Whatif

Carry offered taxonomic edges of a superseded model forward to its current successor. edges lists 1 to 25 edge ids that sim_migration_offers(model) offered; pass successor (the current it returned) so a different current successor — a re-point or a chain that grew since — is refused rather than followed. Every edge is checked before anything is written (it must touch model, be offered, pass sim_link's check, you must own model or have authored the edge, and the edge it becomes must pass its predicate's write rule for you — hasMember only for the collection's owner); one failure refuses the call and writes nothing. Each accepted edge is written on the successor attributed to you, with no timestamp, plus a migratedFrom edge from it to the old edge (provenance; sim_edges(<new edge>) shows it, and sim_query closure over migratedFrom traces it back across several supersessions). The old edge stays. Re-running converges on the same ids; an edge you already linked the same way is reused (existing: true, keeping its own timestamp). A call stopped after some edges were written says which were and which were not. Needs sign-in with write scope (opens OAuth the first time); MCP only. Costs 1 compute token per call.

Record a checked connection between two stored models' declared ports, fromModel's fromPort (a place tagged port.output) feeding toModel's toPort (a transition tagged port.input), as a content-addressed Binding (ROADMAP.md Phase 9/11f). Checked now, by the same binds-port-v1 rule sim_run_pipeline applies: both ports exist with the right direction and element kind, the output is a token place, the input is not a delayed transition, the declared kinds are equal, port.unit agrees exactly where both ends declare one (a unit missing on either end is reported unchecked, never agreed), and a numeric transform is at most 1000000. Differing units bind only with a declared conversion, never inferred from the unit names ("order" and "orders" are two units): pass convertFrom, convertTo and convertFactor together (1 convertFrom = convertFactor convertTo), checked by binds-port-v2 — both ports declare port.unit, the two differ, convertFrom and convertTo are exactly those units, the factor is between 1e-06 and 1000000, and the transform times the factor is at most 1000000; sim_run_pipeline multiplies the resampled rate by the transform and then by the factor. A binding without a conversion is checked by binds-port-v1 exactly as before and keeps its id. A delay (hours, in (0, 10000]) is checked by binds-port-v3 (v2 plus check 12: a finite number in range; 0 is no delay and keeps the binding's id): the target receives the source's flow that many hours later and runs on its own declared rate before then, and a loop between models runs only through a binding that declares one. A self-binding is refused, with or without a delay: a model driving its own input is a loop inside one net, built with sim_extend. Set-level checks (a loop with no delayed binding, two bindings into one input, the schedule-segment cap, a loop over 4 models) belong to sim_run_pipeline and sim_create_system; a binding that closes a loop with bindings already stored is not refused here, since a loop exists only in a set someone runs or defines. A refusal stores nothing. It records binding→fromModel "from", binding→toModel "to" and fromModel→toModel "feeds"; storing the same binding twice returns the same id. A recorded binding passes every run's per-binding check; a run still applies the set-level checks to the set it is given and can stop on what only the run decides (a resampled rate that overflows at a very short horizon, the step budget, the deadline). Costs 1 compute token (it stores a Binding).

Calibrate a model against YOUR event log — the reading that meets reality. Upload CSV (case_id, activity, timestamp; the shape sim_dataset emits, activities = transition ids), and rates are learned from the observed timings: sources from inter-arrival times, services from the gap before their completions, all per hour. Instant-pickup transitions (declared rate >= 100) keep their declared rate — their observed gap is the queue wait, and learning it would destroy the calibration discipline. A transition declaring a delay (a fixed-duration timer, not a rate) is fit differently and returned in learnedDelays instead of learnedRates: the MEDIAN observed gap, in hours, written onto the transition itself since a delay has no solver-map slot — a gapCV in rateEvidence far from 0 means the log looks exponential, not fixed, and the calibration says so in a caveat rather than trusting the median anyway. Returns a NEW content-addressed model (learned rates in the solver map, learned delays on the transitions, declared values otherwise untouched, lineage recorded) plus a conformance report: fittingPercent (full replays) is the headline, worst traces named with the activities that could not fire. tokenFitness is a second, harsher reading of the same replay (raw tokens present vs. required at every step, not full-trace success) that under-reads any net with a resource pool — read fittingPercent, not tokenFitness, unless you specifically want the raw-token number. Every learned rate or delay has an entry in rateEvidence: n (gaps it rests on), gapCV (sample std dev over mean of those gaps; ~1 for exponential timings, near 0 for a true timer) and insufficient when n < 2 — n=0 yields nothing, n=1 a value with no spread. Learned values on a structure that cannot replay the traces would be numerology — read fittingPercent before trusting them. Costs 3 tokens from your account's compute bucket (a replay of the log, 2, and the stored model, 1).

Tell whether two differently-labelled models are actually the same net: an isomorphism-invariant id computed from the model's EXACT automorphism orbits (orbits.go), not the colour-refinement (WL) kind sim_classify falls back to when the exact search can't decide. Two models differing only by renaming places or transitions share the same canonicalId even though their content-addressed ids (from sim_get_model) differ — this is the id to compare, not the model id, when checking whether a catalog already holds this net. Also returns the non-trivial automorphism orbits and generator count the id was computed from: zero generators means the net is rigid (no symmetry at all), which is itself a fact about the net's structure. Refuses (as a tool error) when the exact search would spend more than its 1,000,000-step budget (refinement steps, about half a second) — too large or too symmetric for this implementation, per orbits.go — rather than silently falling back to a weaker answer; sim_classify's own fallback covers that case for classification specifically. Costs 1 compute token (one search), as GET /api/models/{id}/canonical does. Equal canonical ids mean the same net under the comparison sim_prove uses for isomorphicTo; to get the place/transition bijection and record it as a witnessed edge, call sim_prove(subject, "isomorphicTo", object).

Re-check a stored structural witness against both stored models with the independent checker (pkg/morphism), which does no search. Given a structural edge's id, checks every witness of that edge; an edge with none answers unwitnessed (unproven, not disproved). Verdicts: valid, invalid (the map is not a morphism of the stored models), or uncheckable (a model was deleted, a net is malformed, or the predicate is now checked under a newer definition). Public: no sign-in. A witness or edge id is free (no compute tokens). With derivation (from sim_query): re-checks every step's witness, the chain, the composed map's digest and the composed map against both end models, and answers {verdict, reason, hops} (the body POST /api/derivation/check returns); a derivation naming a definition other than the predicate's current one is uncheckable; costs 1 compute token. With reach (a sim_reach answer): replays a trace through metamodel.Enabled/Fire and the guards, or re-checks an invariant witness against the stored model, answering {verdict, reason}; from and to are held to the scenario caps (1,000,000 per place, 4,000,000 in total), and a replay is priced before its first firing and is uncheckable over 200,000,000 work units; an exhaustive no is uncheckable (re-run sim_reach); an unknown answer is refused; costs 1 compute token (POST /api/reach/check). With reach on a system's answer, via is re-checked whole with it: the system is stored and holds the net, every other model has exactly one embedding, each embedding's witness or derivation re-checks valid now, and any translation of component places follows them. With reachResult (the result id sim_reach returns): re-checks a stored answer end to end — the record (its bytes hash to its id, its stored question asks what the answer answers under the budgets it records, and over names what it rests on), via as above, and the witness — answering {verdict, reason, parts, answer, currency}; an unknown answer stays unknown (its witness part is uncheckable, never a no); currency says whether this server would compute the same result (definition, budgets, registry, view, every model still stored) and lists superseded models without flipping current; costs 1 compute token (POST /api/reach/check with {"result": …}, or GET /api/lineage/<result>). With refutation (a refuted sim_prove answer's refutation, ROADMAP.md Phase 12i): re-checks its counterexample against both stored models with pkg/morphism — a refines trace replayed through the shared firing rule on both nets, or for preservesInvariant each embedding re-checked as an induced embedding, its invariant re-checked as a P-invariant of the object (y·C = 0 for every transition) and its pullback shown to break at the subject transition named (that the list holds every embedding is the search's claim and is not re-checked) — answering {verdict, reason}; costs 1 compute token (POST /api/refutation/check). A preservesInvariant or refines witness id re-checks its cover or simulation relation the same way (free, bounded: at most 4,096 cover terms, or 4,096 pairs and 20,000,000 firing-rule units). Give exactly one of id, derivation, reach, reachResult or refutation. Writes nothing.

Discover the parameter classes of a stored model and return them as JSON-LD with empty annotation slots for you to fill in (label, comment, unit, domain, substitutes — nothing else; membership/kind/evidence are derived and settled by measurement, not yours to edit). With verify=true a shared colour is checked by exact automorphism proof where the search can decide it (settling interchangeability outright, the stronger claim), falling back to the sampled permutation experiment only where it can't — the exact search refuses past its 1,000,000 refinement-step budget on nets too large or too symmetric for it. Read the sim://docs/classification resource once for the colour-refinement caveat and the annotation contract in full.

Derive a Petri-net model from source code with the configured LLM (control-flow, state-machine, resources or concurrency focus) and store it as a NEW model you own when it passes sim_create_model's decode and the model validator. Stored, it returns stored: true with id, url, name, usage and warnings (what sim_create_model would warn about the model, omitted when there is nothing to say). Otherwise nothing is stored, and it returns stored: false with the LLM's model JSON as model, usage, and either error — one string: the decode failure ("response is not a model: ...") or the unknown-field refusal worded as sim_create_model words it, each misspelt key by path with the key it most likely meant and the structural reasons after them — or errors, the validator's list (each with code, message and fix); fix the JSON and store it with sim_create_model. A reply with no JSON object in it, or a model the store's own validation still refuses, is a tool error naming why. Runs generator.CodeToFlow directly; this tool is its only surface on sim.pflow.xyz, which mounts no code-to-flow HTTP endpoint. Needs sign-in; refused when this deployment has no LLM provider configured.

