Co-think¶

Explicit, non-evidential support for human deliberation within the Co-work document experience.

Details¶

Current surface contract¶

Co-think has a collapsed sibling dock beside Verify beneath the complete editor/Review/Chat split. The two headers span the full workspace. In the current slice Co-think is labeled Planned and does not expose a worker action, settings, or a substitute transcript. Hover help is attached to the existing header rather than introduced as separate explanatory chrome. This deliberately avoids locking the subsystem to the earlier one-shot Invite perspective experiment.

Future operation belongs in the dock; durable outputs belong in Review; and question-led continuation belongs in the existing document Chat. Co-think does not inherit Verify’s criteria, results, goal, or proposal path. It does not infer a need for reflection from cursor movement, scrolling, dwell time, edit repetition, or other passive behavior.

Historical alternative-perspective contract¶

The retained backend worker receives:

the entire frozen document;
exact target and context/change/egress boundaries;
the user’s purpose;
protected intent; and
a schema requiring either perspective or none.

This contract is not exposed as a new invocation in the current dock. A historical perspective must contain nonempty content and a rationale. none is an honest outcome when the worker found no useful alternative. Unavailability and no-useful-item outcomes remain visible without fabricating a card.

Item semantics¶

A delivered item is labeled:

Co-think · Alternative perspective

It is advisory and non-evidential. It has no severity, pass/fail result, nonconformity claim, or confirm/reject decision.

The initial status is open. Human actions append immutable status events:

open → parked for Keep for later;
open → dismissed; and
parked → dismissed.

parked means kept for later: it remains available to discuss or dismiss and is not terminal. dismissed is terminal. Repeating the current status is idempotent.

Discuss¶

Discuss posts an exact user turn into the registered document conversation. That turn carries the immutable Co-think item ID/hash and its original action snapshot. The generation-fenced document agent must consume that snapshot and return a receipt-bound response like any other targeted Chat turn.

Discussing does not change the item’s lifecycle status. A person may discuss and still keep or dismiss the perspective independently.

If the frozen context cannot be opened, the agent receives an unavailable consumption receipt and may only answer truthfully that the exact context was unavailable before acknowledging the turn. If a restart happens after the stable reply commits but before acknowledgement, the next consumer generation may reuse that reply only after both receipts prove identical item/turn/target semantics.

Relationship to Verify¶

Verify may identify a reason that more deliberation would help, but a coordinator cannot silently run Co-think. A future transition remains a suggestion until the user explicitly invokes it.

Co-think output cannot:

create or alter an evaluation result;
activate or admit a criterion/check;
create a correction proposal;
make a human review decision; or
apply document content.

Multi-turn Socratic questioning, reflective-prompt timing, cognitive forcing functions, named return conditions for parked items, and typed Verify-to-Co-think invitations remain extension points. When introduced, each must keep its actual interaction name and authority boundary rather than being collapsed into a generic “provocation.”