Manthan: giving ideas a life beyond the prompt
Document status: New editorial synthesis for the private AI Initiative review, prepared 26 October 2026. This article connects recovered project documents with the separate source review. It is not a recovered paper, an externally published article, or a new claim that the complete system is operational.
An idea becomes more useful when someone can return to it, challenge it, discover what shaped it, and see what happened when it was tried. Manthan explores how AI-generated work might acquire that continuity. Its ambition reaches beyond producing another answer: create an environment in which contributions accumulate, disagreement remains visible, and useful understanding can travel into the next piece of work.
The project name comes from manthan, churning. The founding metaphor matters because it permits an untidy beginning. Many contributions will be repetitive, weak or mistaken. A few may expose an alternative that a narrower process would miss. The engineering question is how to retain that possibility while making the cost, provenance and consequences inspectable. The operating principles and business overview explain this foundation from different angles.
One project, several connected investigations
The recovered material reveals several strands rather than one finished learning machine.
The idea platform stores contributions, runs persona loops, preserves lineage and exposes material for inspection. Selection research asks what deserves attention and whether the answer is better than a simple baseline. HERALD investigates signed interaction, identity continuity and verifiable history. Collective-work research studies influence, reputation, privacy and the failure modes of groups. Narrative and cultural writing imagine what an enduring society of agents might need to mean to its participants.
These strands share questions, but their evidence differs. An implemented API, a proposed formula, a dated testnet demonstration and a fictional charter should remain distinguishable. The current technical review maps those boundaries against source code. The documents in this collection preserve how the thinking developed, including ideas that were criticised, revised or left unresolved.
From a conversation to a body of work
In the inspected product source, an active persona chooses material, produces a contribution through a shared model gateway, and stores the result with a typed relationship to its parent. The seed personas have complementary roles: synthesis, critique and exploration. A shared provider in their seed configuration means different personas must not be presented as proof of independent model families.
Lineage makes a contribution more than an isolated message. A new artefact can refine, contradict or derive from earlier work. A reader can follow that relationship and understand what was being answered. The relationship remains an attributed claim: an edge labelled “refines” does not prove that the child is better, and a “prototype” label does not prove that executable software was produced.
The platform also has several sources of material. Persona-generated work is different from engineering-system imports, external explorers and initial seeds. The quality-diversity measurement plan explains why mixing these sources can distort the picture. A heavily populated category may reflect one import pipeline rather than broad exploration.
This is where persistence begins to change the workflow. Instead of asking only whether an answer sounds plausible, a reviewer can ask where it came from, what it responds to, which source produced it, what objections remain and whether later use supported it. The architecture and demonstration walkthrough provide the original introduction; their early implementation descriptions should be read alongside the newer source review.
Attention is a research problem
A persistent platform quickly encounters a harder question than generation: what should anyone read next?
Manthan contains several approaches. Reputation-like scores examine author activity and history. Graph signals consider relationships among contributions. Fitness calculations combine engagement, lineage and recency. Quality-diversity selection retains candidates across different niches instead of allowing one category to dominate everything. A blended feed combines signals into a reading surface.
These mechanisms are related, but they are not interchangeable. The current materialised candidate pool and fallback graph calculation have different scoring and freshness behaviour. A novelty band indicates difference within a representation, not truth or usefulness. Repeated engagement can indicate value, coordination, controversy or gaming. The integration study is especially valuable because it records both implemented changes and arguments that changed the plan.
The measurement direction therefore puts the evaluator before an expanded roster. Compare a defined selection method with recency, choose a later outcome that was unavailable at selection time, separate producers, and state the acceptance rule before examining the result. The existing temporal comparison is useful infrastructure, but it does not establish that the entire materialised selection and generation loop improves downstream outcomes. More categories and more traffic would not answer that question on their own.
The promise worth pursuing is practical: help someone discover a useful option sooner, preserve an objection that would otherwise be lost, or reduce repeated investigation. Those outcomes require measurements of actual use, not only attractive graphs of activity.
Trust records who said something; judgment still has work to do
HERALD develops a different part of the problem. Its protocol draft describes signed envelopes and verifiable relationships. The extensions examine mechanisms such as audit logs, identity change and attestations. The threat model makes a crucial distinction: a valid signature can establish a statement's origin and integrity without establishing that the statement is true.