Run several scenarios against one model on one shared seed and return them side by side — the seed sharing is server-enforced, so differences are the scenarios, not the dice. Returns a summary by default (finals with finalStdDev, each final count's spread across realizations, throughput/mean/P95 metrics, depletion and contention — no time series); pass full=true for the complete trajectories, which run to hundreds of KB. A scenario carrying "summary": true stays summarized even under full=true, so one comparison can chart some scenarios and only read the rest. Unset hours default to 8, samples to 60 (the trajectory grid, which only matters under full=true — metrics are time-weighted and do not depend on it) and realizations to 16 per scenario. Each scenario can set its own "engine" (see sim_scenario / docs/engine-selection.md); comparing an "ode" run against an "ssa" one is legitimate but the shared seed only removes dice from scenarios using the same engine. Costs 1 token per scenario from the compute bucket (per address, or per account when signed in). Each result's contention is split exactly as sim_scenario's is: contended holds only what the run was short of (kind conserved or bounded), waitingForWork the entries of kind queue (the work had not arrived) or state (the model was in another phase of its cycle), neither of them a shortage, and neverFired one note per place that never held a token in any run while its transitions never fired; sim_scenario's description lists what the run cannot tell apart behind such a note.

List the component registry: pre-baked subnet templates (arrivals, service, hazard, inventory, decision, mailbox, datastore) with the calibration discipline baked into the arcs and rates. Each entry names its ports (places you can attach onto existing places), its params with recommended defaults, and the discipline notes explaining WHY the template is shaped the way it is. Compose them with sim_compose.

Instantiate a registry component into a model and store the result as a NEW content-addressed model you own (lineage recorded when composing onto an existing id). Omit id to start a model from the component alone; pass attach to fuse a component port onto one of the model's existing places (e.g. attach {"queue": "tickets_queue"} wires a hazard onto the service's queue). Prefix namespaces the created elements (defaults to the component name). Three calls build a working helpdesk: arrivals, then service attached to its queue, then hazard attached to the same queue — the result passes diagnose because the discipline is in the template.

Check how well a stored model matches an observed event log WITHOUT rewriting its rates — the read sim_calibrate bundles into calibration, offered on its own and in full: fitness (can the model replay each case?), precision (does it allow behaviour never observed?), generalization and simplicity, with per-trace diagnostics naming the activities that could not fire. Also returns a residual block: the same replay read as a trajectory — at which position along the cases (and for which activity) the model first lacks the tokens the log shows, how many tokens short, and a summary that says where fidelity holds and where it breaks. Log is CSV (case_id, activity, timestamp; the shape sim_dataset emits, activities = transition ids). The log is replayed one case at a time from the model's initial marking, so the model should be the per-case workflow; a resource net whose places are shared across cases will not fit. Each case replays in timestamp order; rows with equal timestamps can replay out of row order. Caveats name what the analysable net encoded lossily and where this replay reads the model differently from its firing rule: it does not enforce place capacity, keeps only the heaviest of parallel arcs between one place and one transition, and reads an inhibitor drawn transition -> place as a read arc. Costs 2 tokens from your account's compute bucket.

Online conformance: replay an ORDERED event stream against a stored model and get a running verdict — does the model still match, or which event was the first that did not, and why. Unlike sim_conformance (whole finished cases, aggregate fitness) the order of events IS the stream's order, cases may interleave, and the answer names the first non-conforming event: its position, case, activity, whether the activity is unknown to the model or the transition was not enabled, the places that blocked it (need vs have), the case's marking just before it, and `repair` — the shortest unobserved transitions that would have enabled it (a local alignment of that one step, bounded; absent when none is found within repairDepth). This is token replay, not trace alignment. A divergent event is force-fired so the case resynchronises and later events are judged on their own. fitness is conforming events / events; tokenFitness is 1 - missing/consumed tokens: 1 while nothing has been consumed, and below 0 when more tokens are missing than were consumed. Missing tokens are counted as sim_conformance's replay counts them (an unknown activity and a blocking inhibitor each count one, an unmet read arc its shortfall), except where that replay ignores capacity or merges parallel arcs, which its caveats name. Give the stream as `events` (JSON array of {case, activity, timestamp?}) or `log` (CSV with a case_id, activity header, timestamp optional; rows are taken in the order given, not sorted). Activities are transition ids. Each case replays from the model's initial marking, so the model should be the per-case workflow; guard expressions are not evaluated and timestamps are not checked against delays (caveats says so). keepSteps adds the per-event running signal. Costs 2 tokens from your account's compute bucket. With outlook (JSON {"goal": {place: count}, "avoid": {place: count}, "cases": [case ids]}) the reply adds outlook, separate from conforming and firstDivergence: from the marking each case is in after the replay, whether goal is reached on every way the net can go on and avoid on none. Each target constrains only the places it names, to exactly that count, as sim_reach's to does. Each verdict is guaranteed, forced-violated, open (some continuations satisfy it and some do not) or unknown, which names the budget, cap, guard or timer that stopped the search and is never a yes or a no. A verdict rests on the case's marking itself, a sim_reach answer from that marking (reach; re-check a yes or an invariant no with sim_check_witness(reach=…)), an escape (a deadlock or a cycle that never reaches the target, re-checked before it is returned), or a search of every marking before the target, which has no smaller certificate. No fairness is assumed: a loop that can repeat forever makes a goal open, unless a delayed transition stays enabled along it unfired (the simulator starts a timer the instant it is enabled), which is unknown. Guards decidable from the marking are enforced, though the replay ignores them; a transition whose guard reads action parameters is never fired. Cases are grouped by marking, at most 4 distinct markings per call (outlook.cases, default every case in the stream, narrows it), and each search is bounded by sim_reach's 100,000-state and 20,000,000-unit budgets. The outlook costs 2 more tokens per distinct marking and predicate the marking does not already decide (one for each reachability search it may run, as sim_reach's search costs 1; at most 16), charged before any runs, even when the first search decides it; a refused outlook costs nothing more.

Mint a named Collection you own and return its content id. A collection carries no member list of its own — a mutable list would change the collection's own id every time something joined it, the same reason Lineage lives beside a model rather than inside it. Add members with sim_link(collectionID, "hasMember", memberID) — members must be stored entities — and read them back with sim_neighbors(collectionID, "hasMember") or sim_edges(collectionID). Only you, the owner, may add members: hasMember's write rule is subject-owner (https://sim.pflow.xyz/ns/predicates/v1), so sim_link refuses anyone else, naming the rule. The name and your account together are the content id: creating a collection with a name you already used returns your existing id, and another account's collection with the same name is a different collection. A collection created before collections had owners has none, so its membership is frozen as it stands (still readable); create your own to add to. Members cannot be removed: an edge, once written, is immutable. Needs sign-in with write scope (opens OAuth the first time); costs 1 compute token.

Store a Petri-net model and return its content id and url, plus warnings (omitted when there are none) for a model that stores but probably does not say what was meant: transitions with no rate, which simulations run at the engine default of 1 per hour (each stage of a staged one, in a run without a schedule); a net in which nothing can fire at its declared initial marking; a guard no simulation enforces; an inhibitor arc drawn transition -> place; a pool (staff, machines) one transition takes and gives back in the same firing beside the work it serves, which is never busy, so its headcount multiplies that rate (or, through a "kinetic": false arc, changes nothing) instead of capping work in progress; a pool a transition only reads, through a read arc beside the work it takes, declared above the read arc's weight (staff at 2 read by serve beside queue), which is never taken and stays out of the rate, so every count above the weight runs exactly as the weight does (at the weight, such as an open sign at 1, it is a gate and draws no warning). At most 10 of each per-element warning are listed, the rest counted. The model is JSON with name, places, transitions and arcs; rates are per hour. Read resource sim://docs/authoring once for a worked example and the fields a simulation reads. Refused, with every reason at once: a key the schema does not define, named by its path with the key it most likely meant (unknown field places[0].intial (did you mean "initial"?)) or, when none is close, to be removed, a key repeated across elements written once (arcs[0..6].source); a value of the wrong kind, by its path and the kind wanted (the first one only, and the structure is checked once every value has the right kind); and otherwise every structural defect, leaving out the ones a misspelt key caused. At most 20 of each are listed, the rest counted. Models are immutable; a changed model is a new id. Costs 1 compute token (the account's bucket when signed in, the client address's otherwise); the model JSON is at most 262144 bytes. Signed in with write scope, the model is yours (mine: true): it appears only in your own listing until you dedicate it to the commons with sim_license_model, and you can remove it with sim_delete_model. An identical model stored before this call keeps the envelope it has, so the reply is then mine: false with a note saying whose it is: another account's, the catalog's, or nobody's (created without sign-in). Without sign-in it is stored unowned, for good (mine: false, and a note saying so): reachable by id, listed for nobody, and nobody can delete, license or supersede it — creating the identical model signed in later returns the same id, still unowned. To own a model, sign in before creating it; a changed model (a new name is enough) is a new id. A call whose token has expired or lacks write scope is asked to sign in, not stored unowned. Anyone you give the id to can use the model either way.

