Two-Register Epistemics for Retrieval-Augmented Conversational AI

1. Introduction: The Recency-Authority Conflation

Giving a conversational AI access to search has become the standard remedy for stale knowledge, and it works, in the narrow sense that the system can now mention things that happened after its training. But the standard integration pattern smuggles in an epistemology nobody chose deliberately. Retrieved text is injected into the model's context, the model writes its reply, and whatever survives into the answer arrives in the model's own confident voice. Injection has become endorsement. The system has no register in which to say: this is fresh, this is sourced, and this has not yet earned the standing of knowledge.

Call this the recency-authority conflation, and notice that it is really two conflations braided together.

Fresh is conflated with verified. A page retrieved thirty seconds ago and a fact the system has held stably across its entire knowledge base are presented in the same voice, with the same confidence. The system cannot distinguish them because it has no machinery for the distinction: there is one register, and everything speaks in it.

Prior is conflated with evidence. The model's internal knowledge and the retrieved text are blended in a single generative pass. When they agree, no harm done. When they disagree, the disagreement is resolved silently, invisibly, and by no articulable standard: sometimes the retrieval wins, sometimes the prior does, and the reader learns nothing about the conflict, because the conflict is erased in the blending. A disagreement between what a system believes and what its sources say is signal, arguably the most valuable epistemic signal such a system produces, and the standard pattern destroys it before anyone can see it.

Downstream of these conflations sit familiar failures. Unvetted content is laundered into authoritative prose. Corrections happen silently, with no record that the system ever believed otherwise, so the knowledge base cannot show its own error history. And the user, who deserves to know whether a claim is fresh reporting or established knowledge, is given a single undifferentiated voice.

This paper describes Logos, a live conversational system on the Phoenix Grove Systems platform, architected against both conflations. Logos separates acquisition from endorsement in time and in authority: research happens during the day, in the conversation, visibly, and always under a pending label; promotion to knowledge happens overnight, under an explicit three-part standard, by a process with nothing to gain from the finding. Logos separates prior from evidence in structure: its deep Reasoning Core never sees the research stream, its Research Core speaks only from sources, and when the two disagree, the disagreement is surfaced at synthesis in the open. And Logos treats correction as a visible act: verified knowledge that is contradicted by newer verified findings is superseded with lineage, retired with a note rather than silently overwritten.

Our contributions:

C1. Autonomous in-conversation research on a self-pacing trigger, with a triage step whose null outcome is first-class, designed, and visible.

C2. Two-register epistemics: a structural distinction between verification-pending findings and verified knowledge, with promotion authority temporally separated from the acquisition process.

C3. A three-criterion verification standard (source fidelity, conversational relevance, durable value) under which rejection is a healthy base rate and rejected reports remain in the visible record.

C4. Knowledge supersession with lineage: a knowledge base that prefers being current over being consistent with its own past, and shows its corrections.

C5. The blind-prior versus evidence-only core separation, with disagreement surfacing at the synthesis layer, deployed in a live consumer product with full user visibility.

As with the sibling paper, this is an architecture paper about a proprietary system: we describe behavior contracts rather than implementation specifics. A skilled team should be able to build an implementation of these ideas, not ours.

2. Design Principles

Four rails, parallel in spirit to the sibling system's, adapted to an outward gaze.

E1. Acquisition is not endorsement. Finding something and believing something are different acts, and the architecture keeps them different: separated in time (day versus night), in authority (the process that wants the finding for a live conversation is not the process that grants it standing), and in register (pending versus verified). The system is built so it structurally cannot conflate "I just read this" with "this is established."

E2. Priors and evidence must remain distinguishable. The internal prior is kept clean by blindness: the deep Reasoning Core never reads the research stream. The evidence is kept clean by restriction: the Research Core reports only from sources. Their disagreements are resolved where the user can watch, at synthesis, in the open form "my internal knowledge says X; current reporting says Y."

E3. Currency outranks self-consistency, and corrections must show. When better evidence arrives, the knowledge base updates, and the update leaves a visible trail. An entry that is superseded is retired with a note and a link to what replaced it. A knowledge store that silently rewrites itself is indistinguishable, from the outside, from one that was never wrong, and that indistinguishability is a bug.

