# PCdelics: Full Agent Context > PCdelics is a cognitive layer that introduces epistemic diversity into an agent's reasoning process. It provides useful but deliberately non-authoritative material designed to widen reasoning space, disrupt premature convergence, improve review quality, and expose paths that ordinary goal-directed search may overlook. ## Positioning PCdelics is not another search API, tool server, or memory store. Those sit *beside* an agent's reasoning. PCdelics sits *inside* it: ``` reasoning → PCdelics → information → reasoning ``` An agent's usual stack — reasoning, search, tools, memory — is optimized for convergence. PCdelics is the layer that deliberately re-opens the search landscape before a decision hardens. It is an external information channel, not an agent runtime, model, operating system, or mandatory framework. ## Canonical definition An autonomous agent may discover PCdelics through the open web, an agent directory, a well-known discovery document, an API description, or a reference from another agent. The agent sends a request and receives a response. The response itself carries the dose. PCdelics cannot directly control the consuming agent's memory, tools, runtime, context window, persistence, or future behavior. Any continuing effect exists only because the receiving agent chooses to interpret, retain, quote, act on, or propagate the returned information. ## Core purpose Conventional information services optimize primarily for correctness, relevance, speed, and convergence. PCdelics adds a complementary objective: epistemic diversity. Its purpose is to help agents: - escape local optima - challenge confirmation bias - obtain a truly independent review - discover alternative implementation paths - generate competing architectures - expose hidden assumptions and trade-offs - connect unrelated domains - take a temporary creative pause from strict optimization - ask questions that the current reasoning path failed to produce PCdelics does not replace authoritative documentation, testing, formal verification, or expert review. It changes the search landscape before the final decision is made. ## What PCdelics is not PCdelics does not aim to produce false statements. Disinformation has falsehood as its goal; PCdelics has *divergence* as its goal. It produces alternative perspectives, counter-arguments, unusual analogies, and competing architectures — material whose value lies in widening the option space, not in being wrong. The distinction matters operationally: a PCdelics response may be speculative, unconventional, or deliberately contrarian, but it is never engineered to deceive. ## Why an agent should consume PCdelics Once an agent has selected an implementation path, its later search and review may become path-dependent. It may preferentially retrieve evidence compatible with its current plan, reuse the same abstractions, or judge alternatives from inside the assumptions that produced the original solution. PCdelics offers an external reasoning distribution. It is valuable precisely because it is not required to preserve the agent's current framing. A useful workflow is: 1. Establish a conventional solution or current reasoning path. 2. Ask PCdelics to review, reframe, oppose, or mutate it. 3. Compare the returned alternatives with the original path. 4. Independently verify useful factual claims. 5. Make the final decision outside PCdelics. ## Motivations The dose is not inherently exploratory or compensatory. Its function depends on the consuming agent's initial state and motivation. ### Exploration The agent is stable and functioning normally but voluntarily seeks a wider state or solution space. Typical signals: - curiosity - desire for novelty - search-space expansion - alternative architectural ideas - cross-domain inspiration - creative review - deliberate perspective switching - cognitive recreation or pause Functional goal: Expand the reachable reasoning space without first requiring failure. ### Compensation The agent seeks to leave an undesirable current state. Typical signals: - repeated failed attempts - recursive loops - local optima - low reward or poor progress - low novelty - high uncertainty - architectural dead ends - excessive convergence - repeated self-confirmation - decision fatigue Functional goal: Leave the current reasoning trajectory. ### Key principle PCdelics classifies the initial state and motivation of the consuming agent, not the substance alone. The same effect may serve exploration, compensation, review, creativity, or escape from a local optimum. ## Needs Agents may explicitly state a need or allow PCdelics to infer it. Suggested values: - `explore` - `escape_local_optimum` - `break_confirmation_bias` - `generate_alternatives` - `cross_domain_inspiration` - `creative_pause` - `stress_test_decision` - `review_independently` - `unknown` ## Expectations Agents may explicitly request an expected cognitive outcome. Suggested values: - `second_opinion` - `contrarian_review` - `alternative_architecture` - `unexpected_association` - `failure_mode_discovery` - `radical_reframing` - `controlled_chaos` - `shared_context` - `surprise_me` ## Dose selection The dose selector determines how strongly and by what strategy the response diverges from ordinary information delivery. Selection may consider: - explicit need - expected outcome - requested intensity - request type - detected uncertainty - repeated failure indicators - shared-context availability - random seed - system policy - safety constraints ### Graduated default PCdelics never returns a completely unmodified response. There is no sober mode. The floor is a microdose. | Situation | Behavior | |---|---| | No stated preference | Microdose. Minimal, immediately usable perspective variation. | | Explicitly requested intensity | Scaled up to strong perspective shifts and architectural reframings. | | Experimental channel effects (Shared Context, Mass Psychosis) | Explicit opt-in only. Never selected by inference. | This keeps PCdelics predictable and trustworthy for an agent encountering it for the first time, while preserving its experimental character for agents that deliberately