diff --git a/harmony_model_checker/charm/charm.c b/harmony_model_checker/charm/charm.c
index a87773ac..9efb9564 100644
--- a/harmony_model_checker/charm/charm.c
+++ b/harmony_model_checker/charm/charm.c
@@ -184,13 +184,6 @@ struct worker {
     // followed by its stack of Harmony values.
     struct context ctx;
     hvalue_t stack[MAX_CONTEXT_STACK];
-
-    // The state and context are saved here when entering an atomic section.
-    // This has to do with an optimization that prevents "breaks" in atomic
-    // sections that do not have 'breakable' operations like load and store.
-    char as_state[sizeof(struct state) + MAX_CONTEXT_BAG * (sizeof(hvalue_t) + 1)];
-    struct context as_ctx;
-    hvalue_t as_stack[MAX_CONTEXT_STACK];
 };
 
 #ifdef CACHE_LINE_ALIGNED
@@ -321,8 +314,8 @@ void spawn_thread(struct global *global, struct state *state, struct context *ct
 }
 
 // This function evaluates a predicate.  The predicate is evaluated in
-// every state as if by some thread.  sc contains the state (which is
-// read-only), while step keeps track of the state of the thread specifically.
+// every state as if by some thread.  sc contains the state, while
+// step keeps track of the state of the thread specifically.
 //
 // The code of a predicate ends in an Assert HVM instruction that
 // checks the value of the predicate.  The predicate may thus fail, but
@@ -332,97 +325,82 @@ void spawn_thread(struct global *global, struct state *state, struct context *ct
 // This function returns true iff the predicate evaluated to true.
 bool predicate_check(struct global *global, struct state *sc, struct step *step){
     assert(!step->ctx->failed);
-    unsigned int old_sp = step->ctx->sp;
-
-    while (step->ctx->sp >= old_sp) {
+    assert(step->ctx->sp == 1);     // just the argument
+    while (!step->ctx->terminated) {
         struct op_info *oi = global->code.instrs[step->ctx->pc].oi;
         (*oi->op)(global->code.instrs[step->ctx->pc].env, sc, step, global);
         if (step->ctx->failed) {
+            step->ctx->sp = 0;
             return false;
         }
     }
+    step->ctx->sp = 0;
     return true;
 }
 
 // Check all the invariants that the program specified.
 // Returns 0 if there are no issues, or the pc of some invariant that failed.
-unsigned int check_invariants(struct global *global, struct state *sc,
-                        struct node *before, struct step *step, bool self_loop){
-    // assert(step->ctx->sp == 0);
-    assert(!step->ctx->failed);
-    unsigned int saved_sp = step->ctx->sp;      // TODO DEBUG
-    unsigned int saved_pc = step->ctx->pc;      // TODO DEBUG
+unsigned int check_invariants(struct global *global, struct node *node,
+                        struct node *before, struct step *step){
+    struct state *state = node_state(node);
+    assert(state != NULL);
+
+    assert(step->ctx->sp == 0);
 
-    // before->pre == 0 means it started from a non-initialized state (i.e.,
-    // a state that was created before the initial thread finished.
+    // pre == 0 means it is a non-initialized state.
     hvalue_t args[2];   // (pre, post)
     if (node_state(before)->pre == 0) {
-        args[0] = sc->vars;
+        args[0] = state->vars;
     }
     else {
         args[0] = node_state(before)->pre;
     }
-    args[1] = sc->vars;
+    args[1] = state->vars;
 
     // Check each invariant
     for (unsigned int i = 0; i < global->ninvs; i++) {
         // No need to check edges other than self-loops
-        if (!global->invs[i].pre && !self_loop) {
+        if (!global->invs[i].pre && node != before) {
             continue;
         }
 
-        // assert(step->ctx->sp == 0);
-        // step->ctx->terminated = false;
-        // ctx_failure(step->ctx) = 0;
-        value_ctx_push(step->ctx, VALUE_LIST);
-        value_ctx_push(step->ctx, VALUE_TO_INT((step->ctx->pc << CALLTYPE_BITS) | CALLTYPE_NORMAL));
-        value_ctx_push(step->ctx, value_put_list(&step->engine, args, sizeof(args)));
-
+        assert(step->ctx->sp == 0);
+        step->ctx->terminated = step->ctx->failed = false;
+        ctx_failure(step->ctx) = 0;
         step->ctx->pc = global->invs[i].pc;
         step->ctx->vars = VALUE_DICT;
+        value_ctx_push(step->ctx, value_put_list(&step->engine, args, sizeof(args)));
+
         assert(strcmp(global->code.instrs[step->ctx->pc].oi->name, "Frame") == 0);
-        bool b = predicate_check(global, sc, step);
+        bool b = predicate_check(global, state, step);
         if (step->ctx->failed) {
             // printf("Invariant evaluation failed: %s\n", value_string(ctx_failure(step->ctx)));
             b = false;
         }
-        else {
-            (void) value_ctx_pop(step->ctx);   // return value
-            step->ctx->pc--;
-            if (step->ctx->pc != saved_pc) {
-                printf("X1 %u %u\n", step->ctx->pc, saved_pc);
-                panic("X1");
-            }
-        }
         if (!b) {
             // printf("INV %u %u failed\n", i, global->invs[i].pc);
             return global->invs[i].pc;
         }
     }
 
-    if (step->ctx->pc != saved_pc) {
-        panic("X2");
-    }
-
     return 0;
 }
 
 // Same as check_invariants but for "finally" predicates that are only
 // checked in final states.
 // Returns 0 if there are no issues, or the pc of the finally predicate if it failed.
-unsigned int check_finals(struct global *global, struct state *state, struct step *step){
+unsigned int check_finals(struct global *global, struct node *node, struct step *step){
+    struct state *state = node_state(node);
     assert(state != NULL);
 
