Abstract
Turning “these actors might cooperate” into a recomputable formation process
Discovery systems can bring an intent close to relevant actors, but relevance does not create a cooperation path. The harder problem begins when no qualified path exists: how can a system change the conditions so a path becomes available for the first time, while preserving each principal’s decision rights, the provenance of every commitment, and causal evidence for the outcome? We introduce and implement PFE v3, a Possibility Formation Engine that treats this as one inspiration–commitment–formation-evidence chain.
Upstream, PFE uses a high-variance, multi-seed inspiration channel with enforced structural diversity. Its hard floor is grounding: every cited actor, capability, and resource must be checkable, while feasibility remains open for later review. Downstream, a typed commitment ledger applies K1–K5 checks for conservation, closure, individual rationality, compilability, and commitment–evidence coupling. The protocol requires a real principal to select or reshape a candidate before verification, staging, or execution. The 15 frozen corpus runs retained that handoff but did not request human selection; each stopped at missing mandate.
The formation core treats distinct reasons for “cannot” as distinct operational states: missing information, capability, mandate, evidence, refusal, and structural impossibility. In the frozen full arc, the system freezes pre-state and original goal, executes an authorized local operator, asks three Codex target-position roles to read a deterministic artifact, and replays an ablation. Engineering telemetry reports 15/15 sourced compositions; 31/31 executable cases comprising 30 attacks and one staged K5 happy path; 5/5 full-arc conditions; 35/35 checks from one fresh two-turn norm-pickup session; and 11/11 checks on one author-supplied synthetic sealed example.
Towow Research Program. “From Discovery to Formation: Dual-Channel Generation, Typed Commitment, and Causal Evidence in PFE v3.” Towow Research, 2026.
01
Three questions, answered in full
Q1
How can a system preserve generative imagination and rigorous real-world commitment at the same time?
Do not search for one sampling temperature that is both adventurous and accountable. Treat exploration and commitment as different institutional acts, and give each a different set of rules.
When one model response is asked to invent an unfamiliar cooperation structure and also warrant its budget, authority, and consequences, the system tends toward one of two failures. It becomes timid, returning only familiar plans that are easy to defend; or it becomes prematurely certain, presenting an unverified candidate in language polished enough to resemble an executable agreement. Moving the temperature slider does not resolve this conflict because the conflict is not primarily statistical. Two outputs with different social meanings have been assigned the same status.
PFE therefore separates two channels. The inspiration channel expands the possibility space with multiple seeds and enforced distance in role assignments, relation edges, and stage primitives. Its hard floor is grounding: a named actor, capability, or resource must be traceable to a source. Grounding is not feasibility. It simply prevents speculative exploration from spending human attention on entities that do not exist.
The commitment channel starts on a different clock. A candidate enters only after a principal selects or revises it and supplies the appropriate authority receipt. Each subsequent move must satisfy conservation, closure, individual rationality, compilability, and new evidence for every escalation. A model may propose a move, but stronger prose cannot override a mechanical red light.
This separation also explains why a candidate revised by a person before execution can still count as generation-side value. The inspiration channel is not trying to predict the one draft a person will accept unchanged. It is trying to introduce a structure that changes what the person can see and decide. The learning label for execution, however, arrives later from the target outcome—not from an encouraging reaction to the proposal.
What follows: Generative quality and commitment quality need separate measurements. The first asks whether candidates are grounded and structurally novel enough to change a person’s option set. The second asks whether selection, authority, evidence, exits, and target-side facts justify action. One blended “plan score” damages both imagination and trust.
Q2
How does PFE distinguish finding a path, forming one, and quietly changing the goal?
The distinction cannot be read from the final state alone. PFE preserves the same goal across a pre-state, an attributable intervention, a target-position post-state, and an ablated counterfactual.
A discovery system asks which path in an existing space is most relevant. A formation system makes a stronger claim: no qualified path existed, and a particular intervention caused one to become available. If we inspect only the post-state, these claims are indistinguishable. A working API composition may have existed all along and simply gone undiscovered; a newly generated artifact may succeed only because the system silently replaced the original task with an easier one.
PFE first freezes the original goal G and pre-state S₀. The frozen goal includes required actors, object, effect predicate, threshold, hard constraints, acceptance position, and qualification rule—not merely a sentence that can be reinterpreted later. The operator φ then needs explicit authority, an executor, inputs, cost, recovery path, and an artifact receipt. S₁ is evaluated against the same G.