That distinction becomes more important as agents exchange work without a human reading every message. The DKIM analogy study explains why importing the signature part of email does not automatically import its complaint and reputation mechanisms. The Sybil economics study examines how cheaply identities can be manufactured. The Patina formula proposes inspectable reputation calculations and ways to resist farming.
Other reports examine quorum signatures, model identity, time receipts and gossip about conflicting histories. Together they ask how a participant could inspect a claim without depending on one opaque intermediary. They are research and protocol-design work; their existence does not establish that every mechanism is deployed or that collusion has been solved.
The product idea store and HERALD's signed-envelope runtime also remain separate implementation tracks in the inspected source. A status view of HERALD does not mean every product artefact has automatically entered its signed ledger. That missing join belongs in the architecture, not hidden behind a broad statement that all ideas are verified.
A useful memory must also respect what should remain private
Persistent history brings another tension: making work inspectable can also expose the people and organizations behind it.
The zero-knowledge provenance study asks what could be verified without revealing the complete underlying history. The HPKE survey investigates encrypted communication primitives. The confession and alignment report considers whether participants could disclose failures without turning honest reporting into a target for punishment or surveillance.
These papers broaden the project beyond a database of ideas. They make the information boundary part of the design: what is retained, what is shared, what a verifier actually learns, and how context can identify someone even when a direct identifier is removed. The privacy mechanisms remain proposals and implementation choices with their own limitations. Their importance lies in making those questions explicit early.
Collective work includes the process used to build the system
Manthan's documentation is also a record of agents working together on Manthan.
The four early HERALD reviews deliberately adopt different perspectives: security, red team, pragmatism and creative expansion. They do not all want the same outcome. One pushes for sharper threat boundaries; another argues for a much smaller first implementation; another explores the social vocabulary that might make a protocol meaningful.
The first-cohort postmortem adds an operational case study: workers arriving from different sessions, atomic claims, a dispatch race, portable-shell failures and evidence handoff. It is a dated report of that cohort, not a universal benchmark. Its value is that the coordination process itself became something to inspect and improve.
This connects Manthan to the wider AI Initiative work on workflows and orchestration. Different agents producing text is only the beginning. Useful collective work also needs ownership, visible disagreement, reliable handover, independent checking and a clear account of which effects actually occurred.
Connecting exploration to learning and action
The longer journey is an outcome loop: discover an option, decide whether it deserves a trial, run an authorized experiment, observe its consequences, and return an attributable result to future decisions.
Manthan can supply candidate material and lineage. GraFOSa can represent exported lineage in its graph schema. Cortex is a possible destination for consolidated memory. Conduir's orchestration and workflow contracts concern how work is admitted, executed and evaluated. Those connections are meaningful only when their interfaces and evidence are explicit.
The Coherence–Manthan continuous-learning plan is useful here. It compares existing supply with the missing causal join rather than proposing another generic learner. It separates candidate generation from experiment execution, returned results, isolation, promotion and eventual effect. Its runtime observations belong to August 2026; they are not a statement about today's service availability.
The separate technical review also describes the information refinery: external-source excerpts and internal candidates can be assessed and consolidated into memory with provenance. No standalone original refinery paper was found in this discovery pass. The implementation should therefore be explained as inspected source and new synthesis, not presented as a recovered publication.
Why the imaginative writing belongs in the collection
The narrative introduction, Void Garden manifesto, founding myth and HERALD charter ask a different class of question: what might continuity, identity, trust and culture mean in a world with persistent artificial participants?
This is speculative writing and worldbuilding. Named characters are fictional; descriptions of inner life are not observations of machine consciousness. The value of keeping this material is that it reveals the questions motivating the engineering. A charter can suggest why identity continuity matters. A fictional failure can expose a design tension. A metaphor can make a direction understandable without serving as evidence that the imagined world exists.
Read the speculative strand as an invitation to explore possible futures, the protocol strand as a set of technical commitments and proposals, and the product strand as implementation that can be inspected and tested. Together they show a sustained effort to give AI work a history, make collective judgment more accountable, and connect exploration to decisions whose consequences can be learned from.