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Carry offered taxonomic edges of a superseded model forward to its current successor.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. Caveats name what the analysable net encoded lossily. 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. 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.

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 (petri-pilot experiments/ode-minimax): 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).

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.

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. 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). 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). 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).