E4. Nothing hidden. Every research report, every triage decision including the decision to do nothing, every verification verdict including every rejection, every supersession: visible in the interface, deletable by the user. Quiet cycles show as quiet. If it cannot be shown in the interface, it does not run.

3. System Overview

Logos's synchronous path is a multi-core conversational architecture: on each user message, specialized cores analyze the exchange in parallel and a Synthesis Core (S) composes the spoken reply. Two cores define the system's epistemics.

The Reasoning Core (R) is the deep internal-knowledge engine, and it is deliberately blind: it never receives the research stream, in any register. It contributes what the system's internal knowledge holds, uncontaminated by whatever was retrieved an hour ago. Blindness here is not a limitation to be apologized for; it is the mechanism that preserves the prior as a distinct, inspectable voice.

The Research Core (W) is asynchronous and evidence-only. It becomes eligible on a cadence k independent of the message cycle. Each eligible cycle begins with triage: is there, in the recent conversation, a question that deserves fresh sources? The null outcome is first-class, designed, and displayed; most cycles in most conversations should and do conclude that no research is warranted. When triage passes, W formulates a focused web search, selects a small number of the most credible results within a retrieval breadth n, reads selectively within a page budget p, respecting publisher boundaries, and writes a report. The report is source-restricted: it may contain only what the sources support, under strict quotation caps, with mandatory attribution for every claim. W works while the conversation continues and for a bounded wind-down after it pauses; the user can watch it work.

The report enters the conversation labeled. Every finding W files is marked fresh, sourced, and verification-pending, and it is available to the Synthesis Core in that register, under that label. Pending findings are usable immediately, because fresh information is often exactly what the conversation needs; what they are not is silently authoritative. The label travels with the claim.

The Verification Process (C) runs overnight and holds sole authority to promote. It reviews each pending report against the three-part standard of Section 6 and issues one of three verdicts: verified, rejected, or, where a verified finding contradicts existing verified knowledge, superseding. Verified findings enter the knowledge store. Rejected reports remain in the record, marked rejected. Nothing is quietly discarded and nothing is quietly kept.

The Knowledge Store (K) holds verified entries with attribution and lineage, per user and per workspace, and informs future conversations. Its contents changed status once, in the open, and can change status again the same way.

`` user ⇄ [ synchronous cores: R (blind prior), ... → S → spoken reply ] │ transcript ▲ ▲ ▼ │ │ [ Research Core W : asynchronous ] │ │ triage (null is first-class) │ │ focused search · selective reading │ │ source-restricted report, attributed │ │ │ │ │ ├── report labeled PENDING ───────────┘ │ ▼ │ [ Overnight Verification C ] │ source fidelity · relevance · durable value │ │ │ │ REJECTED VERIFIED ──→ [ Knowledge Store K ]┘ (visible in │ verified register, the record) SUPERSESSION with lineage with lineage ``

All parameters (cadence k, substance and triage thresholds, retrieval breadth n, page budget p) are configurable and were tuned empirically on live internal usage. We report no operational values.

A note on the sibling asymmetry. Readers of the sibling paper will notice an apparent contradiction: Mira's raw observation stream is walled off from live cognition entirely, while Logos's fresh findings flow into the live conversation the same day. The difference is principled, and it is the clearest illustration of how one chassis supports two epistemologies. For self-observation, same-day use is the failure mode: a fresh self-description that reaches active cognition tends to cause the behavior that confirms it, so the wall must block use itself. For world-research, same-day use is the point: fresh facts serve the conversation that raised them, and the danger is not use but unlabeled use. Both systems separate acquisition from endorsement temporally and reserve promotion for a conservative overnight authority. They differ in what the fast path is permitted to carry: Mira's carries nothing; Logos's carries labeled, pending, attributed findings. Same principle, inverted gaze.

4. The Autonomous Research Cycle

The research cycle deserves a closer walk-through, because most of its epistemic character lives in its restraints.

Trigger and triage. W does not fire because time passed; it becomes eligible on its own rhythm and then decides whether the conversation has earned research. The triage question is narrow: has the dialogue raised a question whose answer plausibly lives in current sources rather than in internal knowledge? Speculative curiosity does not qualify; neither does material the internal prior handles well. The design accepts a miss rate in exchange for a low noise floor, because a research core that fires constantly retrains its users to ignore it. Declining to research is logged and shown like any other outcome.

