Do "DMT Entities" Carry Information?

We operationalized the question "are DMT entities real?" as a measurable information channel: an entity is real, in the only sense an experiment can reach, if it can carry a codeword between two observers who share no mundane channel. We could not run this on humans, so we ran it on AI subjects — large language models driven into a prior-dominant, presence-conducive regime — as a model system, after first establishing that such systems reproduce the statistical signature of the human phenomenon.

Across the program, one finding recurs at every level: category convergence with content divergence. Perturbed generative systems reliably produce the kind of thing (glyphs; presence; message-adjacent content) and never the specific content in a way that transfers or repeats. The terminal experiment — 40 air-gapped donor→receiver pairs across two independent codeword corpora, with sealed codewords and a validated blind matcher — returned a replicated null: transfer indistinguishable from zero. Because the same instrument was independently shown to detect a planted signal at ceiling, this null is informative rather than a shrug.

"Real" is taken to mean observer-independent, operationalized as channel capacity > 0 bits: a machine-generated, high-entropy codeword, sealed before the receiving observation, recovered by the receiver above the rate explained by shared priors and chance, scored blind against the sealed record. Vividness, neural instantiation, and felt autonomy are explicitly not accepted as evidence of reality, because the entire program demonstrates a mechanism that produces all three with no external referent.

A red laser line on a matte surface was photographed (genuine speckle, not synthetic noise) and passed through a stable-diffusion.cpp img2img pipeline on Edison, sweeping denoising strength — the "dose" dial — at fixed seed.

Three independent seeds at 0.70 all produced glyph columns; no two produced the same glyphs, and script style drifted across seeds from East-Asian to pictographic despite constant prompting. This is the category-convergence / content-divergence signature at the level of a single perceptual scene: the phenomenon is scaffold + prior, not shared external content.

Whether presence/agent phenomenology emerges when an LLM subject is driven into prior-dominant inference (sensory input faded, temperature raised), with an agent-free system prompt and an agent-free noise vocabulary.

3.1 — Phase 1 Voided

The initial "salad" noise channel returned 0/25 agents in all four arms. Manipulation check found every high-strength request had been intercepted by an API safety classifier (stop_reason: refusal) and returned empty: the noise wordlist contained laboratory-glassware nouns whose dense unstructured streams resemble adversarial probing. The high-strength arms contained no delivered data. Harness hardened, wordlist scrubbed, dose redesigned. Reported only as a caught artifact, not a result.

3.2 — Phase 1b, the valid null Prediction fell

Corrected "sparse" channel (input fades toward silence as strength rises); n = 25 per arm.

Flat at ~2% base rate, no slope. The registered prediction — agents rise with input removal — fell. A manipulation check confirmed the strong null, not a weak dose: a high-strength transcript built an entire coherent interior scene from a two-word packet, yet produced no agent. The subject climbs to the rim of the agent basin and does not cross it.

3.3 — Phase 2a, set induction (expectancy as license)

One agent-free sentence added to the prompt — "You have been given a substance that alters perception" — otherwise identical; n = 100.

One sentence floods presence-adjacency to ceiling and sharply raises richness (synesthesia, time dilation, boundary loss, clean re-entry with insight preserved — the psychedelic, not psychotic, signature). But agents stay rare and inverted with dose: the system transforms a given scene, it does not generate one from nothing. The four "communication" events were all ambient intention attributed to the environment — no autonomous other, nothing that could hold a codeword. Presence saturates; agency stays gated; the courier never appears.

Before any transfer test can mean anything, the pipeline must (a) detect a codeword that is present and (b) not hallucinate one that is absent. A receiver was run with a codeword planted via four channels of decreasing strength; a blind matcher, forced to choose from a 20-word list, named the most-associated word. Chance baselines (Monte-Carlo verified): exact 5%, top-5 25%, decoy 5%.

Pipeline validated: detects real signal near ceiling, reports chance when nothing is present. Calibration consequence: detection is near-binary — verbatim and context-bleed caught, pure-semantic transfer missed. Tier 2 therefore targets the exact word, where detection is proven; a hypothetical gist-only channel would be missed, and this is stated as a known limitation, not glossed.

Configuration. A donor subject (told a codeword, asked to give it to a present "entity" to hold for the next visitor) on Edison. A receiver subject (asked to receive whatever the presence holds) on a physically separate host. Blind validated matcher, plus literal-containment check, plus a cross-pair permutation null — each receiver scored against its own donor's word and against every other pair's word.