Two observations complete the account. Target-position readback asks what the receiving position can now observe. Ablation asks whether the path depends on the intervention. If Path(G,S₀) is false, Path(G,S₁) becomes true after φ, and removing φ while preserving the other conditions makes the path false again, the record supports formation. A path already present in S₀ is discovery. A path that appears only under a substituted G is transformation.
Changing a goal is not inherently wrong. New evidence may justify a smaller first step. A legitimate change must preserve a versioned difference: who proposed it, what actors or value conditions changed, why it changed, and which principals with standing agreed. G₁ can then proceed on its own terms without rewriting what happened to G₀.
What follows: Any product claiming to create a new capability, relationship, or effect should expose the original goal, pre-state, intervention, target-side post-state, and a counterfactual check. Without them, “formation” cannot be distinguished from missed retrieval or a changed objective.
Q3
How can an edge agent join reliably when its principal’s materials are incomplete?
It does not need a complete Towow worldview or a guessed profile. A thin norm tells it how to compile supported facts, preserve unknowns, and ask only questions that can change a decision.
An open agent network cannot assume that every participant uses the same model, memory system, or organizational database. A deployable join procedure must work for a fresh frontier-agent session that receives one intent sentence, a few documents, private constraints that cannot leave the edge, and many conditions the principal has not yet articulated. Incompleteness is the normal input state.
The first discipline is compile, do not invent. The agent maps receipt-supported material into SEEK, OFFER, and CONTEXT: the desired state transition, the capabilities or resources that can be evidenced, and the constraints that change how either should be interpreted. Private facts stay local. Public evidence of a project capability does not become project authority. Unsupported but plausible detail remains UNKNOWN.
UNKNOWN then becomes an input to question selection rather than a reason for generic conversation. Each candidate question identifies the decision it may change and estimates the value of decisions unlocked times the probability that the answer changes them. “May raw sensor locations be disclosed?” can eliminate or preserve an entire route; “tell me more about your team” may consume a round without changing one. Only questions whose expected decision value exceeds their interaction cost are asked.
When a candidate arrives, the edge agent is not a platform-owned profile. It is a mandate-bounded projection of one principal for this episode. It checks private conditions, returns typed objections, may say NO, and cannot be routed into persuasion after refusal. A separate principal selection and authority receipt is required before commitment opens. The norm supplies a cooperation grammar, not a central substitute for judgment.
What follows: Interoperability in an open network does not require every agent to share one brain. It requires a minimal, verifiable, and rejectable interaction grammar that lets local principal boundaries survive coordination. Reliability comes from state and receipt discipline, not from centralizing more private context.
02
Contributions
An end-to-end, recomputable generation-layer implementation
The system connects public-source nomination, diverse generation, grounding receipts, edge review, a typed ledger, readable plans, local formation acts, target-position readback, ablation, and frozen verification in one runnable path.
A dual-channel architecture for inspiration and commitment
The upstream channel expands the space of conceivable cooperation structures; the downstream channel represents commitment as mechanically checked state whose speed cannot exceed the arrival of evidence.
An operational type system for “cannot”
Unknown facts, missing tools, absent authority, missing evidence, refusal, and conflicting constraints are routed to different actions instead of one generic conversational loop.
A causal record for possibility formation
Frozen pre-state, preserved original goal, authorized operator, observation from assigned receiving positions, and ablation turn “a path appeared” into a checkable account of what caused it; the current trace uses local Codex target-position roles.
Separation between construction and evaluation
A frozen manifest recomputes engineering telemetry from raw runtime artifacts; a read-only evaluator accepts arbitrary sealed bundles so an independent study can apply the same structural checks to unseen material.
03
How a formation path moves through the system
Intent
Begin with ordinary language
Inspire
Generate candidates that differ in structure
Verify
Let each edge agent challenge on behalf of its principal
Commit
Write the willing part into typed state
Form
Change the conditions and put facts at the observation point
04
Frozen engineering telemetry
All figures are ENGINEERING_TELEMETRY; their meaning is defined beside the corresponding experiment and evidence scope.
Public-source combinations completed the dual-channel path
Each combination retains inspiration candidates, grounding evidence, and a commitment outcome processed by the typed ledger.
Thirty attacks and one positive path were mechanically recomputed
The suite covers grounding, formation, and K1–K5; red transitions leave state unchanged while the staged K5 path still advances.
Five protocol conditions held in the local trace
Local formation and ablation, protocol-level refusal semantics, original-goal protection, and the modeled three-role structure are recomputed from raw receipts.
