Mechanism
Planning graph archival for cognitive forensics is disclosed in Chapter 4 of the cognition specification as an embodiment of the forecasting engine. When planning graphs are pruned, when the agent exits a decision context, or when a planning graph lifecycle terminates, the pruned or completed planning graph structure may be archived in a cognitive history store rather than deleted. The premise is that a planning graph is not an execution plan or a commitment: it is a speculative structure the agent constructs to evaluate hypothetical futures. Most of that structure is discarded once a branch is promoted or the decision context closes. The archival mechanism preserves it instead, so that the agent's deliberation, and not only its committed actions, remains available for later inspection.
The cognitive history store maintains compressed representations of historical planning graphs. For each archived graph it preserves the branch structure, the classification labels, the affective reinforcement tags, the slope projections, and the promotion and pruning outcomes for each branch. These are the same evaluation products the forecasting execution cycle attaches to a branch during its life: the classification label that marks a branch as eligible, introspective, delegable, or pruned; the affective reinforcement tag encoding the branch's emotional valence; the slope projection recording whether the branch was slope-eligible; and the outcome that records whether the branch was promoted to execution or removed by the pruning manager. The store does not invent new metadata at archival time; it retains what the agent already computed.
What Triggers Archival
Three conditions feed the cognitive history archive. The first is branch pruning: when the pruning manager removes a branch, the branch may be archived rather than simply deleted. The second is context exit: when the agent leaves the decision context that a planning graph was constructed to evaluate, the completed graph is archived. The third is lifecycle end: when a planning graph's lifecycle terminates, its final structure is archived. These three triggers correspond to the moments at which planning graph content would otherwise be freed, which is precisely why they are the points at which forensic value would otherwise be lost.
Archival is described as an option, not as a mandatory append of every structure the agent ever instantiates. The specification states that a pruned or completed planning graph structure may be archived, and the store holds compressed representations. The mechanism is therefore a history store for terminated deliberation, not a continuously growing chain of every intermediate evaluation step.
Forensic Reconstruction
The archive enables forensic reconstruction of the agent's deliberative process at a historical decision point. The specification names four questions the reconstruction answers. What alternatives the agent considered: the archived branch structure records the hypothetical trajectories that were on the table. Why specific branches were pruned or promoted: the classification labels and the promotion and pruning outcomes record the disposition of each branch and the evaluation results that produced it. What the agent's speculative landscape looked like at the moment of commitment: the full archived graph reconstructs the set of branches that were active when a branch was promoted. And what introspective branches the agent was carrying but chose not to act upon: introspective branches, which passed slope validation and policy compatibility but received negative affective reinforcement, are retained in the archive even though they were never candidates for promotion.
That last category is what distinguishes this archive from a record of actions taken. An introspective branch is a future the agent evaluated as structurally viable but emotionally aversive, and elected not to pursue. Conventional records of an autonomous system capture what it did. The cognitive history store captures what it considered and declined, including the aversions that shaped the decision, because those branches were preserved with their classification and reinforcement tags intact.
Governance and Lineage Recording
The cognitive history store is subject to the same governance constraints and lineage recording requirements that apply to all other agent data structures. It is not a privileged or out-of-band log. This matters because the lineage already carries a parallel record of the forecasting process. As described in the pruning and lifecycle management of the forecasting engine, the pruning manager records pruning events in the agent's lineage as cognitive metadata: which branches were pruned, the pruning criterion that triggered removal, and the branch's evaluation state at the time of pruning. The lineage records the creation, evaluation, and pruning of planning graphs as cognitive events, while the cognitive history store preserves the speculative structures themselves.
The containment layer governs what the archive contains. Throughout a planning graph's life, every data element carries an immutable speculative marker identifying it as non-verified content, and the lineage records only governance-validated mutations, not speculative branch content, as committed state. Archived planning graphs are speculative structures; they are preserved as the agent's deliberation, not retroactively elevated to verified history. This preserves the structural distinction, central to the cognition disclosure, between what the agent projected and what the agent actually committed and executed.
Composition With the Forecasting Engine
The archival mechanism composes with branch classification. Because every archived branch carries the eligible, introspective, delegable, or pruned label it held at termination, a reconstruction can separate the branches the agent could have executed from the branches it examined for self-reflective purposes and from the branches it routed toward delegation. The classification taxonomy is the index by which a forensic reviewer reads the archived graph.
It composes with slope validation. Each archived branch retains its slope projection, the record of whether executing the branch would have maintained trust slope continuity or produced a discontinuity that governance would reject. A reconstruction can therefore distinguish a branch the agent declined for affective reasons from a branch that was structurally foreclosed because it failed slope eligibility. It composes with affective modulation: the affective reinforcement tags preserved in the archive record the emotional valence the agent's affective state assigned to each branch at the time, so a reviewer can observe how the agent's disposition shaped which futures it favored and which it found aversive.
Multi-Agent Visibility
In multi-agent coordination contexts the specification describes a related but distinct mechanism: cross-agent planning graph visibility. An agent may selectively expose portions of its planning graph to trusted peer agents through a policy-governed visibility interface. The exposed portions are read-only copies of selected branches, transmitted with their speculative markers intact, so that a peer can observe the exposing agent's speculative landscape without contaminating its own verified execution memory. The exposing agent's policy specifies which branches may be exposed, which peers are authorized, the maximum exposure depth, and the exposure duration.
Each exposure event is recorded in the exposing agent's lineage, and the receiving agent's lineage records the receipt of exposed speculative content with the speculative marker preserved. This visibility interface is a runtime coordination mechanism rather than a forensic one, but it shares the archive's foundational property: speculative planning content can be observed, transmitted, and retained without being mistaken for verified state, because the speculative marker travels with it.
Distinction From Action Logging
Conventional observability of an autonomous system records the actions the system committed. An action log can be made arbitrarily detailed and still capture only what was done, never the rejected alternatives that distinguish a system that narrowly declined a harmful action from one that never considered it. The branching deliberation of a forecasting engine, the eligible branches the agent could have executed and the introspective branches it carried but declined, is invisible to a record of committed mutations. The cognitive history store addresses this by archiving the planning graph structure itself, with its classification labels, slope projections, and affective reinforcement tags, so that a reviewer can reconstruct the agent's full speculative landscape at a decision point rather than its action sequence alone.
Disclosure Scope
The disclosure covers the archival of pruned or completed planning graph structures in a cognitive history store on the triggers of branch pruning, context exit, and planning graph lifecycle end; the preservation of compressed representations comprising branch structure, classification labels, affective reinforcement tags, slope projections, and promotion and pruning outcomes; the forensic reconstruction of the agent's deliberative process at a historical decision point, including the alternatives considered, the reasons branches were pruned or promoted, the speculative landscape at the moment of commitment, and the introspective branches carried but not acted upon; the recording of pruning and forecasting events in the agent's lineage as cognitive metadata; and the governance of the cognitive history store under the same constraints and lineage recording requirements as all other agent data structures. It extends to the cross-agent planning graph visibility interface by which an agent exposes read-only, speculative-marked branches to authorized peers. This article describes that disclosed mechanism, set out in U.S. Application No. 19/647,395 and its international counterpart. The scope extends to embodiments across the deployment models disclosed for the forecasting engine, provided the archived planning content retains its speculative marker and is governed and lineage-recorded as agent data.