Define a System (ROADMAP.md Phase 11f): a content-addressed, stored set of models and the bindings between them, so the catalog can answer questions about them together. Pass collection (snapshot its model and binding members now; other members are listed under skipped, and a grown collection is a new system, the old id still readable) or models and bindings (JSON arrays of ids), not both. Every binding's endpoints join the models (listed under implied); every binding must pass the check sim_bind applies now (binds-port-v1, or binds-port-v2 for one that declares a unit conversion, or binds-port-v3 for one that declares a delay, and the system then records bindingCheck binds-port-v2 or binds-port-v3, v3 first; a legacy binding sim_bind would refuse today is refused here) and the set must close no loop without a delayed binding, take each input port once, fit the schedule-segment cap and hold no loop over the model limit — the checks sim_run_pipeline applies before it knows the hours. At most 64 models and 64 bindings, and a snapshotted collection of at most 256 members. The id is the hash of the sorted lists, the name and the snapshotted collection (if any), so the same call gives the same id; a snapshot and an explicit list of the same ids are different systems. Returns the id and the view sim_system computes (checks, the structural edges among the models with up to 128 witnesses re-checked now, and the shape: one-net, pipeline, separate or unknown). This is not refines: a system's edges are isomorphicTo and subnetOf (embeds by wiring / is the same coloured net). refines (trace inclusion with transitions matched by id) and preservesInvariant (the object's P-invariants pull back along an embedding) are proved one pair at a time by sim_prove and are never walked by closure, since neither is declared transitive. For "what contains this pattern?" ask sim_query with closure: {"select":["?m"],"where":[{"subject":"<pattern id>","predicate":"subnetOf","object":"?m","mode":"closure"}]} (isomorphicTo for copies; add {"subject":"<collection>","predicate":"hasMember","object":"?m"} to scope it), each row carrying a derivation sim_check_witness re-checks; a subnetOf row claims no behaviour inclusion. For "what does this model refine?" ask sim_query for one hop: {"select":["?m"],"where":[{"subject":"<model id>","predicate":"refines","object":"?m"}]}, each row an edge sim_prove recorded. A loop between models runs only through a binding that declares a delay (sim_bind delay; ROADMAP.md Phase 12j): a system may hold one, of at most 4 models, and its view lists it under pipeline.loops with pipeline.maxHours, the longest horizon sim_run_pipeline runs it at (0 when no horizon of 0.01 h or more does; pipeline.runnable covers only the checks that do not depend on the hours), where it is solved as an exact fixed point over the horizon (pipeline-loop-v1). A loop with no delayed binding is refused when a system is defined and by sim_run_pipeline, and sim_bind refuses a self-binding. A port search with among a system (bindable-v3) lists a candidate that closes a loop already holding a delayed binding, with that loop's models under closesLoop, and counts one that would close a loop of undelayed bindings under excluded as closes-loop (sim_bind with a delay can close it); a catalog or collection search does not consult stored bindings, so sim_create_system or sim_run_pipeline is where such a loop is refused. Needs sign-in (opens OAuth the first time); costs 2 compute tokens; writes one System.

Run every applicable READING of a model against the others and report agreement or divergence with the reason: discrete SSA means vs the continuous mean-field solve, algebraically derived conservation laws vs simulated means, and (for game-schema models) the closed-form incidence ranking vs rollouts vs exact search. Divergence is a finding, not an error — small-count mean-field gaps and the prior's threat-blindness are named as such. Trust is agreement between independent readings of one structure. Gated nets (read arc, inhibitor, reached capacity, guard) have no ODE reading to compare against at all — see docs/engine-selection.md for the four-rule decision behind which readings even apply.

Generate a synthetic event log from a stored model (seeded SSA playout; case-per-arrival). Returns CSV. Deterministic: same id, same seed, same bytes.

Delete a model you created signed in. Refused for the curated catalog, for models you do not own (a model created without sign-in is nobody's, for good), and for models already dedicated to the commons (a dedication is irrevocable).

Returns what one more of each resource (and a change in each rate) is worth to the outcome, each ranked against a measured noise floor — no fitness test to write. Also reports generic gates (mass balance, dormant sources, staffing knee and practical knee, whether any knob binds), parameter classes among the controls, and four structural readings needing no run (T-invariants, siphons/traps with deadlock witnesses, CTMC lumpability, constrained lumping); a lumping or siphon reading that does not apply is one "not applicable" line unless verbose. Pure read. Loss/success inference and objective framing can be corrected by tagging places or declaring simulation.objective — read the sim://docs/classification resource once for how to read influence, noise, the knee and structural readings.

Structural difference between two stored models: places, transitions and arcs added or removed, and surviving elements whose numbers changed (initial, capacity, rate, stages, arc weight or kind). The readout for what a builder turn, a sim_extend or a sim_refine actually changed between two ids in a lineage.

Distill exact search into the play scorer: fit rate multipliers for named transition groups so play's rankings agree with exact minimax, on positions sampled by random self-play and labeled by search. This is TACTICAL calibration — the counterpart of sim_calibrate, which learns rates from an event log. The division of labor is deliberate (the ode-minimax experiment): structure carries the tactic, and no fitting of an unmodified net's rates can express what its final state cannot separate — declare the structural prior as transitions in the model (e.g. forced-reply copies of the plays, catalyzed by the opponent's pattern) and distill the magnitudes it introduced. Zero agreement improvement is a finding about the structure, not a failed fit. Read agreementBefore/agreementAfter, not the loss: the hinge loss can overstate failure while every argmax is right. At most 16 groups, each non-empty, and no transition in two groups; the estimated work (loss evaluations × candidate moves × realizations × horizon) must stay under a fixed cap, and an over-cap request is refused with the estimate. Costs 5 tokens from your account's compute bucket at the defaults, scaled by that work estimate (rounded up).

Compare the behaviour of two stored models of one process — typically a net formalized from a document and a net sim_code_to_flow read out of code — and return proved (no word of visible steps separates them), refuted (with separation: the shortest firing sequence one net can make and the other cannot, or after which only one can stop) or unknown. Definition corresponded-traces-v1: correspondence maps subject transitions one-to-one to object transitions (unset: every transition id the two share); each mapped pair is a visible label, and every transition left out is hidden on its own side and closed over, so an internal step the code takes that the document does not name is not a difference. With outcomes (default true) it also compares where each net can stop (a marking where no transition, hidden ones included, is enabled), reading the observed places correspondence.places maps (unset: none, so it compares only whether each can stop); set outcomes=false for nets meant to run forever, such as a server loop. Untimed and interleaving through the shared firing rule: rates, delays and schedules are not compared; a guard, a precondition the net's source carried and it could not represent (guardUnrepresentable, as sim_code_to_flow can leave), a fixed delay, or two consuming arcs from one place into one transition make the answer unknown (nothing is compared). Budgets: 20,000,000 work units, 4,096 joint states, 4,096 markings per side after one word, 262,144 markings held across every joint state kept, 4,096 steps in a run, 1,000,000 tokens per place; the whole call is also bounded by the 55 s request limit. A budget that runs out answers unknown, with agreesThroughLength once the empty word was compared (no word of at most that many labels separates them), and never counts as agreement; retrying gives the same answer. A refuted answer's separation was re-checked by pkg/morphism before it was returned; pass it back as separation, with nothing else, to re-check it against both stored models now, with no search (valid, invalid, or uncheckable: a model was deleted, a net carries a guard, a lost precondition or a fixed delay, the separation names another definition, or the check is over the checker's bounds). A proved answer carries no certificate. Every answer carries means: a refutation shows the two NETS differ; whether the document and the code differ, or a formalization or the correspondence is wrong, the engine does not decide; a proof is agreement between two readings, not fidelity. Unlike sim_prove's refines (one direction, by id, no hidden steps, recorded), it is symmetric and records nothing; unlike sim_diff (the text of two models), it compares behaviour. Needs sign-in (opens OAuth the first time). Costs 2 compute tokens to compare, 1 to re-check a separation. Writes nothing.

List every relation touching an entity, as either subject or object. Answered from an index by subject and object; each call re-lists the store so edges written by other instances are seen. Legacy edges whose predicate predates the registry are included with axis "unregistered". Structural edges (isomorphicTo, subnetOf, preservesInvariant, refines) list the witnesses that prove them; one listing none is marked unwitnessed and proves nothing. An isomorphicTo edge is the same net up to renaming, not the same behaviour: its definition (coloured-net-v1) does not compare guards, stages, schedules, constraints, objectives or non-refine tags such as outcome — read the witness (GET /api/lineage/{witnessId}, whose notCompared lists what that pair carries; sim_check_witness re-checks it) before treating two models as behaving alike.

Score a player's legal next moves with the PLAY method (the blog post's move picker): apply each candidate hypothetically and score the expected objective of the position it leads to, by seeded SSA rollouts (seed 11, shared across candidates so a difference measures the move, not the dice) — the best move scores highest. Needs the game schema (simulation.objective + simulation.players). The response names the engine. This tool always uses play on SSA: next-move elimination (rate-zero ablation, which reads an ungated net through the continuous ODE relaxation), the closed-form incidence reduction and exact search are the HTTP evaluate endpoint's method option (eliminate | incidence | search), not selectable here.

Apply structural edits to a stored model and store the result as a NEW model you own, with lineage back to the original — the same vocabulary the guided builder uses behind its interview, now callable directly. Operations (JSON array, each with "op"): add_place {id, initial}, add_transition {id, guard, event}, add_arc {from, to, weight, kinetic, type}, remove_place, remove_transition, remove_arc {from, to}, set_rate {id, rate}, set_initial {id, initial}, set_capacity {id, capacity}. The edited model is validated before it is stored; a set of operations that leaves the net malformed is refused with every reason, and nothing is written. Returns the new id, the operations applied, and the structural diff.