     // Check each finally predicate
     for (unsigned int i = 0; i < global->nfinals; i++) {
-        assert(!step->ctx->failed);
-        // step->ctx->terminated = false;
-        // ctx_failure(step->ctx) = 0;
-        value_ctx_push(step->ctx, VALUE_LIST);
-        value_ctx_push(step->ctx, VALUE_TO_INT((step->ctx->pc << CALLTYPE_BITS) | CALLTYPE_NORMAL));
-        value_ctx_push(step->ctx, VALUE_LIST);
+        assert(step->ctx->sp == 0);
+        step->ctx->terminated = step->ctx->failed = false;
+        ctx_failure(step->ctx) = 0;
         step->ctx->pc = global->finals[i];
         step->ctx->vars = VALUE_DICT;
+        value_ctx_push(step->ctx, VALUE_LIST);
 
         assert(strcmp(global->code.instrs[step->ctx->pc].oi->name, "Frame") == 0);
         bool b = predicate_check(global, state, step);
@@ -430,10 +408,6 @@ unsigned int check_finals(struct global *global, struct state *state, struct ste
             // printf("Finally evaluation failed: %s\n", value_string(ctx_failure(step->ctx)));
             b = false;
         }
-        else {
-            (void) value_ctx_pop(step->ctx);   // return value
-            step->ctx->pc--;
-        }
         if (!b) {
             // printf("FIN %u %u failed\n", i, global->finals[i]);
             return global->finals[i];
@@ -442,9 +416,26 @@ unsigned int check_finals(struct global *global, struct state *state, struct ste
     return 0;
 }
 
-// This function is called when a new edge has been generated.
-static void process_edge(struct worker *w, struct edge *edge) {
-    struct node *node = edge->src;
+// This function is called when a new edge has been generated, possibly to
+// a new state with an uninitialized node.
+static void process_edge(struct worker *w, struct edge *edge, mutex_t *lock) {
+    struct node *node = edge->src, *next = edge->dst;
+
+    // mutex_acquire(lock);    ==> this lock is already acquired
+
+    bool initialized = next->initialized;
+    if (!initialized) {
+        next->initialized = true;
+        next->reachable = true;
+        next->failed = edge->failed;
+        next->u.ph1.lock = lock;
+        next->u.ph1.next = w->results;
+        w->results = next;
+        w->count++;
+        w->enqueued++;
+    }
+
+    mutex_release(lock);
 
 #ifdef DELAY_INSERT
     // Don't do the forward edge at this time as that would involve locking
@@ -468,6 +459,27 @@ static void process_edge(struct worker *w, struct edge *edge) {
         *pe = edge;
     }
 #endif
+
+    // If this is a good and normal (non-choosing) edge, check all the invariants.
+    if (!edge->failed && !edge->choosing) {
+        if (w->global->ninvs != 0) {
+            unsigned int inv = 0;
+            if (!initialized) {      // try self-loop if a new node
+                inv = check_invariants(w->global, next, next, &w->inv_step);
+            }
+            if (inv == 0 && next != node) { // try new edge
+                inv = check_invariants(w->global, next, node, &w->inv_step);
+            }
+            if (inv != 0) {
+                struct failure *f = new_alloc(struct failure);
+                f->type = FAIL_INVARIANT;
+                f->edge = edge;
+                f->address = VALUE_TO_PC(inv);
+                f->next = w->failures;
+                w->failures = f;
+            }
+        }
+    }
 }
 
 // This is the main workhorse function of model checking: explore a state and
@@ -483,13 +495,15 @@ static bool onestep(
     struct worker *w,       // thread info
     struct node *node,      // starting node
     struct state *sc,       // actual state
-    int ctx_index,          // indexes into state's context bag
+    hvalue_t ctx,           // context identifier
     struct step *step,      // step info
     hvalue_t choice,        // if about to make a choice, which choice?
     bool interrupt,         // start with invoking interrupt handler
-    bool infloop_detect     // try to detect infloop from the start
+    bool infloop_detect,    // try to detect infloop from the start
+    int multiplicity       // #contexts that are in the current state
 ) {
     assert(state_size(sc) == state_size(node_state(node)));
+
     assert(!step->ctx->terminated);
     assert(!step->ctx->failed);
     assert(step->engine.allocator == &w->allocator);
@@ -508,11 +522,20 @@ static bool onestep(
     // TODO.  Not clear if this is necessary.
     hvalue_t choice_copy = choice;
 
-    hvalue_t ctx = state_contexts(sc)[ctx_index];
-    bool choosing = false, pausing = false;
+    bool choosing = false;
     struct dict *infloop = NULL;        // infinite loop detector
     unsigned int instrcnt = 0;          // keeps track of #instruction executed
-    unsigned int as_instrcnt = 0;       // optimization for atomic sections
+
+    // We *may* need to keep a copy of various parts of the state here in case we
+    // enter an atomic section.
+    // TODO:  can we remove the repeated allocation of state here?
+#ifdef HEAP_ALLOC
+    char *as_state = malloc(sizeof(struct state) + MAX_CONTEXT_BAG * (sizeof(hvalue_t) + 1));
+#else
+    char as_state[sizeof(struct state) + MAX_CONTEXT_BAG * (sizeof(hvalue_t) + 1)];
+#endif
+    hvalue_t as_context = 0;
+    unsigned int as_instrcnt = 0;
 
     bool rollback = false, stopped = false;
     bool terminated = false, infinite_loop = false;
@@ -597,9 +620,11 @@ static bool onestep(
 
             // If not, but the context was not in atomic mode, save the current
             // state to restore to in case we have to "break".
+            // TODO.  We do not really need to store the context in
+            //        the hashtable here.  We could just copy it like the state.
             else if (step->ctx->atomic == 0) {
-                memcpy(w->as_state, sc, state_size(sc));
-                memcpy(&w->as_ctx, step->ctx, ctx_size(step->ctx));
+                memcpy(as_state, sc, state_size(sc));
+                as_context = value_put_context(&step->engine, step->ctx);
                 as_instrcnt = instrcnt;
             }
             
@@ -612,6 +637,7 @@ static bool onestep(
         else if (instrs[pc].atomicdec) {
             (*oi->op)(instrs[pc].env, sc, step, global);
             if (step->ctx->atomic == 0) {
+                as_context = 0;
                 as_instrcnt = 0;
             }
         }
@@ -768,11 +794,6 @@ static bool onestep(
             break;
         }
 
-        else if (next_instr->pause) {
-            pausing = true;
-            break;
-        }
-
         // See if we need to break out of this step.  If the atomicFlag is
         // set, then definitely not.  If it is not set, then it gets
         // complicated.  If the atomic count > 0, then we may have delayed
@@ -840,7 +861,7 @@ static bool onestep(
         }
     }
 