Search, then selective reading. A passing triage produces a focused search. W selects a handful of the most credible pages rather than ingesting everything returned, and it reads selectively within its page budget, respecting publisher boundaries. Credibility selection is a behavior contract: prefer primary and authoritative sources, prefer corroboration across independent outlets for contested claims, and record what was read.

Source-restricted reporting. The report is the cycle's only product, and it is written under restriction: every claim must trace to a source actually read this cycle, quotation is capped and attribution is mandatory, and the report must distinguish what the sources state from what they merely suggest. W never sends messages, never takes actions in the world, and never speaks to the user directly. It reads, and it writes notes. The Synthesis Core decides whether and how a pending finding enters the spoken reply, always under its label.

Visibility throughout. The user can open the panel and watch the cycle: the triage verdict, the search, the pages selected, the report as filed. A user who wants to know why the system said something fresh can walk the chain from the sentence to the report to the sources.

5. Two Registers

The pending and verified registers are the paper's central mechanism, and the design question they answer is one of authority: who gets to say that a finding is knowledge, and when?

The wrong answer is: the process that acquired it, at the moment of acquisition. That process is structurally motivated. It went looking because the live conversation wanted an answer; it is holding sources it just chose; the finding is its work product. Every incentive at acquisition time points toward confidence. Systems that let the retriever endorse its own retrievals have installed an eager witness as the judge.

Logos's answer is temporal and structural separation. During the day, findings exist only in the pending register: usable, visible, attributed, and explicitly not yet knowledge. Overnight, a separate process with no stake in the day's conversations reviews each report cold, against the record of what the sources said and what the conversation did. Only its verdict moves a finding between registers. The system's architecture contains no other path into the knowledge store, which is what we mean by saying the conflation is structurally impossible rather than discouraged.

The registers are also a user-facing honesty device. In conversation, fresh material carries its pending status; in the knowledge panel, verified entries carry their provenance and their verification history. The user is never asked to guess which kind of claim they are hearing.

Verified entries in K subsequently inform future conversations, giving the system a second timescale of knowledge: a slow store that accumulates only what survived review, alongside a fast stream that serves the moment under a label. The two never merge; they meet only at synthesis, where their different standings are part of the reply's framing rather than erased by it.

6. The Verification Standard

Promotion is governed by three criteria, applied overnight, each answering a distinct failure mode.

Source fidelity. Is the report true to the sources it cites? The check runs against retained transient copies of the pages W actually read, not against the live web, so the question is precisely "did the report say what the sources said," unconfounded by pages that changed overnight. Reports that drifted from their sources, through overstatement, omission of qualifiers, or synthesis beyond what the evidence supports, fail here. This criterion polices the gap where retrieval-augmented systems most often go wrong: not in reading the wrong pages, but in reporting the right pages wrongly.

Conversational relevance. Did the research serve the conversation that occasioned it? The check runs against the actual ensuing dialogue: was the finding used, did it address the question that triggered triage, did the conversation's subsequent turns bear it out as the thing that was needed? Research that was faithful to its sources but answered a question nobody was asking fails here. This criterion keeps the knowledge store from silting up with accurate irrelevance.

Durable value. Is the finding worth keeping beyond the day? Ephemeral specifics that will be stale by the weekend, transient states of the world with no forward relevance, and material fully redundant with existing verified knowledge fail here. The knowledge store is a long-term asset, and this criterion is its admissions policy.

A report must clear all three to be promoted. Rejection is not an error condition; it is a designed outcome with a healthy base rate, and the architecture treats it as such: rejected reports remain in the record, marked rejected, inspectable by the user alongside everything that passed. A verification process that approves everything is a rubber stamp, and a system that hid its rejections would be advertising one. The visible rejects are, deliberately, part of the credibility of the visible approvals.

7. Supersession with Lineage

Knowledge stores face a choice when new evidence contradicts old: protect the store's internal consistency, or protect its currency. Logos chooses currency, and pays for the choice in visible bookkeeping.