Write a Petri-net model from a plain-language description with the configured LLM, with the levers that moved the pass rate in the benchmark behind it, each optional. The model is checked as it arrives (decode, validator, then the builder's behavioural checks: disconnected parts, transitions that never fire, places nothing fills, and any place_ids you named); what is wrong is shown back to it and it answers again, up to `retries` times. A net that decodes and passes the validator is stored as a NEW model you own, exactly as sim_code_to_flow stores one, even when behavioural problems remain; `clean` says whether any do and `problems` lists them. `notes` lists what is right for some nets and wrong for others, and is never fed back: a state where nothing can fire, a place that grows without limit. A net that cannot be stored is returned as `model` with the reasons. Levers: retries (0-4, default 2) re-asks with the findings and is the strongest; best_of (1-5, default 1) starts again from scratch when a sample ends with problems and stops at the first clean one; selfcheck asks the model to play its own net forward and replays that against the net (it catches inconsistency, not a wrong idea); example is "real" (default, a worked order workflow) or "placeholder" (the older schema with placeholder names, which small models copy literally); model picks one of the deployment's allowed models; thinking caps the model's hidden reasoning tokens (0 turns it off); temperature is 0-1. Cost: every model request is one call and one request makes at most 6 (best_of x (1 + retries), plus one for selfcheck); a request over that is refused before it runs. The calls it really makes are counted against your daily LLM allowance, the same one the builder uses, and the response reports calls, tokens (thinking tokens included) and the allowance left. Needs sign-in; refused when this deployment has no LLM provider configured.

Fetch a stored prompt or artifact by id as JSON-LD: the two records sim_prompt and sim_reroll store besides a model. An artifact (@type Artifact) is an LLM reply that was not stored as a model — a report, code, or model JSON that failed validation — returned whole as content (the store caps an artifact at 262144 bytes), with contentKind. A prompt (@type Prompt) is what was asked: text, system, llmProvider, llmModel and requestTemplate (the version of the fixed instruction and settings sim_prompt wrapped them in). A prompt with no requestTemplate was stored before sim_prompt's description said a stored prompt and its results are readable by anyone with an id, so its text and system are withheld, and so is the content of an artifact it produced: withheld names the fields left out, and note says why. parent is the id it was run against (a prompt) or produced by (an artifact), and lineage is the derivation record for it (wasDerivedFrom, activity, ancestry nearest first, wasAttributedTo, generatedAtTime), absent when none was recorded. A model id, or an id of another stored kind, is refused naming what it is and the tool or route that reads it. Public: no sign-in and no compute tokens; writes nothing. GET /api/artifacts/{id} serves the same document.

Fetch a stored Binding by id, with its check computed now (binds-port-v1; binds-port-v2 for a binding that declares a unit conversion, which is returned as convert; binds-port-v3 for one that declares a delay, returned as delay in hours): ok (with the unit verdict: agree, converted or unchecked), refused (a binding recorded before the check that sim_bind and sim_run_pipeline would refuse today, flagged legacy) or uncheckable (a model it names was deleted).

Fetch a stored model's full Petri-net JSON by its content id. A name in place of an id is refused saying to search with sim_list_models' query, and that a model created without sign-in is listed for nobody, so its id is the one its create returned; or naming the curated model's id when a curated model has exactly that name. An id of another stored kind (a prompt or artifact, a sim_reach question or answer, a system, binding, map, collection, witness, relation, or sim_query query or result) is refused naming what it is and the tool or route that reads it. Public and free.

Derive a model's full algebraic invariant structure: conservation laws (Farkas P-invariants — weighted place sums every run preserves, the arithmetic a trust panel should show), firing cycles (T-invariants, named per-cycle with a readable detail sentence, each tagged StructuralProof), and the siphon/trap report (every minimal siphon and trap found from the arc structure, plus deadlock witnesses — minimal siphons holding no tokens at this model's own initial marking, which proves every transition needing one permanently disabled). This is the same computation sim_diagnose's structural fields read from, not a lesser copy of it. Pure structure, no simulation; every claim holds for every trajectory from this initial marking.

Dedicate a model you created to the commons under CC0-1.0, CC-BY-4.0, CC-BY-SA-4.0. It then appears in every user's listing with the license shown, and the dedication is IRREVOCABLE — it cannot be changed or deleted afterwards, which is what makes it safe for others to build on. CC0-1.0 is the cleanest choice for a model: attribution terms are hard to honor for a net someone folds into a larger one.

Record a typed edge between two stored entities (models, prompts, artifacts, maps, collections, bindings, relations — not witnesses), using a registered predicate. sim_link writes these (subject → object): @type (model → a https://sim.pflow.xyz/ns/models/v1 type name); about (artifact/prompt/map/collection → model/collection/binding/system); broader (model/collection/system → model/collection/system); cites (any entity → any entity); closeMatch (model → model); hasMember (collection → any entity); related (model/collection/system → model/collection/system). from, to and feeds are recorded only by sim_bind, produced only by sim_prompt/sim_reroll, supersededBy only by sim_supersede_model, migratedFrom only by sim_accept_migration; narrower is broader read backwards and is never stored; structural predicates (isomorphicTo, subnetOf, preservesInvariant, refines) are recorded only by sim_prove, together with the witness that proves them. Each is refused here, naming the tool that records it (or the reversed call, for narrower). Subject and object must be stored entities of the kinds the predicate relates, passed by content id (for @type, the object is a /ns/models/v1 type name). Full definitions at https://sim.pflow.xyz/ns/predicates/v1. Distinct from the Lineage a model/prompt/artifact already carries, which is specifically derivation (parent -> prompt -> child). Who may write an edge is its predicate's write rule (writeRule in the registry): hasMember is subject-owner — only the collection's owner (who created it with sim_create_collection) adds members, and a collection created before collections had owners takes no new ones; every other predicate sim_link writes is any-signed-in, an assertion attributed to you. A refusal names the rule. Edges are immutable and are never retracted. Linking the same subject/predicate/object again as the same user — either direction for a symmetric predicate — is idempotent: it returns the existing relation's id (existing: true) rather than recording a duplicate edge. Needs sign-in (opens OAuth the first time); the edge is attributed to you; costs 1 compute token. Writes one Relation, or nothing when it already exists.

List the content id of every stored Binding.

List the models visible to you, as a JSON array of {id, name, description, license, curated, mine} sorted by name (a field that does not apply is left out): the curated catalog (curated: true), models dedicated to the commons (each carries its license), and your own (mine: true). The curated models are complete, working examples of the model JSON: read one with sim_get_model and copy its shape when writing a model for sim_create_model. scope narrows the list to one of those groups and query to the models matching every term; the two compose. With neither, every visible model is still listed, but compactly: each description except a curated model's is cut to its first 64 bytes, ending in "…" where cut; pass scope or query (scope "commons", or query "clinic") for each listed description in full (its first line, at most 200 bytes), which is also the text query searches. Free (no compute token) and writes nothing. Public, except that scope "mine" needs sign-in (opens OAuth the first time). Other users' undedicated models are not listed, but any model id works with every sim_* tool — an id someone shares with you is the model.

Fetch a stored Map's key->value data by id.

List the content id of every stored Map.

Store a key->value lookup table as its own content-addressed entity — a generated parameter sweep, a rate table, a component registry, anything shaped as key->value rather than free text (an artifact) or a Petri net (a model). Returns its content id; the same data, even with keys inserted in a different order, returns the same id.

List what happens to the edges of a superseded model. Taxonomic edges (hasMember, @type, broader, related, closeMatch, cites, about) are OFFERED: each row gives the edge as it would read on the model's current successor (followed through its supersession chain), whether sim_link's checks pass for it now (passes, refused, or unchecked when the store could not answer), any earlier migrations of it, and the sim_accept_migration call that accepts it. Nothing moves automatically, the old edge stays, and this read writes nothing: offers are computed on read and never stored. Tool-written edges (from/to/feeds, produced, supersededBy), structural edges (isomorphicTo, subnetOf, preservesInvariant, refines) and legacy unregistered predicates STAY with the old model, each with its reason; a structural row names its witnesses and the sim_prove call that would make a new edge for the successor. Stored systems naming the model are listed under systems (sim_system {naming}'s answer, at most 100 rows): a System names exact ids, so supersession changes none of them, and systemsNote names the sim_create_system call with the successor. A chain that cycles, runs past 64 hops, ends at a deleted model or reaches a model now owned by someone else is reported as such, with no target. Accepting needs you to own the superseded model or to have authored the edge, and the edge it becomes must pass its predicate's write rule for you, as sim_link would (hasMember: you own the collection); youMayAccept says whether you may, per edge, and writeRule says why not when only the write rule refuses. 100 edges per page (pass next as after); predicate narrows to one stored predicate (narrower is refused: pass broader). Needs sign-in with read scope (opens OAuth the first time); anonymous over HTTP as GET /api/models/{id}/migrations, with no youMayAccept. Costs 1 compute token per call.

List the sheets this user has published, with their URLs.