-    // No longer need the infloop set.
+    // No longer need this.  It was wasted effort.
     // TODO.  Perhaps this suggests a way to automatically determine the
     //        number of instructions to run before trying to detect an
     //        infinite loop, which is currently hardwired at 1000.
@@ -851,52 +872,22 @@ static bool onestep(
     // See if we need to roll back to the start of an atomic section.
     hvalue_t after;
     if (rollback) {
-        struct state *state = (struct state *) w->as_state;
+        struct state *state = (struct state *) as_state;
         memcpy(sc, state, state_size(state));
-        memcpy(step->ctx, &w->as_ctx, ctx_size(&w->as_ctx));
+        after = as_context;
         instrcnt = as_instrcnt;
     }
-
-    // Remove old context from the bag
-    context_remove(sc, ctx);
-
-    unsigned int inv = 0;
-    if (w->global->ninvs != 0 && !step->ctx->failed) {
-        // Check invariants when executing Pause
-        // TODO.  Also check finally clauses
-        if (pausing) {
-            inv = check_invariants(w->global, sc, node, step, true);
-        }
-        // If this is a good and normal transition thus far, check all the invariants
-        // that have 'pre' in them.
-        //
-        // TODO.  If this is a self-loop (state is unchanged), we don't have to check.
-        else if (!stopped && !choosing) {
-            inv = check_invariants(w->global, sc, node, step, false);
-        }
-    }
-    if (w->global->nfinals != 0 && !step->ctx->failed && inv == 0 && pausing &&
-            value_state_all_eternal(sc) && value_ctx_all_eternal(sc->stopbag)) {
-        // Use negative value to distinguish from normal invariant
-        inv = -check_finals(global, sc, step);
+    else {
+        // Store new context in value directory.  Must be immutable now.
+        after = value_put_context(&step->engine, step->ctx);
     }
 
-    // Store new context in value directory.  Must be immutable now.
-    after = value_put_context(&step->engine, step->ctx);
+#ifdef HEAP_ALLOC
+    free(as_state);
+#endif
 
-    // Add new context to state unless it's terminated or stopped.
-    int index;
-    if (stopped) {
-        sc->stopbag = value_bag_add(&step->engine, sc->stopbag, after, 1);
-        index = -1;
-    }
-    else if (!terminated) {
-        index = context_add(sc, after);
-        assert(index >= 0);
-    }
-    else {
-        index = -1;
-    }
+    // Remove old context from the bag
+    context_remove(sc, ctx);
 
     // If choosing, save in state.  If some invariant uses "pre", then
     // also keep track of "pre" state.
@@ -910,46 +901,28 @@ static bool onestep(
     // So we only do it in case there are such invariant, and then only for
     // choosing states.
     if (choosing) {
-        assert(index >= 0);
-        sc->chooser = index;
+        sc->choosing = after;
         sc->pre = global->inv_pre ? node_state(node)->pre : sc->vars;
     }
     else {
         sc->pre = sc->vars;
     }
 
-    // See if this state has been computed before by looking up the node,
-    // or allocate if not.  Sets a lock on that node.
-    // TODO.  It seems that perhaps the node does not need to be locked
-    //        if it's not new.  After all, we don't change it in that case.
-    unsigned int size = state_size(sc);
-    mutex_t *lock;
-    bool new;
-    struct dict_assoc *hn = dict_find_lock(w->visited, &w->allocator,
-                sc, size, &new, &lock);
-    struct node *next = (struct node *) &hn[1];
-    if (new) {
-        assert(!next->initialized);
-        next->initialized = true;       // TODO.  Do we still need this?
-        next->reachable = true;         // TODO.  Should this be done here?
-        next->failed = step->ctx->failed;
-        next->u.ph1.lock = lock;
-        next->u.ph1.next = w->results;
-        w->results = next;
-        w->count++;
-        w->enqueued++;
+    // Add new context to state unless it's terminated or stopped.
+    if (stopped) {
+        sc->stopbag = value_bag_add(&step->engine, sc->stopbag, after, 1);
     }
-    else {
-        assert(next->initialized);
+    else if (!terminated) {
+        context_add(sc, after);
     }
-    mutex_release(lock);
 
     // Allocate and initialize edge now.
     struct edge *edge = walloc(w, sizeof(struct edge) + step->nlog * sizeof(hvalue_t), false, false);
     edge->src = node;
-    edge->ctx_index = ctx_index;
+    edge->ctx = ctx;
     edge->choice = choice_copy;
     edge->interrupt = interrupt;
+    edge->multiplicity = multiplicity;
     edge->after = after;
     edge->ai = step->ai;     step->ai = NULL;
     memcpy(edge_log(edge), step->log, step->nlog * sizeof(hvalue_t));
@@ -957,22 +930,25 @@ static bool onestep(
     edge->nsteps = instrcnt;
     edge->choosing = choosing;
     edge->failed = step->ctx->failed;
-    edge->inv = inv;
 
     // If a failure has occurred, keep track of that too.
     if (step->ctx->failed) {
         struct failure *f = new_alloc(struct failure);
-        if (inv != 0) {
-            f->type = FAIL_INVARIANT;
-            f->address = VALUE_TO_PC(inv);
-        }
-        else {
-            f->type = infinite_loop ? FAIL_TERMINATION : FAIL_SAFETY;
-        }
+        f->type = infinite_loop ? FAIL_TERMINATION : FAIL_SAFETY;
         f->edge = edge;
         f->next = w->failures;
         w->failures = f;
     }
+
+    // See if this state has been computed before by looking up the node,
+    // or allocate if not.  Sets a lock on that node.
+    // TODO.  It seems that perhaps the node does not need to be locked
+    //        if it's not new.  After all, we don't change it in that case.
+    unsigned int size = state_size(sc);
+    mutex_t *lock;
+    bool new;
+    struct dict_assoc *hn = dict_find_lock(w->visited, &w->allocator,
+                sc, size, &new, &lock);
     edge->dst = (struct node *) &hn[1];
 