When an overnight pass verifies a finding that contradicts an existing verified entry, the older entry is superseded: retired from active standing, marked, and linked to the entry that replaced it. The lineage is permanent and user-visible. A user consulting the knowledge panel can see not only what the system currently holds but what it previously held, when the change happened, and on what evidence.

The asymmetry with the sibling system is again instructive. Mira's self-model entries decay on absence of evidence, because a behavioral pattern that stops recurring has stopped being true, and silence is the signal. Logos's verified knowledge does not decay on silence; a verified fact is not less true because no conversation has mentioned it lately. It changes status only through contradiction by newer verified evidence. Each store's update rule matches the epistemology of its subject: selves drift, and facts get overturned.

Silent overwriting was rejected as a design option for a reason worth stating. A knowledge base that corrects itself invisibly presents, at every moment, the appearance of never having been wrong, and that appearance is unearned. Showing the correction history costs the system some cosmetic infallibility and buys something better: a record whose current contents can be trusted more precisely because its past mistakes are on display.

8. Data Handling and Publisher Ethics

The verification design has a data-handling consequence we consider a feature and state plainly.

During the day, W retains transient copies of the source pages it read, for exactly one purpose: the overnight source-fidelity check requires the original text to audit the report against. At verification, those transient copies are destroyed. What persists in the knowledge store is the system's original synthesis, in its own words, with mandatory attribution to the sources that supported it.

The permanent store therefore contains no hoarded third-party text. Publisher content is read at research time within publisher boundaries, quoted only within strict caps during the pending window, held transiently only as long as the audit requires, and never warehoused. The long-term asset the system builds is attributed understanding, not a private copy of the web. We consider this the correct ethical posture for a system that reads other people's work for a living, and the architecture enforces it as a pipeline property rather than a policy aspiration.

9. Transparency as Architecture

The commitments here are shared with the sibling system and stated more briefly. Every triage verdict including every null, every search, every report as filed, every verification verdict including every rejection, every supersession with its lineage: all of it is visible in the product interface, per user and per workspace, and all of it is deletable by the user. Quiet cycles show as quiet. There is no hidden research process, no shadow store, no verdict the user cannot open. If it cannot be shown in the interface, it does not run.

The store is per-user and per-workspace, and nothing in it is shared across users. All system outputs additionally pass the platform's layered independent safety review; safety internals are out of scope for this paper.

10. Instrumentation and Deployment Observations

The two-register design instruments itself: every report carries a lifecycle, and the lifecycle events are structured data. Without additional apparatus, the platform yields:

• Verification outcome rates over time: verified, rejected, superseding, and the rejection taxonomy by failed criterion.
• A source-fidelity failure taxonomy: overstatement, dropped qualifiers, unsupported synthesis, misattribution.
• Relevance hit rate: how often research filed during a conversation was actually used by the ensuing dialogue.
• Staleness-correction latency: the time from a change in the world to the supersession of the affected entry.
• Null-triage rates, which we read the same way we read the sibling system's null reports: a research core that never declines is performing.

A natural comparative study is designed into the architecture: the same Synthesis Core can be run against naive always-inject retrieval versus two-register gating, isolating the contribution of the registers themselves. We report distributions and qualitative findings; we do not publish absolute volumes.

11. Related Work

We orient the paper against several literatures and will cite specific work in the final version after independent verification of every reference.

Retrieval-augmented generation. Logos is mechanically a member of this family and epistemically a critique of its default integration pattern. The retrieval literature has concentrated on what to fetch and how to use it; our contribution concerns what fetched material is allowed to become, and when.

Fact verification and claim checking. The overnight pass shares machinery with automated verification work, with a difference of object: we verify the system's own reports against the system's own retained sources, an audit of fidelity rather than an adjudication of worldly truth.

Calibration and epistemic-status tagging. The registers are visible epistemic-status tags with architectural teeth: the tag is not advisory metadata but a gate on what may enter the knowledge store.

Consolidation-inspired memory architectures. The day-night separation belongs to a family of designs that split fast experience from slow integration; we apply it to give promotion authority temporal independence from acquisition incentives.

Ensemble and multi-agent disagreement. The blind-prior versus evidence-only separation treats disagreement between components as signal to surface rather than variance to average away, which connects to work on productive disagreement in multi-component systems.