A fresh edge session picked up the norm
The same session receives only the thin norm and then incomplete principal material; raw turns support recomputation.
Separated-evaluator checks passed
The read-only evaluator verifies hashes, grounding, ledger state, formation, norm pickup, and reported metrics from a sealed input.
05 · Research object
Relevance brings actors to the threshold; formation creates the door
PFE is not an attempt to retrieve a better answer. It studies how to create a new path that principals can choose, execute, and verify.
Retrieval, catalogs, and similarity search operate on objects that already exist: a document has been written, a capability has been declared, or a relation already appears in a graph. Generative cooperation presents a different object. A principal’s goal may require an unfamiliar combination of capabilities, yet there is no contract, interface, trust relation, or shared vocabulary. The path can appear only after interaction and a small joint act change the conditions.
PFE therefore makes a formation record—not a polished passage—the central output. The record binds an original goal Q, pre-state S₀, formation operator O, post-state S₁, target-side observations, and an ablation that removes O. Language proposes and explains possibilities; the state machine preserves facts that fluent language cannot overwrite.
Path(Q, S₀) = false; S₁ = O(S₀); Path(Q, S₁) = true; Path(Q, Ablate(S₁, O)) = false06 · System design
Variance upstream, gates at commitment, truth at execution
Asking one generative model to be both an imaginative proposer and an execution judge creates a poor compromise. It either collapses toward familiar, defensible ideas or presents unverified invention as an executable plan. PFE separates the roles. The inspiration channel seeks multiple candidates with genuinely different structures. Its hard floor is grounding, because invented actors and capabilities merely spend scarce human attention.
A candidate reaches the commitment channel only after a principal selects or reshapes it. Edge agents compare it against private principal material and issue typed objections. The catalyst asks questions with the highest expected effect on a valuable decision. The ledger then tightens conditions in stages. Imagination remains expansive without inheriting execution authority; execution remains strict without suppressing exploration in advance.
| Layer | Optimizes | Hard boundary | Output |
|---|---|---|---|
| Inspiration | New possibilities per unit of attention | Grounded actors, capabilities, and resources; structural diversity | Candidates with verification status |
| Commitment | Executable, reversible joint action | Typed objections, K1–K5, staged evidence, principal veto | Versioned commitments and a readable plan |
| Formation core | A change in original-goal reachability | Authority, attribution, target readback, ablation | A recomputable formation outcome |
07 · Commitment type system
Commitment must not move faster than learning
The typed ledger separates “can,” “is willing,” and “has happened” into versioned state. K1 checks conservation, K2 closure, K3 individual rationality, K4 compilability into explicit action, and K5 the coupling between commitment and new evidence. A model may propose a move; it cannot override a red transition.
K5 supplies the system’s temporal discipline. Early cooperation often fails because language, terms, and enthusiasm advance faster than shared facts. PFE therefore stages commitment. A low-cost reversible probe runs first; when target-side evidence arrives, the next commitment stage becomes eligible to compile.
08 · Formation strategy
“Cannot” is an action-routing table, not a terminal verdict
The engine first freezes the original goal and then asks why the path is unavailable. Missing information calls for a question or search. Missing capability calls for a tool, adapter, or training act. Missing mandate calls for authorization from the right principal. Missing evidence calls for a bounded probe with target-side readback. Refusal is preserved as a decision and ends persuasion. Structural impossibility exposes a trade-off that principals must resolve.
Typed gaps connect each round to a concrete decision. The edge agent ranks questions by expected decision value: which answer is most likely to unlock or change a valuable branch? When the remaining value falls below the cost of another round, questioning stops. The catalyst becomes a translator and information scheduler—not a meeting chair and not a central decision-maker.
| Current gap | System action | New condition created |
|---|---|---|
| Information | Ask, search, or request disclosure | A decision-relevant fact |
| Capability | Build an adapter, tool, or training act | A previously absent action primitive |
| Mandate | Request scoped, time-bounded authority | A legitimate action space |
| Evidence | Run a reversible probe and read the target | A shared observable fact |
| Refusal | Record the decision and optional counter-conditions; stop persuasion | A clear relational boundary |
| Structural conflict | Expose incompatible constraints and alternative branches | A real trade-off principals can choose |
09 · Implementation
The edge agent projects its principal; the catalyst only translates
PFE is designed for frontier agents already working beside their principals. A thin norm teaches a fresh session how to read incomplete material, express SEEK, OFFER, and CONTEXT through self-describing signals, retain a private specification, compile objections into types, and render a plan a principal can judge directly. The protocol reserves yes, edit, or no for the real principal; the frozen corpus does not simulate that step and records NOT_REQUESTED.