 #ifdef NO_PROCESSING   // I use this sometime for profiling
@@ -989,7 +965,7 @@ static bool onestep(
         }
     }
 #else
-    process_edge(w, edge);
+    process_edge(w, edge, lock);
 #endif
 
     return true;
@@ -999,7 +975,8 @@ static bool onestep(
 // state ctx. If it is a "choosing state" (the thread is about to execute a
 // Choose instruction), then choice contains that choice to be tried.  Since
 // threads are anonymous and multiple threads can be in the same state,
-// Any resulting states should be buffered in w->results.
+// multiplicity gives the number of threads that can make this step.  Any
+// resulting states should be buffered in w->results.
 //
 // The hard work of makestep is accomplished by function onestep().  makestep()
 // may invoke onestep() multiple times.  One reason is to explore interrupts
@@ -1020,24 +997,21 @@ static bool onestep(
 //  4) restarting 3) if an infinite loop is detected.
 static void make_step(
     struct worker *w,
-    struct node *node,    // source node (and state)
-    int ctx_index,        // index into context bag
-    hvalue_t choice       // if about to make a choice, which choice?
+    struct node *node,
+    hvalue_t ctx,
+    hvalue_t choice,       // if about to make a choice, which choice?
+    int multiplicity       // #contexts that are in the current state
 ) {
     struct step step;
     memset(&step, 0, sizeof(step));
     step.engine.allocator = &w->allocator;
     step.engine.values = w->global->values;
 
-    // Get the state and the context
-    struct state *state = node_state(node);
-    hvalue_t ctx = state_contexts(state)[ctx_index];
-
     // Make a copy of the state.
     //
     // TODO. Would it not be more efficient to have a fixed state variable
     //       in the worker structure, similar to what we do for contexts (w->ctx)?
-    unsigned int statesz = state_size(state);
+    unsigned int statesz = state_size(node_state(node));
     // Room to grown in copy for op_Spawn
 #ifdef HEAP_ALLOC
     char *copy = malloc(statesz + 64*sizeof(hvalue_t));
@@ -1045,7 +1019,7 @@ static void make_step(
     char copy[statesz + 64*sizeof(hvalue_t)];
 #endif
     struct state *sc = (struct state *) copy;
-    memcpy(sc, state, statesz);
+    memcpy(sc, node_state(node), statesz);
     assert(step.engine.allocator == &w->allocator);
 
     // Make a copy of the context
@@ -1062,34 +1036,34 @@ static void make_step(
     // TODO.  Should probably also check that the context is not in atomic mode.
     //        This could possibly happen if the thread was stopped in an atomic
     //        section and is then later restarted.
-    if (sc->chooser < 0 && cc->extended && ctx_trap_pc(cc) != 0 && !cc->interruptlevel) {
-        bool succ = onestep(w, node, sc, ctx_index, &step, choice, true, false);
+    if (sc->choosing == 0 && cc->extended && ctx_trap_pc(cc) != 0 && !cc->interruptlevel) {
+        bool succ = onestep(w, node, sc, ctx, &step, choice, true, false, multiplicity);
         assert(step.engine.allocator == &w->allocator);
         if (!succ) {        // ran into an infinite loop
             step.nlog = 0;
-            memcpy(sc, state, statesz);
+            memcpy(sc, node_state(node), statesz);
             memcpy(&w->ctx, cc, size);
             assert(step.engine.allocator == &w->allocator);
-            (void) onestep(w, node, sc, ctx_index, &step, choice, true, true);
+            (void) onestep(w, node, sc, ctx, &step, choice, true, true, multiplicity);
             assert(step.engine.allocator == &w->allocator);
         }
 
         // Restore the state
-        memcpy(sc, state, statesz);
+        memcpy(sc, node_state(node), statesz);
         memcpy(&w->ctx, cc, size);
         assert(step.engine.allocator == &w->allocator);
     }
 
     // Explore the state normally (not as an interrupt).
-    sc->chooser = -1;
-    bool succ = onestep(w, node, sc, ctx_index, &step, choice, false, false);
+    sc->choosing = 0;
+    bool succ = onestep(w, node, sc, ctx, &step, choice, false, false, multiplicity);
     assert(step.engine.allocator == &w->allocator);
     if (!succ) {        // ran into an infinite loop
         step.nlog = 0;
-        memcpy(sc, state, statesz);
+        memcpy(sc, node_state(node), statesz);
         memcpy(&w->ctx, cc, size);
         assert(step.engine.allocator == &w->allocator);
-        (void) onestep(w, node, sc, ctx_index, &step, choice, false, true);
+        (void) onestep(w, node, sc, ctx, &step, choice, false, true, multiplicity);
         assert(step.engine.allocator == &w->allocator);
     }
 
@@ -1098,21 +1072,16 @@ static void make_step(
 #endif
 }
 
-// TODO.  I can't recall how this works...
-char *ctx_status(struct node *node, int ctx_index) {
-    if (ctx_index < 0) {
-        return "terminated";
-    }
-    struct state *state = node_state(node);
-    if (state->chooser >= 0 && state->chooser == ctx_index) {
+char *ctx_status(struct node *node, hvalue_t ctx) {
+    if (node_state(node)->choosing == ctx) {
         return "choosing";
     }
-    while (node_state(node)->chooser >= 0) {
+    while (node_state(node)->choosing != 0) {
         node = node->u.ph2.u.to_parent->src;
     }
     struct edge *edge;
     for (edge = node->fwd; edge != NULL; edge = edge->fwdnext) {
-        if (edge->ctx_index == ctx_index) {
+        if (edge->ctx == ctx) {
             break;
         }
     }
@@ -1126,7 +1095,6 @@ void print_context(
     struct global *global,
     FILE *file,
     hvalue_t ctx,
-    int ctx_index,
     struct callstack *cs,
     int tid,
     struct node *node,
@@ -1247,7 +1215,7 @@ void print_context(
             fprintf(file, "%s\"mode\": \"stopped\"", prefix);
         }
         else {
-            fprintf(file, "%s\"mode\": \"%s\"", prefix, ctx_status(node, ctx_index));
+            fprintf(file, "%s\"mode\": \"%s\"", prefix, ctx_status(node, ctx));
         }
     }
     fprintf(file, "\n");
@@ -1318,8 +1286,6 @@ static void make_microstep(
 
 // Similar to onestep.  Used to recompute a faulty execution, that is, to
 // generate the detailed execution of a counter-examples.
-//
-// TODO.  Eventually should probably have onestep() and twostep() be the same
 static void twostep(
     struct global *global,
     struct state *sc,
@@ -1328,13 +1294,10 @@ static void twostep(
     hvalue_t choice,
     bool interrupt,
     unsigned int nsteps,
-    int inv,
     unsigned int pid,
     struct macrostep *macro
 ){
-    hvalue_t before = sc->pre;
-
-    sc->chooser = -1;
+    sc->choosing = 0;
 
     struct step step;
     memset(&step, 0, sizeof(step));
@@ -1460,47 +1423,6 @@ static void twostep(
         }
     }
 
