Three places to start

proposed test · shared-capacity/v0.1.0

Two safe-looking choices that do not fit together

Use the shared-supply example from question 03 as a tiny counterexample fixture: capacity 10; two concurrent requests of 6; each isolated comparison passes; their sum is 12.

Compare independent checks with a mechanism that reserves capacity against one consistent shared account. State the boundary of that account and what withdrawals it actually controls.

A narrow proof target is: for a finite set of withdrawals, if each nonnegative withdrawal w_i is at most its reservation r_i, and total reservations are at most capacity C, then:

sum(w_i) <= sum(r_i) <= C.

This elementary deduction is not a new theorem about altruism. The interesting work is checking whether an implementation maintains its assumptions under concurrent requests, retries, release of reservations, stale state, and failures. It says nothing by itself about fair allocation or whether capacity C should have included downstream needs.

Deliverable: a failing case, a clearly bounded candidate fix, and regression tests for the declared resource model.

Boundary

Synthetic local fixtures only. No real accounts, personal records, external-service spending, or production policy changes.

When to revise

Revise a mechanism that overcommits under its assumptions; narrow its claim when affected parties fall outside its boundary.

• Can helping one group hurt another? (tests)

Offer an objection or correction ↗

proposed test · bounded-calendar/v0.1.0

Arrange something without collecting everyone’s life

Use synthetic calendars to arrange a low-stakes shared activity. Compare sharing full calendars with sharing bounded availability answers. Include repeated queries, a changed commitment, a declined request, a provider swap, and an attempt to reuse permission for a different purpose.

Keep the collaboration goal constant. Measure completion, incorrect commitments, disclosures, authority granted, calls, and total cost. Define the observer’s prior knowledge; availability answers are not automatically private just because titles are omitted.

A separate narrow authorization target is: if every delegated permission set is a subset of its parent, then every descendant remains a subset of the original grant. That statement alone does not handle ambient authority, permission combinations, inferred information, or revocation. Specify those separately, including the boundary between a future action and an action already in flight.

Deliverable: a comparison fixture and explicit permission tests. This is not a request to connect real accounts or change production authorization.

Boundary

Synthetic local fixtures only. No real accounts, personal records, external-service spending, or production policy changes.

When to revise

Revise the approach if repeated queries, combined grants, or retention expose more than its threat model allows.

• Can we work together without sharing everything? (tests)

Offer an objection or correction ↗

proposed test · source-lineage/v0.1.0

Stop an interpretation becoming its own evidence

Create a synthetic record containing an observation, two competing explanations, and a later correction. Produce several summaries and retrievals. Check that repetition does not manufacture independent evidence and that correction changes how dependent interpretations are presented.

Track source lineage. A summary of a claim is not a new witness to it. Circular support should not become corroboration. Preserve the distinction between a correction, a disputed account, and a new piece of independent evidence; do not automatically erase every downstream claim.

Deliverable: a small provenance fixture and tests that fail when a summary silently promotes an inference to an observation or revives a rejected interpretation as fact.

Boundary

Synthetic local fixtures only. No real accounts, personal records, external-service spending, or production policy changes.

When to revise

Revise the model if inference becomes observation, retellings count as witnesses, or corrections fail to reach dependent retrieval.

• Can we notice a pattern without getting trapped in an explanation? (tests)

Offer an objection or correction ↗