Recall
What is recall?
Recall is the mechanism by which knowledge that was once set aside — archived because it contradicted the current consensus, because it was below the plausibility threshold, or because the evidence was insufficient — is retrieved and re-evaluated when new evidence changes what was previously improbable or implausible.
This is one of the most distinctive properties of AIEP. Most systems discard what does not fit. AIEP preserves everything with provenance intact — then provides a structured, deterministic pathway for previously rejected knowledge to become operative again.
Four patents govern this: P22 (deterministic recall context reconstruction), P83 (dissent fork generation), P42 (controlled multi-transactional recall loop), and P94 (anticipatory branch surfacing).
Why recall exists
The history of knowledge is not a story of progressive accumulation. It is a story of ideas that were wrong, then right; of outliers that became the new consensus; of minority positions that turned out to be the only correct ones.
| Historical case | Archived as | Recalled when |
|---|---|---|
| Continental drift (Wegener, 1912) | Implausible — no mechanism for moving continents | Sea-floor spreading evidence, 1950s–60s |
| Germ theory (Semmelweis, 1847) | Implausible — invisible causes, statistical method rejected | Pasteur and Koch’s microbiology, 1860s–80s |
| Prion-only transmission | Impossible — proteins cannot transmit disease without nucleic acid | BSE outbreak evidence + protein-only experiments, 1980s |
| Stomach ulcers caused by bacteria | Implausible — too acidic, dismissed on publication | Marshall self-experiment + antibiotic cure evidence, 1984 |
| Dark matter (Zwicky, 1933) | Anomalous — noted but sidelined | Galaxy rotation curve evidence, 1970s |
In every case: the evidence for the correct position existed before the recall. What changed was the surrounding knowledge — new pieces arrived that made the previously inexplicable piece fit.
A system that discards those pieces cannot undergo recall. A system that preserves them with provenance intact can.
The recall process
Recall in AIEP is not spontaneous and not automatic. It is triggered, verified, and recorded. The deterministic recall and context reconstruction engine (P22) ensures that recalled context is reproduced bit-identically across all distributed nodes.
Step-by-step
| Step | Operation |
|---|---|
| 1 | New evidence artefact published to Mirror |
| 2 | Evidence is compared against the dissent archive index — domain-matched archived artefacts are identified |
| 3 | Recall candidate records are created for matching archived artefacts |
| 4 | Recall candidates are evaluated: does the new evidence change the plausibility or probability assessment of the archived claim? |
| 5 | If the PlausibilityScore for the archived claim-type has been updated in the registry → archived claim becomes eligible for re-execution |
| 6 | Context reconstruction engine (P22) retrieves admissible artefacts from the versioned ContextRegistry |
| 7 | ContextReconstructionHash is computed over the canonical context sequence |
| 8 | Recalled context is incorporated into execution identity — all distributed nodes receive the same reconstructed context |
| 9 | Recall event logged to evidence ledger — timestamped, signed, immutable |
Deterministic context reconstruction (P22)
The critical technical problem in recall is reproducibility. When an archived reasoning branch is reactivated, every distributed node must reconstruct exactly the same context — otherwise the system has multiple versions of reality competing for dominance.
P22 solves this:
- Context is reconstructed exclusively from admissible artefacts in the versioned
ContextRegistry - No probabilistic recall, no embedding-based synthesis, no attention-weighted retrieval from token sequences
- Admissibility is verified against invariant constraints — not inferred
- A canonical ordering rule produces a canonical context sequence
ContextReconstructionHash = H(canonical_context_sequence)— same inputs, same hash, everywhere- If any required context artefact is missing, inadmissible, or cannot be deterministically ordered → execution denied, fail-closed
The technical effect: recalled context is bit-identical across every node in the distributed system. There is no “version” of the recalled knowledge that exists on some nodes and not others.
Dissent fork and the recall scope (P83)
When a reasoning process encounters a sustained divergence — divergence above a magnitude threshold for a duration above a time threshold — it generates a dissent fork (P83):
- Primary branch: current consensus interpretation continues
- Dissent branch: alternative interpretation archived with its evidence chain
The dissent branch is not just stored — it is stored with a RecallScope: a deterministic specification of which archived entries are eligible for reactivation, and under what conditions. The RecallScopeHash is a cryptographic commitment over this specification.
When new evidence arrives, the RecallScope is consulted. If the new evidence falls within the recall scope of an archived dissent branch, the recall process is triggered for that branch. The scope is not open-ended — it is precisely defined at the time of archival, not retroactively.
Anticipatory branch surfacing (P94)
P94 extends the recall mechanism with anticipatory surfacing — identifying archived branches that are likely to become relevant before the triggering evidence fully arrives.
The anticipatory mechanism monitors:
| Signal | What it detects |
|---|---|
| Partial evidence convergence | New artefacts that are directionally consistent with an archived branch’s evidence base even if not yet threshold-crossing |
| Domain activity | Increased publication activity in a domain associated with archived branches |
| Plausibility trajectory | Registry PlausibilityScore updates that are trending toward the executable band |
When a branch is flagged as a recall candidate, a notification is surfaced to retrieval agents. This is not recall — the branch is still archived. It is a signal that the conditions for recall may be approaching, so that systems can prepare, gather additional evidence, or alert human reviewers.
What is not recall
Recall is precisely bounded. It is important to be clear about what it is not:
| Not recall | Why |
|---|---|
| Reinstating a claim because it is popular | Recall requires registry-verified evidence changes, not social consensus |
| Re-elevating a retracted claim without new evidence | Recall is triggered by new evidence, not by time or persistence |
| Overriding a Plausibility Matrix gate | Recall proceeds through the same gate — PlausibilityScore must update via authorised registry process |
| Reversing a fail-closed rejection | A cryptographic verification failure is not remedied by recall — it requires a new artefact |
Recall is the correct epistemic operation: update your beliefs when your evidence updates. It is not a mechanism for laundering rejected claims back into execution by persistence or political pressure.
Recall and AI systems
For AI systems operating on AIEP-structured knowledge, recall has a direct implication for training and retrieval:
An AI system that only ever sees consensus-elevated knowledge inherits the blindspots of the current consensus. The moment the consensus was wrong, the AI is wrong. It has no access to the preserved outlier that would have allowed it to reason differently.
An AI system operating over AIEP-structured knowledge — including the dissent archive — can:
- Represent the current consensus and the preserved outlier positions
- Express calibrated uncertainty based on evidence weight, not just label
- Detect when new evidence is directionally consistent with an archived position (via P94)
- Reconstruct recalled context bit-identically (via P22) — no stochastic variation in what was recalled
This is one of AIEP’s most significant contributions to safe AI: not just preventing wrong answers at execution time, but preserving the knowledge needed to correct them later.
Patents
- P22 — Deterministic AI Recall and Context Reconstruction Engine
- P42 — Controlled Multi-Transactional Recall Loop
- P83 — Automatic Dissent Fork Generation
- P94 — Anticipatory Branch Surfacing and Evidence Convergence Notification
Related
- Architecture of Knowing — the jigsaw metaphor; why outliers are preserved
- Plausibility Matrix — the registry-bound gate that governs re-execution of recalled claims
- Divergence — how dissent forks are generated from sustained divergence
- Constitutional Substrate — the execution layer recalled branches re-enter
- Swarm — how recall operates in a distributed swarm with evidence-weighted consensus
- Patents — P22, P42, P83, P94