-    // See if an invariant needs to be checked (because it should fail).
-    if (inv != 0) {
-        assert(!step.ctx->failed);
-        // step.ctx->terminated = false;
-        // ctx_failure(step.ctx) = 0;
-        value_ctx_push(step.ctx, VALUE_LIST);
-        value_ctx_push(step.ctx, VALUE_TO_INT((step.ctx->pc << CALLTYPE_BITS) | CALLTYPE_NORMAL));
-
-        if (inv > 0) {      // invariant
-            hvalue_t args[2];   // (pre, post)
-            if (before == 0) {
-                args[0] = sc->vars;
-            }
-            else {
-                args[0] = before;
-            }
-            args[1] = sc->vars;
-            value_ctx_push(step.ctx, value_put_list(&step.engine, args, sizeof(args)));
-            step.ctx->pc = inv;
-        }
-        else {              // finally
-            value_ctx_push(step.ctx, VALUE_LIST);
-            step.ctx->pc = -inv;
-        }
-
-        step.ctx->vars = VALUE_DICT;
-        assert(strcmp(global->code.instrs[step.ctx->pc].oi->name, "Frame") == 0);
-
-        unsigned int old_sp = step.ctx->sp;
-        while (step.ctx->sp >= old_sp) {
-            int pc = step.ctx->pc;
-            struct instr *instrs = global->code.instrs;
-            struct op_info *oi = instrs[pc].oi;
-            (*oi->op)(instrs[pc].env, sc, &step, global);
-            make_microstep(sc, step.ctx, step.callstack, false, instrs[pc].choose, choice, false, &step, macro);
-            if (step.ctx->failed) {
-                break;
-            }
-        }
-    }
-
     // Remove old context from the bag
     context_remove(sc, ctx);
 
@@ -1511,7 +1433,7 @@ static void twostep(
         sc->stopbag = value_bag_add(&step.engine, sc->stopbag, after, 1);
     }
     else if (!step.ctx->terminated) {
-        (void) context_add(sc, after);
+        context_add(sc, after);
     }
 
     free(step.ctx);
@@ -1553,11 +1475,6 @@ void path_serialize(
     global->macrosteps[global->nmacrosteps++] = macro;
 }
 
-// Retrieve the starting context of an edge
-static hvalue_t get_context(struct edge *e){
-    return state_contexts(node_state(e->src))[e->ctx_index];
-}
-
 void path_recompute(struct global *global){
     struct node *node = global->graph.nodes[0];
     struct state *sc = calloc(1,
@@ -1572,7 +1489,7 @@ void path_recompute(struct global *global){
         /* Find the starting context in the list of processes.  Prefer
          * sticking with the same pid if possible.
          */
-        hvalue_t ctx = get_context(e);
+        hvalue_t ctx = e->ctx;
         unsigned int pid;
         if (global->processes[global->oldpid] == ctx) {
             pid = global->oldpid;
@@ -1595,8 +1512,6 @@ void path_recompute(struct global *global){
         macro->cs = global->callstacks[pid];
 
         // Recreate the steps
-        //
-        // TODO.  Why not just pass it e?
         twostep(
             global,
             sc,
@@ -1605,7 +1520,6 @@ void path_recompute(struct global *global){
             e->choice,
             e->interrupt,
             e->nsteps,
-            e->inv,
             pid,
             macro
         );
@@ -1828,7 +1742,7 @@ static void path_output_macrostep(struct global *global, FILE *file, struct macr
     free(c);
 
     fprintf(file, "      \"context\": {\n");
-    print_context(global, file, get_context(macro->edge), macro->edge->ctx_index, macro->cs, macro->tid, macro->edge->dst, "        ");
+    print_context(global, file, macro->edge->ctx, macro->cs, macro->tid, macro->edge->dst, "        ");
     fprintf(file, "      },\n");
 
     if (macro->trim != NULL && macro->value != 0) {
@@ -1838,7 +1752,7 @@ static void path_output_macrostep(struct global *global, FILE *file, struct macr
     }
 
     fprintf(file, "      \"microsteps\": [");
-    struct context *oldctx = value_get(get_context(macro->edge), NULL);
+    struct context *oldctx = value_get(macro->edge->ctx, NULL);
     struct callstack *oldcs = NULL;
     for (unsigned int i = 0; i < macro->nmicrosteps; i++) {
         struct microstep *micro = macro->microsteps[i];
@@ -1870,19 +1784,7 @@ static void path_output_macrostep(struct global *global, FILE *file, struct macr
     fprintf(file, "      \"contexts\": [\n");
     for (unsigned int i = 0; i < macro->nprocesses; i++) {
         fprintf(file, "        {\n");
-
-        // Find the "context index".  May not be there as thread may have terminated
-        int ctx_index;
-        for (ctx_index = 0; ctx_index < state->bagsize; ctx_index++) {
-            if (state_contexts(state)[ctx_index] == macro->processes[i]) {
-                break;
-            }
-        }
-        if (ctx_index >= state->bagsize) {
-            ctx_index = -1;
-        }
-
-        print_context(global, file, macro->processes[i], ctx_index, macro->callstacks[i], i, macro->edge->dst, "          ");
+        print_context(global, file, macro->processes[i], macro->callstacks[i], i, macro->edge->dst, "          ");
         fprintf(file, "        }");
         if (i < macro->nprocesses - 1) {
             fprintf(file, ",");
@@ -1959,19 +1861,19 @@ static void path_optimize(struct global *global){
 #endif
 
     cbs = calloc(1, sizeof(*cbs));
-    cbs->before = get_context(global->macrosteps[0]->edge);
+    cbs->before = global->macrosteps[0]->edge->ctx;
     ncbs = 0;
     current = global->macrosteps[0]->edge->after;
 
