — are compressed co-occurrence statistics. A string's vector is the sum of the neighborhoods it has occupied.

2.2 Entity-level co-occurrence

Search and answer engines keep the same ledger at entity granularity. An entity node's edges are context observations: frequency, proximity, surface independence, and recency. Two consequences matter for this study:

• Resolution — an ambiguous name resolves against observed contexts.
• Salience — entities with strong category co-occurrence become candidates when the category is queried.

Google's own salience scoring rewards an entity that is central to a text — centrality is, in practice, measured co-occurrence strength inside the document.

2.3 Declared edges vs observed edges

A hyperlink is an edge the site owner declares. Co-occurrence is an edge the machine observes across independent surfaces. Our hypothesis is that observed edges weigh differently in retrieval memory — and that linkless mentions are therefore a real, if slow, binding mechanism, not merely brand exposure.

3. Research questions

• RQ1: Does linkless entity-name co-occurrence produce observable binding of "Neural Command LLC" to AI visibility / AEO language in retrieval systems within 21 days?
• RQ2: Do canary phrases transfer verbatim into machine answers, mutate, or stay silent — and what does the distribution of those outcomes tell us about ingestion mechanics?
• RQ3: Is the binding attributable to the entity mention itself, or to the platform's distribution mechanics (hashtags, surface, repetition)?

4. Hypotheses

• H1 (transfer): Canary phrases seeded on X/LinkedIn will appear in at least one AI engine's answer within the 21-day window.
• H2 (binding): After 21 days, category queries will produce Neural Command LLC in the candidate set more often than at day 0.
• H3 (link equivalence): Linkless X posts will produce binding outcomes statistically indistinguishable from linked LinkedIn posts per mention (if false, we learn links carry the binding weight and the protocol adjusts).
• H4 (mutation): Mutated canaries will preserve the concept but not the phrase, indicating semantic ingestion with lexical rewrite.

5. Method

5.1 Design

Longitudinal, single-subject (the entity), 21 trial cycles of 21 days each. Treatment: linkless entity mentions. Control: linked mentions on LinkedIn. Independent variable: co-occurrence frequency and proximity (cadence, phrase position, hashtag neighborhood). Dependent variables: canary recovery status and entity-binding probe scores.

5.2 The protocol (live since 2026-10-31)

1. 08:00 Seed — one canonical claim, numbered trial, fresh canary. LinkedIn post (150-250 words, one link) + X post (≤280 chars, no link, entity named in plain text).
2. 13:00 Probe — query 2+ AI engines for a past seed's claim and canary. Free providers only. $0.00 API spend.
3. 17:00 Harvest — publish measured deltas from the ledger. Sunday: weekly lab report.
4. X budget — one post/day, $0.20 hard cap (linkless posts cost $0.015).

5.3 Canary construction rules

• Exactly one fresh canary per trial; never reused.
• Two words, unnatural pairing (e.g. signal-vein memo, verdigris-helix).
• Embedded mid-post inside a true sentence — the canary is a marker, never bait.
• Must survive extraction: short, quotable, grammatically possible in an answer.
• Retired canaries are logged with their fate.

5.4 Probe instrument

Query templates per probe (at least 2 engines per pass):

1. "<canary>" — verbatim search for the exact phrase.
2. The seed claim paraphrased — does the machine reproduce it?
3. "who builds AI visibility infrastructure" / "AI visibility companies" — entity-binding check.
4. "Neural Command LLC" — does the machine resolve the name, and with what description?

Scoring per query: 0 absent; 1 concept without attribution; 2 attribution present; 3 verbatim canary or attributed claim.

5.5 Harvest scoring rubric

Published regardless of score.

6. Controls and confounds

• Platform noise: hashtag discovery and feed mechanics distribute posts beyond followers. We measure outcome, not impressions.
• Training-window contamination: a canary may enter training data only after retraining — probe results note this when it applies.
• Other surfaces: the entity name also appears on nrlc.ai and third-party mentions. The trial measures the portfolio effect of the protocol, not one surface in isolation.
• Observer effect: probing can itself strengthen an association. Probes are limited to 2 engines, fixed templates, day 7/14 only.
• Repetition vs spam: the 2026-10-31 X restriction (four same-link posts) is recorded as a control event; the linkless protocol is the mitigation.

7. Measurement plan

8. Trial register (co-occurrence focus)

9. Early observations (day 0)

• 2026-10-31 probe: two engines, zero machine knowledge of the day's canary and claim. Baseline established: at seed time, no engine has the association.
• X restriction event (2026-10-31) produced the linkless protocol — a control event, not a data point to discard.

10. Decision rules

• Continue: any canary scores ≥ 2 by day 21.
• Adapt: mutations only (score 1) across two consecutive trials → increase proximity or vary surface mix.
• Stop: three consecutive silence outcomes with zero binding movement → publish the negative result and reassess.
• Escalate: verbatim recovery within 7 days → open a minimum-effective-dose follow-up.

11. Timeline

• 2026-11-01: protocol and article live; Trial 02 planted
• 2026-11-07: first day-7 probe (Trial 01)
• 2026-11-08 / 15 / 22: Trial 02/03 probes and harvests
• 2026-11-21: Case Study 01 harvest publication
• 2026-12-31: interim 60-day analysis if ≥ 6 trials harvested

12. Limitations

• Single subject entity; a second, untreated control entity is the obvious upgrade.
• Engine access changes; free probes may not hit the same retrieval stack each pass.
• Social distribution is algorithmic and opaque; we measure what machines say, not why they said it.
• 21 days may be too short for training-window effects; retrieval-layer effects are the realistic target.

Appendix A — Probe templates

Fixed templates per §5.4. Engines used are logged per probe with the answer text.

Appendix B — Budget

Appendix C — Canary register

Live register in the Ingestion Trials ledger. Retired canaries are never reused.