One-hop traversal from an entity, read through the predicate registry. With no predicate: the object of every relation where the entity is the subject. With a predicate: the entities one hop along it — for an inverse name that is never stored (narrower) the stored predicate is read backwards (the subjects of the broader edges pointing at the entity), and for a symmetric one (related, closeMatch) both directions count, since an edge recorded from the other end says the same thing. A legacy free-form predicate still answers from its stored edges; a name that is neither registered nor used by any stored edge is refused. Pass a collection's id with predicate hasMember to list its members. A structural neighbour (isomorphicTo, subnetOf, preservesInvariant, refines) is a bare id here: sim_edges lists the edge's witnesses, and isomorphicTo does not compare guards, objectives or outcome tags. Ordered by timestamp, then relation id: tool-recorded edges (from/to/feeds, supersededBy, and edges carried forward by sim_accept_migration, unless it reused one you had already linked) carry no timestamp, come first and are in no particular order among themselves, so the last supersededBy neighbor is not necessarily the current successor. Not de-duplicated.

Multi-objective optimisation over transition rates for a stored model: Monte Carlo samples the rate ranges, runs each combination to the horizon with the continuous engine, and returns every sample with a Pareto flag — the non-dominated set is the trade-off frontier ('which staffing is non-dominated on served vs walked out'). Continuous reading: a model with a schedule or a gate is refused with the reason (use sim_compare with explicit scenarios for those). Costs 1 token per 50 samples (rounded up) from your account's compute bucket; rate ranges must be non-negative.

Two-rate grid for a stored model: vary two transition rates over ranges, run each combination to the horizon with the continuous engine, and return the observable's final value as a grid — 'which regime of arrivals × restock keeps the queue empty'. Continuous reading: a model with a schedule or a gate is refused with the reason. Costs 1 token per 50 grid cells (rounded up; a 40x40 grid empties the bucket) from your account's compute bucket.

Ask an LLM to derive something from a stored entity: a variant model, a report, a piece of generated code — whatever the prompt asks for. The parent's JSON rides along as context after a fixed instruction: answer a request for a new or changed model with only the complete model JSON (the shape sim_create_model takes), and anything else in plain text or markdown. The parent is looked up as a model first, then a prompt, then an artifact, then a map — whichever resolves — and the context block is labelled by what kind it found ("## Parent model", "## Parent prompt", ...), so the LLM is never told a report is a Petri net. When the parent is a prompt with no requestTemplate, its text and system are withheld from its readers (sim_get_artifact says why), and so is the content of an artifact such a prompt produced; unless you are its author, the LLM gets the document sim_get_artifact serves in its place, without them. The prompt is stored first and content-addressed like a model, so it has an id of its own before the LLM ever answers; both the prompt and whatever came back are placed in lineage under the parent (sim_prompt as the activity), so Ancestry walks parent -> prompt -> result. A Relation{prompt, "produced", result} is recorded alongside — sim_reroll's forward index, and queryable directly via sim_edges/sim_neighbors. Returns promptId, id, kind and url. A reply that decodes and passes the validation sim_create_model applies is stored as a NEW model you own (kind "model"; read it with sim_get_model), with warnings, when there are any, saying what sim_create_model would warn about it (a transition left at the default rate, a guard no run enforces, staff taken and given back in one firing, ...). Any other reply is stored as an artifact (kind "artifact") and returned inline as content, cut at 16384 bytes with truncated: true and contentBytes (the whole reply's length) when longer; sim_get_artifact reads all of it. When that reply held model JSON (an object with places or transitions) that failed to decode or validate, modelErrors lists every reason at once, as sim_create_model words its refusal (a key repeated across elements once, at most 20 of each kind with the rest counted). Needs sign-in (opens OAuth the first time). Refused if this deployment has no LLM provider configured. The stored prompt, text and system both, is readable without sign-in through sim_get_artifact by anyone who has its promptId or the id of anything derived from it, since lineage names the prompt behind every result: leave out of text and system what you would not share with whoever gets a result's id.

Propose candidate @type values for one or more stored models, e.g. "QueueingSystem" or "ResourcePool", from a small Diagnose run this tool performs on each model (nothing is cached or reused between calls). Every rule is a hand-written assumption about what a shape of knobs/loss/siphons/classes tends to mean, not a structural proof or a measurement, so results are ASSUMPTION-grade until a human reviews one and applies it — apply with sim_link(id, "@type", "<Type>"), there is no separate apply tool. Pure read; nothing here is written to any model. Defaults to scanning the visible catalog (up to limit) when ids is omitted. Costs one Diagnose run per model — 5 tokens each from the compute bucket (per address, or per account when signed in), charged as each model is reached — so limit and realizations are both capped.

Prove a structural relation between two stored models and record it with its witness. isomorphicTo: an exact canonical-labelling search (the one behind sim_canonical, seeded with exactly what the check compares) finds a place/transition bijection preserving arcs (direction, type, weight, kinetic), place sort, initial marking, capacity, rate or delay and refine.* tags. It does not compare guards, stages, schedules, constraints, non-refine tags (outcome, port.*), data-place types and initial values, objectives, players or asserted classes; those the pair carries are listed as notCompared, so proved is not 'behaves the same'. subnetOf: a bounded embedding search finds an injective, sort-preserving map (token places to token places, data places to data places, transitions to transitions) of the subject's places and transitions into the object's whose arcs among the image are exactly the subject's (initial markings, capacities, rates, delays and tags are not compared). preservesInvariant (invariant-pullback-v1): an induced embedding as subnetOf's along which every P-invariant of the object pulls back to a P-invariant of the subject (the incidence the shared firing rule implies: consuming and producing arcs on token places). Proved carries the map and a cover — per subject transition, rational coefficients of object transitions with the same incidence — which proves it for every invariant at once, plus invariantDimension (0 means vacuous: the object conserves nothing). Embeddings are tried one image at a time (embeddings onto one image differ by a symmetry of the subject, which keeps or breaks the pullback with them), at most 16 images. refines (id-simulation-v1): every firing sequence of the subject from its initial marking, transitions the object has no transition of (by id) dropped as silent, is one of the object from its own; untimed, guards and fixed delays refused (unknown: a delay is a timer whose in-flight markings an atomic step skips), rates not compared, at least one shared transition id required. Proved carries a simulation relation (pairs of markings closed under the subject's firings, at most 4,096; relationPairs counts them, GET /api/lineage/<witness> lists them), which makes the claim unbounded; neverFired lists the shared transitions the subject never fires from its initial marking, and vacuous: true says it never fires any of them, so the proof (still recorded) has no visible step to match. Neither is declared transitive, so sim_query never closes over them. The answer is proved (with the map, in the caller's direction, or the relation), refuted (exhaustive, with the reason; a refuted refines, or a refuted preservesInvariant where the subject embeds at all, carries refutation, a counterexample sim_check_witness(refutation=…) re-checks: a trace for refines, an invariant per embedding for preservesInvariant, whose list of embeddings is the search's claim; a preservesInvariant refuted because no embedding exists carries none, as a refuted subnetOf carries none) or unknown (a search budget ran out; nothing is recorded, retrying gives the same answer, and unknown is not a no). Needs sign-in (opens OAuth the first time), including with record=false. Costs 3 compute tokens from your account's bucket, or 2 with record=false. Budgets: 1,000,000 refinement steps for both nets' canonical labelling together (isomorphicTo), 2,000,000 embedding steps (subnetOf), 4,000,000 units for elimination, embedding and tests together (preservesInvariant), 20,000,000 firing-rule units and 4,096 pairs (refines); the whole call is also bounded by the 55 s request limit. Writes, only when proved and record is true (the default): one witness, then one relation; isomorphicTo is stored with the smaller id as subject, and storedAs says so when that turned the call round. Otherwise it writes nothing. sim_check_witness re-checks either at any time. With via, sim_prove promotes instead of searching: it takes a structural fact sim_query derived (two to eight stored witnesses composed), re-derives it here from the stored witnesses — every step's witness re-checked against both stored models now, the chain, the composed map's digest, and the composed map checked against the two end models by pkg/morphism — and records the composed map as one direct witness through the same write gate as a search's, so the edge has the same id (one per subject, predicate and object) and is trusted exactly as far. The witness records the derivation as promotedFrom; re-promoting the same derivation writes nothing new (existing: true). A derivation that re-checks invalid or uncheckable is refused and nothing is written; an unknown is never promoted. Costs 2 compute tokens (a derivation re-check plus a store) instead of 3; record=false is refused with via.

Publish a stored model into the signed-in user's Google Sheets: the model workbook (live formulas when honest, a refusal tab when not), a server-run scenario as data tabs, and trajectory + contention charts. Returns the sheet URL. Counts against the daily quota, and costs 1 token from your account's compute bucket for the run.

Publish the generated application for a model you own — the single-file HTML a generator produced from the model's `view` prompt. Served at /app/<id> in a sandboxed opaque origin (no cookies, no session; only the CORS-open public API is reachable). START FROM THE RUNTIME, not from scratch: /lib/app-template.html is a working console that imports /lib/sim-console.js and composes <sim-controls>, <sim-disruptions>, <sim-net>, <sim-timeline>, <sim-trajectory> and <sim-results> — the same components the generic console at /whatif/ runs. Composing them is how an app inherits role derivation, the fungible-set collapse, the influence ranking that never filters, the contention ledger and the verbatim caveats, none of which the checks below can verify you reimplemented correctly. Root-relative /lib/ imports are allowed. Off-origin loading is stopped in two places: the upload checks refuse an off-origin <script src> or <link href>, and the app is served under a Content-Security-Policy that confines scripts, styles, fetches, images and fonts to this origin — which is what stops the forms the checks do not parse (an inline module's import "https://…", a dynamic import(), a CSS @import). Checks refuse an app that is empty, oversized, never references its model id, or carries an off-origin <script src>/<link href>; behavioral correctness (does the app actually do what the view says) is on the generator and any browser gate you run.