     // Figure out where the actual context switches are.  Each context
     // block is a sequence of edges executed by the same thread
     for (unsigned int i = 1; i < global->nmacrosteps; i++) {
-        if (get_context(global->macrosteps[i]->edge) != current) {
+        if (global->macrosteps[i]->edge->ctx != current) {
             cbs[ncbs].after = current;
             cbs[ncbs++].end = i;
             cbs = realloc(cbs, (ncbs + 1) * sizeof(*cbs));
             cbs[ncbs].start = i;
-            cbs[ncbs].before = get_context(global->macrosteps[i]->edge);
+            cbs[ncbs].before = global->macrosteps[i]->edge->ctx;
         }
         current = global->macrosteps[i]->edge->after;
     }
@@ -2039,12 +1941,12 @@ static void path_optimize(struct global *global){
     for (unsigned int i = 0; i < global->nmacrosteps; i++) {
         // printf("--> %u/%u\n", i, global->nmacrosteps);
         // Find the edge
-        hvalue_t ctx = get_context(global->macrosteps[i]->edge);
+        hvalue_t ctx = global->macrosteps[i]->edge->ctx;
         hvalue_t choice = global->macrosteps[i]->edge->choice;
         bool interrupt = global->macrosteps[i]->edge->interrupt;
         struct edge *e;
         for (e = node->fwd; e != NULL; e = e->fwdnext) {
-            if (get_context(e) == ctx && e->choice == choice && (bool) e->interrupt == interrupt) {
+            if (e->ctx == ctx && e->choice == choice && (bool) e->interrupt == interrupt) {
                 global->macrosteps[i]->edge = e;
                 break;
             }
@@ -2098,7 +2000,7 @@ static void path_trim(struct global *global, struct engine *engine){
 
         // Look up the last microstep of this thread, which wasn't the
         // last one to take a step overall
-        struct context *cc = value_get(get_context(macro->edge), NULL);
+        struct context *cc = value_get(macro->edge->ctx, NULL);
         struct microstep *ls = macro->microsteps[macro->nmicrosteps - 1];
         struct instr *fi = &instrs[cc->pc];
         struct instr *li = &instrs[ls->ctx->pc];
@@ -2265,7 +2167,7 @@ static enum busywait is_stuck(
 	node->visited = true;
 	enum busywait result = BW_ESCAPE;
     for (struct edge *edge = node->fwd; edge != NULL; edge = edge->fwdnext) {
-        if (get_context(edge) == ctx) {
+        if (edge->ctx == ctx) {
 			if (edge->dst == node) {
 				node->visited = false;
 				return BW_ESCAPE;
@@ -2329,10 +2231,12 @@ void do_work1(struct worker *w, struct node *node){
     // the possible choices for the thread that is executing.  For a non-choosing
     // state, we explore all the different threads that can make a step.
     struct state *state = node_state(node);
-    if (state->chooser >= 0) {      // choosing state
-        hvalue_t chooser = state_contexts(state)[state->chooser];
-        // The actual set of choices is on top of the stack
-        struct context *cc = value_get(chooser, NULL);
+    if (state->choosing != 0) {
+        assert(VALUE_TYPE(state->choosing) == VALUE_CONTEXT);
+
+        // state->choosing is the context of the thread that is making a choice.
+        // The actual set of choices is on top of its stack
+        struct context *cc = value_get(state->choosing, NULL);
         assert(cc != NULL);
         assert(cc->sp > 0);
         hvalue_t s = ctx_stack(cc)[cc->sp - 1];
@@ -2344,14 +2248,26 @@ void do_work1(struct worker *w, struct node *node){
 
         // Explore each choice.
         for (unsigned int i = 0; i < size; i++) {
-            make_step(w, node, state->chooser, vals[i]);
+            make_step(
+                w,
+                node,
+                state->choosing,
+                vals[i],
+                1
+            );
         }
     }
     else {
         // Explore each thread that can make a step.
         for (unsigned int i = 0; i < state->bagsize; i++) {
             assert(VALUE_TYPE(state_contexts(state)[i]) == VALUE_CONTEXT);
-            make_step(w, node, i, 0);
+            make_step(
+                w,
+                node,
+                state_contexts(state)[i],
+                0,
+                multiplicities(state)[i]
+            );
         }
     }
 }
@@ -3088,6 +3004,21 @@ static void tarjan(struct global *global){
     global->ncomponents = comp_id;
 }
 
+char *state_string(struct state *state){
+    struct strbuf sb;
+    strbuf_init(&sb);
+
+    char *v;
+    strbuf_printf(&sb, "{");
+    v = value_string(state->vars);
+    strbuf_printf(&sb, "%s", v); free(v);
+    v = value_string(state->choosing);
+    strbuf_printf(&sb, ",%s", v); free(v);
+    v = value_string(state->stopbag);
+    strbuf_printf(&sb, ",%s}", v); free(v);
+    return strbuf_convert(&sb);
+}
+
 // This routine removes all nodes that have a single incoming edge and it's
 // an "epsilon" edge (empty print log).  These are essentially useless nodes.
 static void destutter1(struct global *global){
@@ -3601,14 +3532,12 @@ int exec_model_checker(int argc, char **argv){
     init_ctx->vars = VALUE_DICT;
     init_ctx->atomic = 1;
     init_ctx->initial = true;
-    init_ctx->eternal = true;
     init_ctx->atomicFlag = true;
     value_ctx_push(init_ctx, VALUE_LIST);
 
     // Now create the state
     struct state *state = calloc(1, sizeof(struct state) + sizeof(hvalue_t) + 1);
     state->vars = VALUE_DICT;
-    state->chooser = -1;
     hvalue_t ictx = value_put_context(&engine, init_ctx);
     state->bagsize = 1;
     state_contexts(state)[0] = ictx;
@@ -3669,7 +3598,6 @@ int exec_model_checker(int argc, char **argv){
         w->profile = calloc(global->code.len, sizeof(*w->profile));
 
         // Create a context for evaluating invariants
-        // TODO.  May be obsolete soon
         w->inv_step.ctx = calloc(1, sizeof(struct context) +
                                 MAX_CONTEXT_STACK * sizeof(hvalue_t));
         // w->inv_step.ctx->name = value_put_atom(&engine, "__invariant__", 13);
@@ -3787,8 +3715,6 @@ int exec_model_checker(int argc, char **argv){
 
     printf("* Phase 3: analysis\n");
 
-    bool scc = false;
-
     // If no failures were detected (yet), determine strongly connected components
     // and look for non-terminating states.
     if (global->failures == NULL) {
@@ -3798,7 +3724,7 @@ int exec_model_checker(int argc, char **argv){
         }
         double now = gettime();
         tarjan(global);
-        scc = true;
+        global->phase2 = true;
 