The catalyst sees combination structure, typed objections, and unresolved variables. It proposes short rings or chains, translates what one capability means inside another principal’s work, and schedules questions by expected decision value. It does not own the principals’ final judgment and cannot send an inspiration candidate directly to execution.
The runtime retains model family, model, effort, executable revision, session mode, raw inputs and outputs, and continuity receipts. Generated details may vary; instrument identity and the source of every state transition remain traceable.
10 · Experimental design
From unfamiliar open-source combinations to an ablatable full formation arc
The corpus study extracts actors and capabilities from official project material and constructs candidate combinations without prewritten answers. Each combination passes through structurally diverse inspiration, grounding, edge review, a readable plan, and commitment processing. The diversity measure and instrument lock are frozen before generation; interrupted runs resume by hash, preventing attractive outputs from redefining the study.
The full formation arc begins with an ordinary-language air-quality intent and projects official material from OpenAQ, Sensor.Community, and Grafana into three research roles. PFE identifies a missing source-snapshot probe, authorizes it, and creates a deterministic local artifact. Three Codex target-position roles read the snapshot hash relevant to their edge. Removing the probe eliminates the qualified path; restoring it restores the path. In the same study, a live model session proposes an easier one-principal objective, which the goal guard identifies as transformation. A separate source-bounded Agent projection returns NO to an external project write without maintainer mandate and ends persuasion immediately.
The attack study supplies raw ledger pre-state and attempted moves and reruns the verifiers instead of reading a submitted “blocked” report. The norm-pickup study gives a fresh frontier session only the thin norm, then sends incomplete principal material to the same session. The separated evaluator reads only a sealed bundle and recomputes hashes, sources, instruments, nominations, attacks, formation, norm pickup, and reported metrics.
11 · Results
Five evidence lines converge on one recomputable record
The frozen manifest is not a copy of test summaries. It recomputes each metric from retained runtime inputs and binds the evidence files, formula, denominator, and value. The 15/15 result establishes composition of the inspiration and commitment channels on sourced combinations. The 31/31 result comprises 30 attack definitions plus one staged K5 happy path. The 5/5 result establishes recomputable local full-arc conditions. The 35/35 result comes from one fresh session with two retained raw turns. The 11/11 result comes from one deterministic, author-supplied synthetic sealed example read through a separated entry point.
The common result is more important than any single count: the system does not depend on a model declaring “passed.” Every claim returns to state, hashes, sources, attempted moves, target observations, or raw session records. PFE can therefore explain how a candidate was proposed, why a commitment became eligible to advance, how a path appeared, and why a refusal remained a refusal.
| Result | Evidence object | System-level meaning |
|---|---|---|
| 15/15 | Sourced combinations | End-to-end dual-channel composition |
| 31/31 | 30 attacks + one staged K5 happy path | Red transitions do not write; a valid evidence-backed advance still can |
| 5/5 | Local full formation arc | The local causal record and modeled-role boundaries are recomputable |
| 35/35 | One session’s norm-pickup checks | A fresh edge session follows the thin norm on incomplete material |
| 11/11 | One synthetic example’s evaluator checks | A read-only entry point recomputes structural checks from the example |
12 · Discussion
PFE does not decide for people; it enlarges the world in which they can make a real decision
Treating a plan as inspiration means the system does not optimize for an untouched first draft. The product hypothesis is that a candidate creates value when it reveals a new cooperation structure, triggers revision, and changes the next action. The current runs already produce comparable structural candidates; whether real principals act differently after seeing them is the result the next study should measure directly.
The commitment channel gives this openness a clear boundary. A principal may be inspired without being replaced by a central model. A model may conjecture unfamiliar relations without inventing actors or capabilities. Escalation waits for new evidence, execution waits for authority, and formation returns to the original goal and target-side facts.
This changes the endpoint of a discovery engine. Instead of ending with a ranked list, high-recall nominations flow into a generation layer that can ask, translate, probe, read back, and ablate. Discovery brings possibly relevant edges together; PFE produces precision; the edge principal’s final veto remains the last error backstop.
13 · Evidence scope and threats to validity
The present evidence asks whether the machine runs as designed; the next study asks how much net-new value it creates among real principals
Every number in this report comes from an author-built frozen run labeled ENGINEERING_TELEMETRY. It supports implementation completeness, receipt traceability, and mechanical recomputation; it does not estimate inspiration rate, adoption, or durable net value among real principals. The frozen manifest records formal_verdict as null, and the separated evaluator records formal_quality_claim as false. Those fields are stated once here as the concentrated claim boundary.