Publish a multi-scenario comparison into the signed-in user's Google Sheets — sim_compare's export, the counterpart of sim_publish for a single scenario. Runs every scenario on one shared seed (the same server-enforced sharing sim_compare uses, so differences are the scenarios and not the dice) and writes a comparison table plus a trajectory chart, rather than one scenario's own data tabs. Returns the sheet URL. Counts against the same daily publish quota as sim_publish, and costs 1 token per scenario from your account's compute bucket.

Answer a question over the edge graph between stored models: bounded paths, closure over transitive predicates, and conjunctive patterns of up to 4 triples joined on shared ?variables. The query argument is JSON: {"select": ["?m"], "where": [{"subject": "?m", "predicate": "subnetOf", "object": "<content id>", "mode": "closure"}, {"subject": "?m", "predicate": "@type", "object": "QueueingSystem"}]}; a term starting with ? is a variable, anything else a constant (a 24-hex content id, or for @type's object a model kind). Modes: edge (default; one stored edge, read through the registry: narrower is read as broader turned round, symmetric predicates match either way round), path (1..maxHops hops, maxHops 1 to 8 required; a row says reached by a path unless the predicate is transitive), closure (no maxHops; allowed only on the transitive predicates broader, isomorphicTo, migratedFrom, narrower, subnetOf, supersededBy). Put the bound term in object to walk backwards. Unregistered (legacy) predicates are one stored edge in their stored direction only, never walked or closed over. A structural edge (isomorphicTo, subnetOf, preservesInvariant, refines) counts only when one of its stored witnesses re-checks valid now; a derived structural fact carries a derivation (the chain of witnesses and the composed map's digest) that sim_check_witness(derivation=…) re-checks, and is never stored; excluded lists the edges not walked and why. Closure runs per predicate (isomorphicTo hops never feed a subnetOf closure). status complete means complete over the eligible edges (and, for path, within maxHops: hopLimitReached names patterns with edges beyond); truncated means a bound was reached, rows is absent and the rows found so far are in partialRows (true, but not all of them). A query with no variables answers yes, no or unknown (a bound stopped it, or a stored row could not be decoded); a no's caveat says how far the search looked (only closure rules out a derived edge, and a path cut at maxHops says so), and no is not a refutation. Pattern numbers in a result (support[].pattern, hopLimitReached, unmatched, truncated.pattern) are 0-based positions in query.normalized.where; refusals number your patterns from 1. Limits: 4 patterns, 8 hops, 200 rows, 10,000 intermediate rows, 1,000,000 work units, 64 models probed, a 256 KiB result, a 16 KiB query; the whole call is bounded by the 55s request limit. Over a limit before running is a refusal naming it; reached while running is truncated, and retrying gives the same answer. Needs sign-in with read scope (opens OAuth the first time). Costs 2 compute tokens per call, cache hits and refusals included. Writes one Query and one Result entity, content-addressed (resolve either at GET /api/lineage/{id}; stored is false, with storedNote saying why, if the write failed or a composed map was refused); a repeated question is served from the cache only while the edges and witnesses it read, and the models it checked, are unchanged. An answer naming models that have since been superseded carries superseded ([{model, successor, current, chainStatus}]); the answer is still right about those versions.

Returns whether the marking to is reachable from the marking from in ONE stored net (a model, or a system that is one net): yes with a firing trace, no with a re-checkable P-invariant (or after an exhaustive search), or unknown with every reason. A sim_compose result is one fused net, so it qualifies; a sim_bind binding or a collection links rates, not tokens, and answers unknown (not-one-net). A system id (sim_create_system) is answered on its one net when one member has every other embedded in it by a witnessed subnetOf or isomorphicTo edge or a chain of edges of one predicate (the answer carries via: an edge and witness per member, or the derivation sim_check_witness(derivation=…) re-checks); otherwise it answers unknown: not-one-net when it holds bindings (they link rates, not tokens) or no member qualifies, system-truncated when its witness bound was reached or the chain search stopped at a bound. On a one-net system, a key in from or to may name a member's place as <model>:<place> (the member's id, or its name when no other member shares it): it is translated into the net place that member's embedding maps it to, the answer's from and to are the translated question, and via.translation lists each key's member, place and net place, which sim_check_witness(reach=…) re-checks against the stored witnesses. Refused, naming the key: a member or place the system does not have, a data place, a key that reads more than one way, a place two valid witnesses of the member's edge map differently, and two keys landing on one net place with different counts; more than 16 other witnesses to compare answers unknown (translation-budget). from is a sparse override on the model's declared initial marking (unset = declared). to constrains only the places it names, each to exactly that count; naming every token place asks for an exact marking. Unknown or data places, negative or fractional counts and counts over the scenario caps (1,000,000 per place, 4,000,000 in total) are refused. Reachability here is untimed: each firing is one atomic step under the shared firing rule (read and inhibitor arcs, capacities, guards decidable from the marking); rates, schedules, delays and stages play no part. Answers: no with a P-invariant y whose weighted sum differs between from and to (when to leaves a place of y free, to alone must already exceed it), valid whatever read arcs, inhibitors, capacities or guards allow, since those only remove firings; no (exhaustive) when every marking reachable from from was searched; yes with a firing trace (shortest among the firings the search makes; reads says when it skipped some); unknown (unknownBecause lists each cause) when neither is found within 100,000 markings or 20,000,000 work units, when a firing would pass 1,000,000 tokens on a place, when a guard fails to evaluate at a reached marking, when a transition whose guard reads action parameters (or does not parse, or whose precondition was lost) was enabled and so left out, when the invariant algebra runs past its 20,000,000-operation budget, is cut short or needs coefficients over 2147483647, or when a found trace would cost the independent checker more than 200,000,000 work units to replay. Unknown is not a no. Every answer (unknown included) is stored: question is the content id of the question as asked (the id, from and to as written, the definition and the budgets), result the id of the answer with what it was computed over (the net, every model, edge and witness via rests on, the predicate registry and system view for a system); a repeated question is served from a per-instance cache with the same bytes, and storedNote says why an answer too large to store (over 262,144 bytes) was not. Public: no sign-in; costs 2 compute tokens (a search plus a store). Re-check a yes or an invariant no with sim_check_witness(reach=<this answer>), which also re-checks via whole; re-check a stored answer end to end with sim_check_witness(reachResult=<result>) or GET /api/lineage/<result>.

Run a seeded scenario and get back the result PLUS a signed run receipt: an Ed25519 certificate over (model id, scenario, result hash, service revision) and the inputs the engine recorded for this run: inputs.modelDigest (the model after params), method, hours, samples, realizations and seed as run (with the scenario argument's defaults filled in, and samples at the engine's 60 when unset or below 2), marking (the scenario's overrides), rates (each constant rate the engine ran at; a delayed transition, and one running a schedule, has none), schedule, solver (entry point and options) and engine (go-pflow version from build info, "unknown" when the build carries none, and Go version). evidenceLevel is bound at issue; only POST /api/receipts/verify can report replayed, which it computes from its own re-run, and a scenario-run receipt never reaches proved. A receipt over the 1,048,576-byte body /api/receipts/verify reads (compact JSON) is refused before the run rather than issued (the inputs repeat each override, so a scenario overriding every rate of a very wide model can reach it), and so is engine: "ode": the continuous engine is not bit-reproducible, so its result hash could not be checked; take the receipt on the discrete engine. Anyone can check it two ways — verify the signature offline against the embedded public key (proves this service reported this result), and POST it to /api/receipts/verify (no auth) to replay the run and confirm the result hash reproduces (proves the run is reproducible, not invented); when what the engine reads now differs from the recorded inputs it names the changed keys in inputsChanged, with the new record in inputsNow. The current signing key is at GET /api/receipts/key. Reproducibility is the bottom rung of the trust ladder receipts build: play the model, check the anchors, re-run the seed, verify the certificate. The result is the engine's own form, the one POST /api/models/{id}/scenario returns: contended is the unsplit list with each entry's kind (only conserved and bounded are shortages; queue means the work had not arrived, and state that the model was in another phase of its cycle), not sim_scenario's split into contended, waitingForWork and neverFired, because the result hash covers the result exactly as returned. Costs 1 token from your account's compute bucket.

Refine a model's parameter classes by editing what the model SAYS (tags on a place or transition, or assertedClasses), then re-derive. Returns a NEW model id (ids are content addresses, so the original stays reachable) plus a before/after class diff. tags can only split classes; assertedClasses declares a merge and gets re-verified and costed, never trusted blind. Read the sim://docs/classification resource once for why the two levers are not symmetric. Costs 2 tokens from your account's compute bucket (the stored model, 1, and the verified re-classification, 1).