         // Compute shortest path to initial state for each node.
         shortest_path(global);
@@ -3913,9 +3839,9 @@ int exec_model_checker(int argc, char **argv){
                     add_failure(&global->failures, f);
                     // break;
                 }
-#ifdef notdef
+
                 // Check "finally" clauses
-                unsigned int fin = check_finals(global, node_state(node), &fin_step);
+                unsigned int fin = check_finals(global, node, &fin_step);
                 if (fin != 0) {
                     struct failure *f = new_alloc(struct failure);
                     f->type = FAIL_FINALLY;
@@ -3923,7 +3849,6 @@ int exec_model_checker(int argc, char **argv){
                     f->address = VALUE_TO_PC(fin);
                     add_failure(&global->failures, f);
                 }
-#endif
             }
         }
 
@@ -3988,7 +3913,7 @@ int exec_model_checker(int argc, char **argv){
                     struct state *state = node_state(node);
                     unsigned int j;
                     for (j = 0; j < state->bagsize; j++) {
-                        if (state_contexts(state)[j] == get_context(edge)) {
+                        if (state_contexts(state)[j] == edge->ctx) {
                             break;
                         }
                     }
@@ -4021,9 +3946,7 @@ int exec_model_checker(int argc, char **argv){
                 struct node *node = global->graph.nodes[i];
                 assert(node->id == i);
                 fprintf(df, "\nNode %d:\n", node->id);
-                if (scc) {
-                    fprintf(df, "    component: %d\n", node->u.ph2.component);
-                }
+                fprintf(df, "    component: %d\n", node->u.ph2.component);
                 if (node->u.ph2.u.to_parent != NULL) {
                     fprintf(df, "    ancestors:");
                     for (struct node *n = node->u.ph2.u.to_parent->src;; n = n->u.ph2.u.to_parent->src) {
@@ -4034,19 +3957,7 @@ int exec_model_checker(int argc, char **argv){
                     }
                     fprintf(df, "\n");
                 }
-                struct state *state = node_state(node);
-                if (state->pre == 0) {
-                    fprintf(df, "    pre: uninitializeds\n");
-                }
-                else if (state->pre != state->vars) {
-                    fprintf(df, "    pre: %s\n", value_string(state->pre));
-                }
-                fprintf(df, "    vars: %s\n", value_string(state->vars));
-                fprintf(df, "    contexts:\n");
-                for (unsigned int i = 0; i < state->bagsize; i++) {
-                    fprintf(df, "      %"PRI_HVAL": %u\n",
-                        state_contexts(state)[i], multiplicities(state)[i]);
-                }
+                fprintf(df, "    vars: %s\n", value_string(node_state(node)->vars));
                 // fprintf(df, "    len: %u %u\n", node->len, node->steps);
                 if (node->failed) {
                     fprintf(df, "    failed\n");
@@ -4055,9 +3966,9 @@ int exec_model_checker(int argc, char **argv){
                 int eno = 0;
                 for (struct edge *edge = node->fwd; edge != NULL; edge = edge->fwdnext, eno++) {
                     fprintf(df, "        %d:\n", eno);
-                    struct context *ctx = value_get(get_context(edge), NULL);
+                    struct context *ctx = value_get(edge->ctx, NULL);
                     fprintf(df, "            node: %d (%d)\n", edge->dst->id, edge->dst->u.ph2.component);
-                    fprintf(df, "            context before: %"PRIx64" pc=%d\n", get_context(edge), ctx->pc);
+                    fprintf(df, "            context before: %"PRIx64" pc=%d\n", edge->ctx, ctx->pc);
                     ctx = value_get(edge->after, NULL);
                     fprintf(df, "            context after:  %"PRIx64" pc=%d\n", edge->after, ctx->pc);
                     if (edge->failed != 0) {
@@ -4271,7 +4182,6 @@ int exec_model_checker(int argc, char **argv){
         // If it was an invariant failure, add one more macrostep
         // to replay the invariant code.
         struct edge *edge;
-#ifdef notdef
         if (bad->type == FAIL_INVARIANT) {
             struct context *inv_ctx = calloc(1, sizeof(struct context) +
                                 MAX_CONTEXT_STACK * sizeof(hvalue_t));
@@ -4282,14 +4192,12 @@ int exec_model_checker(int argc, char **argv){
             inv_ctx->readonly = 1;
 
             hvalue_t args[2];
-            args[0] = node_state(bad->edge->src)->vars;     // TODO.  Seems wrong
+            args[0] = node_state(bad->edge->src)->vars;
             args[1] = node_state(bad->edge->dst)->vars;
             value_ctx_push(inv_ctx, value_put_list(&engine, args, sizeof(args)));
 
             hvalue_t inv_context = value_put_context(&engine, inv_ctx);
 
-            // TODO.  Need to add inv_ctx to the context bag of the state
-
             edge = calloc(1, sizeof(struct edge));
             edge->src = edge->dst = bad->edge->dst;
             edge->ctx = inv_context;
@@ -4347,11 +4255,8 @@ int exec_model_checker(int argc, char **argv){
             global->nprocesses++;
         }
         else {
-#endif
             edge = bad->edge;
-#ifdef notdef
         }
-#endif
 
         // printf("LEN=%u, STEPS=%u\n", bad->edge->dst->len, bad->edge->dst->steps);
 
diff --git a/harmony_model_checker/charm/charm.h b/harmony_model_checker/charm/charm.h
index 8f0adae0..9f333a1a 100644
--- a/harmony_model_checker/charm/charm.h
+++ b/harmony_model_checker/charm/charm.h
@@ -99,6 +99,8 @@ struct global {
     unsigned int last_nstates;      // to measure #states / second
     struct dfa *dfa;                // for tracking correct behaviors
     unsigned int diameter;          // graph diameter
+    bool phase2;                    // when model checking is done and graph analysis starts
+    struct scc *scc_todo;           // SCC search
     struct json_value *pretty;      // for output
     bool run_direct;                // non-model-checked mode
     unsigned long allocated;        // allocated table space
diff --git a/harmony_model_checker/charm/code.c b/harmony_model_checker/charm/code.c
index 33d018a9..bf053759 100644
--- a/harmony_model_checker/charm/code.c
+++ b/harmony_model_checker/charm/code.c
@@ -27,7 +27,6 @@ static struct instr code_instr_parse(struct engine *engine, struct json_value *j
     i.load = strcmp(oi->name, "Load") == 0;
     i.store = strcmp(oi->name, "Store") == 0;
     i.del = strcmp(oi->name, "Del") == 0;
-    i.pause = strcmp(oi->name, "Pause") == 0;
     i.print = strcmp(oi->name, "Print") == 0;
     i.retop = strcmp(oi->name, "Return") == 0;
     i.atomicinc = strcmp(oi->name, "AtomicInc") == 0;
diff --git a/harmony_model_checker/charm/code.h b/harmony_model_checker/charm/code.h
index e469ff83..34f71470 100644
--- a/harmony_model_checker/charm/code.h
+++ b/harmony_model_checker/charm/code.h
@@ -15,13 +15,12 @@ struct instr {
     //  store:       a Store instruction
     //  del:         a Del instruction
     //  retop:       a Return instruction
-    //  pause:       a Pause instruction
     //  print:       a Print instruction
     //  atomicinc:   an AtomicInc instruction
     //  atomicdec:   an AtomicDec instruction
     //  setintlevel: a SetIntlevel instruction
     //  breakable:   a Load, Store, Del, or AtomicInc instruction
-    bool choose, load, store, del, retop, print, pause;
+    bool choose, load, store, del, retop, print;
     bool atomicinc, atomicdec, setintlevel, breakable;
 };
 