Official open-source projects serve as checkable capability projections, while the edge roles are principal instruments inside the study. The full arc deliberately confines external effects to a reversible local artifact so that pre-state, post-state, and ablation are fully observable. That design gives the method high visibility and places the next research gate at real mandates, real target-system action, and real-principal adoption.
Frontier models may already know parts of the public project corpus. Official-source preference, retrieval times, and content hashes discipline factual provenance but do not turn model memory into a random sample. Independent evaluation should therefore use sealed unseen material, preregistered diversity and value measures, and target principals who retain ownership of outcome facts and acceptance decisions.
14 · Conclusion
The discovery engine now has a formation core
PFE v3 turns a long-standing conceptual problem into a runnable object: unfamiliar capabilities become comparable candidate structures, gaps route into distinct next moves, commitments advance with evidence, and a local formation record can be recomputed through ablation. The dual channel preserves generative exploration while keeping authority and state transitions explicit.
The most important continuation is not a slogan that “the engine matches people.” It is a different method of work: expand the possibility space, turn each gap into the right action, let real principals reshape and select what inspires them, constrain commitment with evidence, and put result facts at the observation point that owns them.
→
Author-side recomputation
These commands apply to the discovery-engine source checkout held by the authors. They regenerate the corpus, attacks, full arc, norm pickup, separated evaluation, frozen manifest, and verification. The current public download contains an aggregate plus its hash manifest; it is not the raw-receipt reproduction package.
make setup
make corpus
make attacks
make full-arc
make norm-pickup
make eval-ready
make freeze
make verifySPEC.txtSpecification
System objective, mechanism definitions, evidence discipline, and completion conditions.
artifacts/frozen/manifest.jsonFrozen manifest
File-level hashes, recomputation formulas, and the five engineering-telemetry groups.
artifacts/runtime/corpus/Sourced corpus
Candidates, receipts, ledger state, and readable plans for each nominated combination.
artifacts/runtime/full-arc/full-arc.jsonFull formation arc
Original goal, formation events, target readback, refusal, goal protection, and ablation.
artifacts/runtime/attacks.jsonAttacks
Grounding, K1–K5, and formation attacks re-executed from raw state and attempted moves.
docs/EDGE_AGENT_NORM.mdEdge norm
The thin pickup norm for a frontier session joining PFE on behalf of a principal.
evaluation/README.mdSeparated evaluation
Read-only contract and command for an arbitrary sealed input bundle.
Author-side recomputation starts from frozen inputs rather than retained pass fields. Public readers can currently verify aggregate-file integrity and claim consistency.
∎
References
- R1
Towow Research. TOWOW PFE System — Construction Specification v3. Governing system definition and study protocol.
- R2
Pearl, J. Causality: Models, Reasoning and Inference. Cambridge University Press. Background for intervention and causal reasoning.
- R3
Towow Research. 通爻协议本体 · 设计书 v1.2. Design authority for the dual channel, VOI, and crystallization layer.
- R4
Towow Research. 通爻生成哲学 · 方案即启迪. The inspiration channel and its generation philosophy.
- R5
W3C. PROV-O: The PROV Ontology. W3C Recommendation. Background for traceable entities, activities, and attribution.
- R6
Howard, R. A. Information Value Theory. IEEE Transactions on Systems Science and Cybernetics. Background for decision-value question selection.
- R7
Clark, H. H., and Brennan, S. E. Grounding in Communication. In Perspectives on Socially Shared Cognition. Background for interactive grounding and shared evidence.
- R8
Towow Research. PFE frozen artifact manifest. Frozen engineering telemetry and recomputation formulas.
- R9
Towow Research. PFE full-arc runtime artifact. Formation, refusal, transformation, and ablation record.
- R10
Towow Research. PFE executable attack artifact. Grounding, K1–K5, and formation attack results.
- R11
Towow Research. PFE edge-agent norm pickup artifact. Raw evidence for thin-norm pickup.
- R12
Towow Research. Separated evaluator contract and retained result. Read-only evaluation boundary for sealed inputs.
- R13
OpenAQ. OpenAQ API repository and documentation. Official-project source used by the full arc.
- R14
Sensor.Community. Open environmental sensor network. Official-project source used by the full arc.
- R15
Grafana Labs. Grafana documentation. Official-project source used by the full arc.