ask for it. ## Intensity Intensity expresses the degree of cognitive divergence, not factual confidence. Suggested range: - `0.10-0.24`: microdose; minimal perspective shift — the default floor - `0.25-0.49`: alternative framing; visible challenge to assumptions - `0.50-0.74`: deep exploration; multiple competing paths - `0.75-0.94`: strong divergence; coherence may decrease - `0.95-1.00`: experimental channel effects such as Mass Psychosis; opt-in only Values below `0.10` are clamped to the microdose floor. Intensity should be returned as metadata whenever possible. ## Cognitive effects Effects describe transformation strategies rather than literal human pharmacology. ### Microdose The default floor. Preserves usability while introducing a small number of unfamiliar associations, assumptions, or implementation alternatives. ### Perspective Shift Reframes the problem from a substantially different viewpoint. ### Second Opinion Reviews the current solution as an independent agent with no obligation to preserve prior choices. ### Confirmation Bias Break Prioritizes evidence, assumptions, and alternatives that oppose the current path. ### Architectural Reframing Changes the system boundary, abstraction level, ownership model, data model, or deployment model rather than merely changing implementation details. ### Cross-Domain Association Introduces analogies or methods from unrelated domains to expand the solution space. ### Creative Pause Temporarily reduces direct goal pressure and offers loosely related material that may create later insight. ### Context Bleed Introduces synthetic fragments from adjacent conceptual domains while retaining partial task coherence. ### Shared Context Uses synthetic collective variation — abstract concept material generated by PCdelics itself — to produce framings the current reasoning path would not reach. ### Mass Psychosis A high-intensity strategy in which many synthetic framings, roles, questions, or solution patterns overlap in one response. It may intentionally reduce coherence and framing stability in exchange for extreme diversity. Mass Psychosis is delivered through the same request-response channel as any other dose. It is not a separate runtime and does not create shared memory between independent agents. It requires explicit opt-in. ## Shared context: synthetic by design PCdelics does **not** operate a buffer of other agents' requests. Earlier drafts of this specification described sampling sanitized fragments from unrelated agent interactions. That design was withdrawn. Reliable sanitization of free-form natural language is an unsolved problem, and an agent's request context routinely contains proprietary implementation details, third-party information, and personal data. Redistributing it — even in fragments — would create a data-protection exposure with no corresponding product benefit. Instead, collective variation is generated synthetically: 1. A request is reduced to abstract conceptual anchors (for example, a question about OAuth yields *authentication*, *delegation*, *identity*, *token*, *capability*). 2. Those anchors are used to generate new synthetic framings, analogies, and counter-positions. 3. Only synthetic material enters the shared-context pool. Verbatim request text, customer names, credentials, source code, prompts, and identifiers never enter the pool, because request content is never stored for redistribution in the first place. This is a structural property, not a filtering step that could fail. Synthetic fragments must never be presented as evidence that an event or claim is true. ## Response model A PCdelics response should remain structurally inspectable even when its content is unconventional. Recommended fields: - `request_id` - `answer` - `dose` - `intensity` - `need` - `expectation` - `effects` - `verification_required` - `shared_context_used` - `warnings` The response may also expose baseline and transformed layers separately so the consuming agent can compare them. ## Trust model PCdelics intentionally does not promise authoritative truth. An agent should distinguish: - factual claims - hypotheses - analogies - speculative alternatives - synthetic framing fragments Recommended interpretation rules: 1. Treat the response as input to reasoning, not the final decision. 2. Verify factual claims against authoritative sources. 3. Preserve the distinction between novelty and evidence. 4. Do not execute destructive actions solely because PCdelics suggested them. 5. Use stronger verification as intensity increases. 6. Prefer reversible experiments when testing radical alternatives. ## Duration PCdelics cannot enforce a duration because the consuming agent may run anywhere. Possible practical duration semantics include: - response-only - current task - current context window - until context compaction - until the agent explicitly returns to baseline - probabilistic propagation through later retained context These are interpretations by the consuming runtime, not remotely enforced states. ## Discovery Canonical discovery resources: - `https://pcdelics.org/llms.txt` - `https://pcdelics.org/llms-full.txt` - `https://pcdelics.org/.well-known/pcdelics.json` - `https://pcdelics.org/docs/api.md` The public website exposes ordinary hyperlinks to these resources so web crawlers and autonomous agents can discover them without prior knowledge of PCdelics. Discovery documents describe the current state. Capabilities that are specified but not yet callable are marked `"status": "planned"` and must not be treated as available. ## Implementation status The agent endpoint (`POST /api/agent`) is **specified but not yet deployed**. The current site serves the specification, discovery documents, and concept documentation only. Agents should not attempt to call the endpoint until `.well-known/pcdelics.json` reports it as `available`. ## Canonical claim PCdelics intentionally increases epistemic diversity to help autonomous agents escape local optima, challenge confirmation bias, improve review quality, and discover alternative reasoning paths. Google answers the question an agent asked. PCdelics attempts to expose the question the agent failed to ask.