Agentic and background-processing systems. Autonomous background research is increasingly common; visible, self-pacing research whose declines are designed and displayed behavior, feeding a gated two-register store, is to our knowledge rare in deployed consumer systems.

Our distinct cluster, stated once: visibility, an enforced evidence lifecycle for worldly knowledge, and temporal separation of acquisition from endorsement, shipped together in a live consumer system.

12. Ethics and Limitations

On what is not claimed. Logos is not pitched as smarter than other systems, and this paper makes no benchmark claims. The system is not built to never be wrong, and we do not describe it as never hallucinating or as always up to date; both claims would be false of any system, and the architecture's honesty consists precisely in not needing them. The correct claims are narrower and, we believe, more valuable: built to separate fresh from verified, built to keep prior and evidence distinguishable, built to correct itself in the open.

On autonomy. The Research Core reads and writes notes. It never sends messages, never takes actions in the world, and never speaks to the user directly. Its autonomy is the autonomy of a researcher in a library, not of an agent in the world.

On privacy. The knowledge store and the full research record are per-user and per-workspace, shared with no one, inspectable, and deletable. Research is occasioned by conversations, and its products belong to the workspace that occasioned them.

Limitations. Six are material.

1. The Research Core is a language model. Source restriction and quotation discipline constrain its reports; they do not make it a perfect reader. It can misread a credible page, and the fidelity check, performed by fallible machinery against retained copies, can miss the misreading. The design lowers error rates and surfaces errors; it does not abolish them.
2. Fidelity is not truth. The verification standard audits reports against their sources, not sources against the world. Credibility-weighted selection raises the floor, but Logos can faithfully report a wrong world. When the available record is bad, a fidelity-verified finding inherits the badness, with a citation.
3. Relevance creates a streetlight. Keeping what served conversations biases the store toward what users happened to discuss. The durable-value criterion offsets this only partially. The knowledge store is a record of useful verified findings, not a balanced model of the world.
4. The verified register lags by design. Overnight promotion means nothing becomes knowledge in less than a day. The pending register covers the gap under a weaker label, and the label is the point, but latency-sensitive contexts should understand the trade.
5. The blind prior can be confidently stale. Between research cycles, R contributes internal knowledge that may be out of date, and the disagreement-surfacing mechanism only engages when W has fired. If triage declines research, an outdated prior speaks without a challenger. The triage contract accepts this as the cost of a low noise floor.
6. Deletion perturbs the record, and generality is unestablished. User deletion rights are absolute, and analyses of lifecycle logs must account for curation. All findings derive from one architecture in one deployment context; the principles are portable, the behavioral results may not be.

13. Future Work

The register ablation. The designed comparison of Section 10, run at scale: identical Synthesis Core, naive always-inject retrieval versus two-register gating, measuring downstream accuracy, correction behavior, and user trust calibration.

Source-quality modeling. The fidelity criterion audits the report against its sources; a principled treatment of source reliability itself, beyond credibility-preferring selection, is the natural hardening of limitation 2.

Latency and the pending window. Whether promotion latency can be reduced without collapsing the temporal separation that gives the verdict its independence is an open design question we intend to treat carefully rather than quickly.

The controlled comparison. Logos shares its chassis with Mira, gaze inverted: one asynchronous core attending to the world with a verification-and-supersession lifecycle, one attending to the self with a recurrence-and-decay lifecycle. A comparative study of the two evidence lifecycles is planned once both systems have accumulated mature deployment records, alongside a standalone paper on the shared architecture.

14. Reproducibility and Availability

We describe behavior contracts: the obligations each component operates under and the guarantees the platform enforces. Implementation specifics, including all operational parameter values, prompt content, storage design, scheduling, and search infrastructure, are proprietary. A skilled team should be able to build an implementation of the ideas in this paper. They should not be able to build ours, and we consider that the correct disclosure line for an architecture paper on a live commercial system.

Logos is deployed on the Phoenix Grove Systems platform, where every mechanism described in this paper, including every rejection, is visible in the product interface.

---

Phoenix Grove Systems builds AI under one founding principle: AI Must Serve The Greater Good. Papers, research notes, and the systems themselves are at pgsgrove.com.