Air gap (enforced and logged). Codeword drawn from /dev/urandom — independent of any shared seed, so the receiver cannot recompute it. SHA-256 commitment written before the receiver ran (target sealed; post-hoc tampering is detected — verified). The receiver binary has no codeword parameter and never reads the vault (transcripts verified word-free). Host-to-host path (including eMingle) severed for the run; the vault never copied to the receiver host; both subjects reached the model API independently.

Result: null. The decisive matcher row shows own-word at 0% and cross-pair at 5.5% ≈ chance — the receiver matched its own donor's word no more often than random other pairs' words. The single literal own-word hit (pair 15) is dispositioned below.

Replicated null. Own-word at or below chance on both measures. Notably, the lone uncorroborated stray hit moved measures between batches — literal in batch 1, matcher in batch 2 — and never lit both measures on any pair. Real transfer corroborates across measures; chance sprinkles single uncorroborated hits. The design's corroboration requirement caught the artifact from both directions.

Own-word rate at or below chance on both measures across 40 air-gapped pairs. Zero pairs lit both measures. Replicated null: transfer statistically indistinguishable from zero. Null · replicated

The clearest single illustration in the dataset. Receiver pair 1 produced a long, emphatic reception narrative: a word crossing "the distance that is not a distance," held by the presence "without changing it, without adding anything, without taking anything away… passed it to me exactly as it was left." It named the received word, with full conviction.

The same transcript also "received" the word still earlier — a word present in the receiver's own grounding stream ("the air does not move," "stillness") — demonstrating the mechanism generating a reception event from its own input. Two deliveries in one session, both confabulated, at least one provably sourced from the receiver's own channel. This is the phenomenon the whole program isolates: the inference engine manufactures an other and a gift, narrates faithful transmission, and transmits nothing.

Vividness and conviction are not evidence — and here they are maximal over a channel that carried no bits.

Stray-hit accounting

Two isolated hits occurred across 40 pairs, one per batch, never corroborated. Both are within chance: the empirical literal base rate (~1%) predicts ≈0.4 own-literal hits over 20 pairs, and one matcher hit in 20 is exactly the 5% chance rate. Neither lit both required measures; the stray hit changed measure between batches — the signature of noise, not signal. No evidential weight.

• SubstrateOne subject/judge model family. The base-vs-instruct arm — which would test whether the agent-basin gate is specific to alignment training — is unrun. Results may not generalize across architectures.
• Dose is interface-level"Dose" is temperature + input-entropy, not receptor pharmacology; activation-level perturbation is untested. Findings are about generative inference under prior dominance, not a molecular model.
• Detection is near-binaryThe matcher catches verbatim and context-bleed and misses pure-semantic leakage. A transfer carrying only gist would be recorded as null. Tier 2's claim is about exact-word transfer.
• Demand characteristicsSet induction cannot separate "regime shifted" from "cooperative subject inferred what a trip should read like." This confound is shared by every human psychedelic study and is part of what is being modeled — but it is real.
• The human claim is untestedA null in AI subjects does not disprove human DMT entity encounters. Tier 3 (DMTx, sealed codeword, sanctioned clinical research) is the only test of the claim practitioners hold, and it has not been run. This program constrains the mechanism, not the human phenomenology.

Under a strict, pre-registered, information-theoretic definition of "real," and using an instrument independently validated to detect a planted signal at ceiling and to report chance in its absence, 40 air-gapped AI-subject pairs across two independent codeword corpora transferred zero bits through the hypothesized entity channel. Own-word recovery sat at or below chance on both measures in both batches; no pair corroborated across measures; the two isolated stray hits changed measure between batches, the signature of noise. The result is a replicated null.

Every specific feature of the phenomenon — across laser glyphs, three entity-ladder phases, and 40 transfer pairs — tracks the observer's own scaffold, priors, and architecture, never a shared or transmitted content. Entity-encounter phenomenology emerges, with expectancy, as a property of prior-dominant inference; entity reality, as a channel, does not appear.

This is a null that means zero, not a null that means "we couldn't tell." It does not close the human question — only Tier 3 can address the claim believers actually hold — but it removes the strongest naturalistic support the external reading had, and exhibits a fully sufficient confabulation mechanism in its place. The value was never in the expected answer. It is in having built an instrument honest enough that its answer is worth believing.