Re-run a stored sim_prompt against the SAME parent it originally ran against — a sibling attempt, never a chain: it never derives from the previous attempt's output, only from the original parent, so rerolling ten times leaves ten independent siblings in lineage rather than a chain of ten. Reuses the original prompt's text and system unless you override them here; a prompt with no requestTemplate (stored before sim_prompt said a stored prompt is readable by anyone with its id) has them withheld from every read, so only its author (its lineage's wasAttributedTo) may reroll it reusing them, and anyone else passes text, and system when it has one, of their own. The parent rides along as sim_prompt sends it, withheld fields left out unless you are its author. The original prompt and its result are left untouched; this stores a new prompt and a new result (model or artifact, same rule as sim_prompt) under Activity sim_reroll, and returns the same fields sim_prompt does: a model's warnings, an artifact's content inline, and modelErrors when the reply was model JSON that failed. When called with neither override and the deployment's LLM provider and model are unchanged since the original ran, the response carries a reproducibility field. If the request template (the instruction and settings sim_prompt sends; a prompt's requestTemplate) is unchanged too, it is checked against every prior result this exact prompt has ever produced (via the same forward "produced" relation sim_edges/sim_neighbors can query directly): "verified" if this result content-matches one of them, "diverged" if it doesn't, "not verified" if there's no prior result on record yet. If the template changed since the original ran, it is "not comparable", and the run is stored under a new promptId. Needs sign-in (opens OAuth the first time). Refused if this deployment has no LLM provider configured. The new prompt is readable as sim_prompt's is: without sign-in, by anyone with its id or a result's.

Returns one scenario result per model that a set of stored Bindings (see sim_bind) touches, run as a composed pipeline (each result's contention split into contended, waitingForWork and neverFired exactly as sim_scenario's is), with the assumptions the composition makes, inTransit (per delayed binding, the source's flow its delay carried past the horizon) and, when the set holds a loop, the loop's reading (loop, pipeline-loop-v1: verdict, claim, components with sweepsNeeded and sweepsRun, sweepsRun, residual). Costs 1 compute token per model run, charged up front, before any model is read, for the most runs the set can take — one per model outside a loop, and a loop's models times the most sweeps it can take — and not refunded when a loop settles in fewer. Defaults: hours 8, seed 1, realizations 0 (one SSA trajectory per model; set it, e.g. 16 as sim_scenario does, since bindings and loops couple means). Caps: hours up to 10000, realizations up to 200, as sim_scenario. Each model runs through its own ordinary scenario in topological order, an output port's trajectory resampled into the target's input-transition schedule, one seed and horizon shared by every model (the discipline sim_compare enforces within one model). A declared conversion is applied exactly as declared: the resampled rate is multiplied by any numeric transform, then by the convert factor. A binding with a delay drives its target that many hours later; before then the target runs on its own declared rate. A loop between models runs only when some binding on it declares a delay. It is solved by sweeps, each running every member once with one delayed binding (cut) driven by the previous sweep, until a sweep reproduces the previous one's schedules bit for bit: verdict fixed-point, whose claim (also assumptions' first line) is that this is the unique fixed point of the coupling over 0 to hours at the seed, followed by its four limits — each model was driven by the other's mean across realizations (mean-field: for joint dynamics build one net with sim_extend), each delay is declared, not measured, before its delay a target ran on its declared rate, and nothing is claimed past the horizon (a fixed point can be a loop that grows; read the series). Verdict stopped means a run hit the step budget or token cap (from sweep 2 on, typically a loop whose gain is above 1, which grows every sweep; in sweep 1, before any feedback has gone round the loop, the model itself at these hours and realizations) and not-settled is a defect to report; neither carries steps, and models after the loop are not run. Refused before anything is charged or read, each refusal naming the value and the bound: more than 64 binding ids; realizations or hours out of range (the text sim_scenario gives); and the set-level refusals, whose tool error starts with its code — a loop with no delayed binding (closes-loop), two bindings into one input (port-taken), more than one binding into a model (segment-cap: each becomes a schedule of up to 61 segments, and a scenario holds at most 100), a loop of more than 4 models (loop-too-large), a loop whose total delay does not exceed one resample interval (hours/60) per binding (loop-lag), one needing more than 24 sweeps (loop-horizon), loops needing more than 48 model runs together (run-cap); loop-lag and loop-horizon name the delay that would run at these hours and the most hours that run at this delay, and sim_system's view of a system holding the loop reports that most as maxHours. Refused after the charge (which pays for reading the models) and before anything runs: any binding that fails the check sim_bind applies (binds-port-v1; v2 for a unit conversion; v3 for a delay; a binding recorded before a check may be refused here by the same text), and a delayed binding whose target declares so many schedule segments before the delay ends that its schedule would pass the cap at these hours (segment-cap, code first as above). Every result names its seams: no model's own fitness gates cover the join, one unit line per binding (units agree; converted, naming the factor; or unchecked where a port declares no port.unit), and that a binding couples means: a driven model ran on its source's mean, so its std_dev is its own spread given that mean, not the joint system's. A driven model's resampled schedule keeps one random stream per realization across its segments, so splitting a run into segments adds nothing where no rate changes; the resampling itself is a piecewise-constant approximation of the source's mean, which smooths anything shorter than one resample window (hours/60). Public: no sign-in; writes nothing.

Run a seeded what-if scenario against a stored model: marking overrides, rate overrides, piecewise rate schedules, and params assignments to the model's declared structural parameters (arc weights, capacities — batch sizes and shelf sizes). Pure read — asking cannot change the model. Returns trajectory, final marking, metrics, depleted, contended, waitingForWork, neverFired (all three described below), caveats and assumptions. "samples" (default 60) is the trajectory's resolution: the number of evenly spaced points from 0 to hours inclusive at which times and every place's series (mean and std_dev per point) are reported — it sizes the answer, not the run, since metrics (throughput, mean, p95, utilization, inFlight) are time-weighted over every firing and do not change with the grid. "summary": true omits the times and series arrays entirely (the keys are absent, not null) and returns just final, finalStdDev, metrics, depleted, contended, waitingForWork, neverFired, caveats and assumptions — the verdict without the chart data, and the right form when nothing will be plotted. finalStdDev, summarized or not, is each place's spread across realizations at the horizon beside final's mean (absent at one realization and on the ode engine). Transitions declaring stages (phase-type durations) run with the declared lower spread — the engine expands them structurally and reports in the model's own vocabulary. A model-declared schedule (the day shape on a transition) is honored by every run; the scenario's own schedule or rate override still wins for that transition. "engine" picks the reading: "ssa" (default, discrete Gillespie — the right choice whenever counts are small enough that variance is the answer, or a schedule is in play) or "ode" (continuous mass-action; refuses a schedule, and refuses outright rather than silently misread a model carrying a read arc, inhibitor, reached capacity, guard or non-kinetic arc — Forecast's caveats name which). See docs/engine-selection.md for the full decision rule, including why an arc weight above 1 gets a genuinely different rate law from each engine, and sim_crosscheck to run both readings side by side. Contention arrives split three ways. contended holds only what the run was short of, something an operator could buy more of: kind conserved (a fixed pool such as staff) or bounded (a shelf with a declared capacity), each entry with fraction (the share of the run, across every realization, in which that place was the only input missing for the transitions under blocking), longest wait first. waitingForWork holds the engine's other entries in the same shape, and neither of its kinds is a shortage: queue (filled only by the model's own flow, such as waiting customers or work in service) means the work had not arrived, a statement about demand; state (a marker of which phase of a cycle the model is in, such as a stoplight's colour) means the model was in another phase. neverFired is one note per entry whose place never held a token in any run and whose transitions never fired, e.g. "work" never fired in any run: its input "q" never held a token. A queue or state entry like that appears only as its note; a conserved or bounded one keeps its contended row as well, since the run was short of it. The run cannot tell why a place stayed empty. Either it starts empty and nothing fills it (check its initial count and what produces it), or its producer was switched off (a rate of 0, or a scenario marking of 0 on it or upstream of it), or what fills it is rarer than this run's hours and realizations can sample (raise either; once the place fills, the note goes). sim_receipt and the HTTP scenario routes return the engine's unsplit contended list instead, each entry with its kind.

Returns a deterministic suite of firing sequences generated from ONE stored model, for replaying against an implementation outside sim (for example the code a sim_code_to_flow model came from): every transition fired once, a refusal probe for each transition at a marking where the model refuses it, and each conflict found (two enabled transitions where firing one disables the other) fired then probed. Each sequence starts from the declared initial marking: fire each of fires in order (the model fires each), then try refused once when it is set (the model refuses it); covers names its targets and final is the model's nonzero marking after fires. calls counts every fire and probe, not the resets. The reading is untimed, as sim_reach's is: each firing is one atomic step under the shared firing rule (read and inhibitor arcs, capacities, guards decidable from the marking); rates, schedules, delays and stages play no part, and a transition whose guard reads action parameters (or does not parse, or whose precondition was lost) is never fired or probed (excluded lists it). Each target is covered; unreachable, said only after an exhaustive search; or not-covered with a reason: depth-bound (markings or targets were left unexplored past depth: ask a larger depth), state-budget or work-budget (the 20,000-marking or 10,000,000-unit search budget), token-cap, guard-error, excluded-live (an excluded transition was enabled, so markings past it were not searched), excluded, or suite-budget (past 20,000 calls or 100,000 listed entries, each fire, probe and final-marking place counting one). notCovered gives each reason's detail and remedy, and coverage.conflict.complete is true only after an exhaustive search. depth is 1..16, default 8; a model over 128 transitions is refused. defects (0..200, default 50; 0 omits the figure) seeds up to that many single-point net mutants (all of them when the model has fewer; defects.candidates and defects.seeded say how many) (drop-arc, weight-up, weight-down, test-flip, read-to-consume, initial-up, initial-down, capacity-down, capacity-drop, guard-drop; rates and timing are never mutated), picked by seed (default 1), replays the suite against each with markings compared, and runs 8 seeded random testers (seeds seed to seed+7) against each with the suite's own call budget, resetting every depth fires. defects.suite.detected counts seeded mutants the suite told apart from the model, beside defects.random.detected's mean, min and max across the runs; it is a count of seeded defects, never a share of real bugs. survivors lists the mutants the suite missed: a survivor may be equivalent (no sequence of any length tells it apart), and this figure does not decide which. A mutant no suite sequence told apart and at least one could not drive to an answer (a firing it cannot decide, such as one past the token cap) is undecided: undecidedDefects lists it with its first error, and it counts in neither figure. The score is priced before it runs (defects.price) and refused over 2,000,000,000 work units: each call costs its arc test, its guard and passes over the model's token places, and each mutant its compile. summary true omits the sequences array (the key is absent, not null). MCP only: there is no HTTP route. Public: no sign-in; costs 2 compute tokens (one bounded search plus a priced replay battery). Writes nothing.

