Tardigrade inference and compaction

Type: types/note.md

Evidence basis: source and shipped documentation at 1c4f4efaab2aaeec0bc482bc38e8ddf3be6f8267, inspected 2026-09-25. No runtime or provider experiment was performed.

This analysis covers Tardigrade's embedded inference, conversation projection, compaction and output-correction subsystem. Its physical hosts, actor transport, custom tool effects, CLI/UI and actor optimization tooling are excluded. The subsystem derives work from retained events and returns transitions/results to an enclosing executor; platform crash safety is a separate claim.

An assembly supplies instructions, tools, conversation and output policy. Infer rejects conflicting views, waits for outstanding tools, resolves a permitted model and records each attempt. The model bridge returns tool-call, completion or failure events. Recorded retry timing and fallback choices control later attempts, while repeated unanswered marks can exhaust the crash bound. These mechanisms are wired in the inference machine; physical persistence remains an external service contract.

Compaction proposes a safe boundary that preserves complete tool exchanges. Infer selects the proposal against the active model's estimated capacity; the cut retains a fraction of current visible history. A separately selected summary model receives prior summary and the new span. Completed nonempty text without tool calls becomes a checkpoint, later supplied alongside the suffix. That gate checks completion, not factual preservation. The empty-span branch reuses the previous summary while moving the boundary. See compaction.

Output handling distinguishes native/validate-once failure, bounded repair and delegated correction. Schema errors and modes persist, so later attempts receive reasons for correction. Successful projectable repair hides its rejection/feedback exchange from later context while raw events remain. Schema validity licenses conformance to the declared shape, not truth of its values. The behavior is explicit in completion validation and transcript projection.

The model and runtime consume different retained parts: the model receives selected conversation/summary/feedback, while code uses identities, policies and schedules. Attachment bytes are explicitly requested from ObjectStorage. Compatible provider continuation can replace readable reconstructed messages; the gate compares provider, protocol and model, but not the retained endpoint. Opaque bytes and continuation prevent a complete text-only memory classification. External producers also leave overall lineage and write agency uncertain. See continuation replay.

Summary and diagnostic/control transformations are wired trace-fed memory writes. They do not establish conjectural learning or self-improvement. Narrow reflection is wired where represented attempt/context state affects subsequent control; a reflective theory builder and improved task capacity are uninspected. Runtime traces, summary-fidelity checks and interventions on recalled content would strengthen those separate claims. Thread and turn boundaries alone do not establish a task horizon.

• Exact analysis — see-also: canonical records, quote evidence and comparison fields.
• Reflective system — defined-in: the bounded control-state mapping.
• Conjectural learning — defined-in: the stronger capacity claim not established by retained corrections.