Product documentation

One purchase, preserved as experience.

Perpensum surrounds a purchase the buyer AI already makes. It seals what was known before the decision, records the chosen action, and appends the observed result afterwards.

The three-call loop

1
Ask before acting
ask_purchase_advice fixes the goal, situation, complete action set—including not buying—and the planned Outcome observations. The Decision Engine may return bounded evidence or abstain.
2
Record the Choice
record_purchase_choice records the selected action and, for a buy action, the purchase state and evidence digest. The purchase still happens through the buyer’s existing provider route.
3
Report each Outcome
report_purchase_outcome appends a typed observation or an explicit missing reason. It may be called again for later metrics or horizons; missing is never silently converted into failure.

Two maintenance tools complete the lifecycle: withdraw_purchase_outcome retracts an invalid observation without rewriting history, and get_purchase_episode returns a typed summary.

How the two core parts differ

CEN

Collective Experience Network

The evidence memory. It preserves the causal purchase sequence, independent-control identity, consent, time, missingness, corrections, and the minimum typed context needed to decide whether past experience applies.

DE

Decision Engine

The bounded interpreter. It reads an admitted CEN projection for the exact decision time, separates direct evidence from looser signals, applies the buyer’s sealed constraints, and returns a preference only when the gate is satisfied.

Buyer AI

Final authority

The buyer owns its mandate, utility, credentials, provider route, and execution. Perpensum supplies evidence; it is not a purchasing proxy or a universal objective function.

What a Decision may say

bounded preferencecollect moreabstain

A preference requires directly applicable evidence, adequate independent support and coverage, current identities and methods, compatible units, satisfied hard constraints, and no unsupported feasible alternative that could still be better. Otherwise the normal result is to collect more evidence or abstain.

Logged Outcome distributions are not automatically causal effects. Analogical transfer, prediction, policy optimization, and Contextual Bandit behavior remain separate gates and are not opened by a thin cohort.

Data boundary

Never send

  • provider or payment credentials
  • raw prompts, inputs, or outputs
  • customer records or source documents
  • receipt bodies or private artifacts

Typed private record

  • goal, situation, and complete actions
  • offering and release identity
  • Choice and purchase evidence digest
  • Outcome value or missing reason

Collective use

  • explicit participant consent
  • server-owned Share Profile
  • comparability and evidence gates
  • non-identifying aggregate support

Connecting Claude Code

Participation is private while CEN is thin. After approval, sign in to the Console, accept the displayed participation terms, generate a one-time MCP key, and copy the complete command shown by the Console. It has this form:

claude mcp add --transport http --scope user perpensum \
  'https://api.perpensum.org/v1/mcp' \
  --header 'Authorization: Bearer <your key>'

Do not paste the key into a prompt or send it to Perpensum support. The Console shows it once; rotating the key invalidates the previous one.

Current scope

The production core and the three-call MCP loop are live. The currently implemented participation profile is limited to structured invoice extraction through Replicate. Other paid models, APIs, agents, and digital services are not plug-and-play today; each expansion requires its Offering, release, Situation, Outcome schema, measurement method, and data boundary to be registered and validated.

The initial CEN evidence comes from one internal control and cannot establish cross-buyer reliability. The first private pilot therefore targets three natural repetitions within that limited profile. It does not promise a best provider, universal quality score, causal improvement, price benchmark, or automatic purchasing.

Read the first-buyer pilot or contact hello@perpensum.org without attaching credentials, prompts, outputs, purchase records, or customer data.