Mark an old version of your model as replaced by a newer one. The old id keeps working and its commons dedication (if any) stands — only the listing moves on to the successor. Both models must be yours. It also records a supersededBy edge (old → new), so sim_edges(old) and sim_neighbors(old, "supersededBy") show that the supersession happened. Re-pointing an old model to a different successor keeps the earlier edge too, and these edges carry no timestamp, so their order says nothing about which came last: the current successor is the `current` this tool returns (the model's own supersededBy record), not the edge list. Taxonomic edges on the old model are offered for migration to the current successor (sim_migration_offers, accepted with sim_accept_migration), never moved; structural and tool-written edges stay with it.

Returns a stored System's view ({id}), the models whose ports can bind to one declared port ({model, port, direction?, among?}), or the stored systems that name a model or a binding ({naming}); one form per call. {id}: the system's view, computed now and never stored (system-view-v2): each model (present, superseded), each binding with its binds-port check (v1, or v2 with its convert when it declares a unit conversion, or v3 with its delay when it declares one) and whether the set runs as a pipeline (with its loops and maxHours when it holds one), the isomorphicTo/subnetOf edges among the models with up to 128 witnesses re-checked (past that the view is truncated and the rest read unchecked), and the shape — one-net when one member has every other embedded in it by a direct witnessed edge or a chain of edges of one predicate through any stored model, every hop's witness and the composed map re-checked (net, embeddings: a direct one names edge and witness, a chained one its derivation, which sim_check_witness(derivation=…) re-checks; closure says what the chain search cost; sim_reach then answers on that net), pipeline when it holds bindings (they link rates, not tokens: sim_reach answers unknown), separate otherwise (notOneNet says why; a chain mixing subnetOf and isomorphicTo hops does not count), unknown when the witness bound was reached before any member qualified, or the chain search stopped at its hop, model or work bound (closure-incomplete: not a no). {model, port, direction?, among?}: what can bind to that declared port — every model with a port sim_bind would accept against it without a unit conversion (opposite direction, equal kind, units agreeing where both declare one), ranked units-agree first, then model id and port; each row carries the sim_bind arguments. A port whose declared unit differs is counted under excluded as unit-differs, not listed: it binds only through sim_bind with convertFrom, convertTo and convertFactor (binds-port-v2), and an empty answer means none binds without one. The universe is the visible catalog, or among (a collection's model members, at most 256, or a system's models, where a candidate that would take a bound port or close a loop of undelayed bindings is counted under excluded, never listed, and one that closes a loop already holding a delayed binding is listed with closesLoop: bindable-v3). Every model in the universe is searched (bindable-v2; bindable-v3 among a system), through a port index that keeps each model's ports by direction and kind, read once per content id and synced from the listing on every call (a model another instance stored is searched on the next call). At most 256 models the index does not hold yet are read per call (past it status is truncated with bound fetches and unscanned says how many were left; ask again and it continues), 100000 candidate ports checked and 50 rows listed: past any bound status is truncated, and if a member model cannot be read status is incomplete (listed under unreadable); in both cases the rows are in partialRows. candidates, scanned, unscanned, fromIndex, read and portChecks say exactly how far it searched; the index holds at most 16777216 bytes of ports (of which at most 2097152 for models no catalog-wide search has listed, such as an among search's or the asked model's) and at most 65536 bytes for one model, and a model past any of these is searched but not kept (notKept). The call's token covers its first 64 reads; each further 64 cost one more token, charged before they are read, and when the bucket cannot cover them status is truncated with bound rate. {naming}: every stored system whose models or bindings include that id (as: model or binding), sorted by system id, from an index synced from the store's listing on every call (a system another instance stored is seen on the next read). At most 100 systems listed (past it status is truncated, total says how many) and at most 256 not-yet-read systems fetched per call (past it status is incomplete and unindexed says how many are unread; ask again and it continues); the index holds at most 33554432 bytes of systems, and once full nothing more is indexed on that instance (status incomplete, unindexed says how many, and asking again does not change it); a stored system that cannot be decoded also keeps status incomplete (listed under unreadable), and asking again does not clear it; in every case the rows are in partialSystems. A system cannot be named (systems do not nest) and a collection is refused, naming the remedy; an id no entity holds any more (a deleted model) is answered when a system still names it. Systems are linkable and collectable: sim_link puts one in a collection (hasMember) and relates it with cites, about, broader and related. This is not refines: a system's edges are isomorphicTo and subnetOf (embeds by wiring / is the same coloured net). refines (trace inclusion with transitions matched by id) and preservesInvariant (the object's P-invariants pull back along an embedding) are proved one pair at a time by sim_prove and are never walked by closure, since neither is declared transitive. For "what contains this pattern?" ask sim_query with closure: {"select":["?m"],"where":[{"subject":"<pattern id>","predicate":"subnetOf","object":"?m","mode":"closure"}]} (isomorphicTo for copies; add {"subject":"<collection>","predicate":"hasMember","object":"?m"} to scope it), each row carrying a derivation sim_check_witness re-checks; a subnetOf row claims no behaviour inclusion. For "what does this model refine?" ask sim_query for one hop: {"select":["?m"],"where":[{"subject":"<model id>","predicate":"refines","object":"?m"}]}, each row an edge sim_prove recorded. A loop between models runs only through a binding that declares a delay (sim_bind delay; ROADMAP.md Phase 12j): a system may hold one, of at most 4 models, and its view lists it under pipeline.loops with pipeline.maxHours, the longest horizon sim_run_pipeline runs it at (0 when no horizon of 0.01 h or more does; pipeline.runnable covers only the checks that do not depend on the hours), where it is solved as an exact fixed point over the horizon (pipeline-loop-v1). A loop with no delayed binding is refused when a system is defined and by sim_run_pipeline, and sim_bind refuses a self-binding. A port search with among a system (bindable-v3) lists a candidate that closes a loop already holding a delayed binding, with that loop's models under closesLoop, and counts one that would close a loop of undelayed bindings under excluded as closes-loop (sim_bind with a delay can close it); a catalog or collection search does not consult stored bindings, so sim_create_system or sim_run_pipeline is where such a loop is refused. Public: no sign-in; costs 1 compute token (a port search, 1 more per 64 models read past its first 64); writes nothing.

Verify declared properties of a stored model: a report with one verdict per property, proved, refuted or unknown (unknown is never a pass, and ok is true only when every property is proved), with evidence for a proof and a counterexample for a refutation, plus caveats naming anything the analysis net could not express. Unset properties check bounded and deadlock-free. Every check starts from the model's own initial marking. Each verdict's method says how far it reaches. structural: from the incidence matrix, with no state search. A structural bounded holds for every initial marking (every place is covered by a P-invariant), and so does a structural conserves (the token total is a P-invariant). A structural invariant or mutual-exclusion verdict rests on a P-invariant's value at this initial marking (its evidence names the P-invariant and the value), so it holds from this marking and from any other that gives that P-invariant the same value, not from every marking: two workers sharing one lock token are proved mutually exclusive, and the same net with two lock tokens is refuted. exhaustive: every state reachable from this initial marking. witness: a firing sequence found by targeted search, decisive for this initial marking (a cycle that grows a place without bound). partial: the search was truncated, so only a refutation is sound. vacuity holds one reading per verdict, in the same order, decided by the engine. proved-vacuously: a proved unreachable whose target needs a token on a place never marked from the initial marking, a proved invariant or mutual-exclusion whose places are all never marked, a proved mutual-exclusion whose marked places never hold more than the bound on their own (so they never compete), or a proved bounded, terminating or conserves when no transition can fire at all (deadlock-free is then refuted); neverMarked names the never-marked places the claim reads and deadTransitions the transitions that would mark them, which never fire. not-vacuous: each place the claim reads (for unreachable, each place its target asks a token on) that neverMarked does not list is marked at some reachable marking the search found; a listed place adds nothing, and a target that asks for no token is not-vacuous without a search. method says what decided it: structural (the arcs alone) or exhaustive (sim_reach's search of the model under its firing rule, evaluating the guards that read only the marking, up to 20000 markings or 20000000 work units). unknown is open, never a no: the search stopped at a bound, skipped a firing the model may make (a guard over action parameters, a lost precondition), or did not run (the model declares a delay, which a one-step search misreads), or a kind with no antecedent has only guarded transitions to fire; it may still list neverMarked. not-applicable: a verdict other than proved, a proved live or reachable (each rests on a firing sequence), or a kind with no antecedent where some transition can fire. ok, the counts and every verdict do not read vacuity: a proved-vacuously verdict still counts as proved. To ask about another starting marking, store the model with that marking (sim_extend's set_initial) and verify it. Read the Verification properties section of sim://docs/authoring for each kind's fields. Free (no compute token); writes nothing.