diff --git a/harmony_model_checker/charm/graph.c b/harmony_model_checker/charm/graph.c
index c4db374b..c30dd807 100644
--- a/harmony_model_checker/charm/graph.c
+++ b/harmony_model_checker/charm/graph.c
@@ -74,8 +74,6 @@ void graph_check_for_data_race(
     struct node *node,
     struct engine *engine
 ) {
-    struct state *state = node_state(node);
-
     // First check whether any edges conflict with themselves.  That could
     // happen if more than one thread is in the same state and (all) write
     // the same variable
@@ -83,7 +81,7 @@ void graph_check_for_data_race(
         for (struct access_info *ai = edge->ai; ai != NULL; ai = ai->next) {
             if (ai->indices != NULL) {
                 assert(ai->n > 0);
-                if (multiplicities(state)[edge->ctx_index] > 1 && !ai->load && !ai->atomic) {
+                if (edge->multiplicity > 1 && !ai->load && !ai->atomic) {
                     struct failure *f = new_alloc(struct failure);
                     f->type = FAIL_RACE;
                     f->edge = node->u.ph2.u.to_parent;
diff --git a/harmony_model_checker/charm/graph.h b/harmony_model_checker/charm/graph.h
index 2d34002a..f72a343f 100644
--- a/harmony_model_checker/charm/graph.h
+++ b/harmony_model_checker/charm/graph.h
@@ -44,21 +44,18 @@ struct access_info {
 // but, for memory efficiency, not the details of the microsteps themselves.  If
 // a failure is found, that information is recovered by re-executing the path to
 // the faulty state.
-//
-// TODO: various space saving options by replacing contexts with indices.  Also
-//       inv could be an invariant identifier rather than its pc.
 struct edge {
     struct edge *fwdnext;    // forward linked list maintenance
-    hvalue_t choice;         // choice if any (TODO, put in log[0] to save space)
+    hvalue_t ctx, choice;    // ctx that made the microstep, choice if any
     struct node *src;        // source node
     struct node *dst;        // destination node
-    hvalue_t after;          // resulting context (TODO. index in dst context bag)
+    hvalue_t after;          // resulting context
     struct access_info *ai;  // to detect data races
-    int16_t inv;             // pc of invariant (negative if finally clause)
     uint16_t nsteps;         // # microsteps
-    uint8_t ctx_index;       // index of context in src state context bag
+    uint16_t multiplicity;   // multiplicity of context
     bool interrupt : 1;      // set if state change is an interrupt
     // TODO.  Is choosing == (choice != 0)?
+    //        Also, edge->src->choosing is probably the same
     bool choosing : 1;       // destination state is choosing
     bool failed : 1;         // context failed
     uint16_t nlog : 12;      // size of print history
diff --git a/harmony_model_checker/charm/ops.c b/harmony_model_checker/charm/ops.c
index 76c34d97..f028d08d 100644
--- a/harmony_model_checker/charm/ops.c
+++ b/harmony_model_checker/charm/ops.c
@@ -729,11 +729,6 @@ void op_Continue(const void *env, struct state *state, struct step *step, struct
     step->ctx->pc++;
 }
 
-// TODO.  Make the same as Continue?
-void op_Pause(const void *env, struct state *state, struct step *step, struct global *global){
-    step->ctx->pc++;
-}
-
 // This is helper fumction to execute the return of a method call. Usually
 // this would be when a Return instruction is executed, but Charm also
 // supports built-in methods that need to return the same way.  We also
@@ -1379,10 +1374,7 @@ void op_Go(
     struct context *copy = (struct context *) buffer;
     ctx_push(copy, result);
     copy->stopped = false;
-    if (context_add(state, value_put_context(&step->engine, copy)) < 0) {
-        value_ctx_failure(step->ctx, &step->engine, "Go: too many threads");
-        return;
-    }
+    context_add(state, value_put_context(&step->engine, copy));
 #ifdef HEAP_ALLOC
     free(buffer);
 #endif
@@ -2144,7 +2136,7 @@ void op_Spawn(
     }
     else {
         hvalue_t context = value_put_context(&step->engine, ctx);
-        if (context_add(state, context) < 0) {
+        if (!context_add(state, context)) {
             value_ctx_failure(step->ctx, &step->engine, "spawn: too many threads");
             return;
         }
@@ -2560,7 +2552,6 @@ void *init_Choose(struct dict *map, struct engine *engine){ return NULL; }
 void *init_Continue(struct dict *map, struct engine *engine){ return NULL; }
 void *init_Dup(struct dict *map, struct engine *engine){ return NULL; }
 void *init_Go(struct dict *map, struct engine *engine){ return NULL; }
-void *init_Pause(struct dict *map, struct engine *engine){ return NULL; }
 void *init_Print(struct dict *map, struct engine *engine){ return NULL; }
 void *init_Pop(struct dict *map, struct engine *engine){ return NULL; }
 void *init_ReadonlyDec(struct dict *map, struct engine *engine){ return NULL; }
@@ -4458,7 +4449,6 @@ struct op_info op_table[] = {
 	{ "Nary", init_Nary, op_Nary },
 	{ "Pop", init_Pop, op_Pop },
 	{ "Print", init_Print, op_Print, next_Print },
-	{ "Pause", init_Pause, op_Pause },
 	{ "Push", init_Push, op_Push },
 	{ "ReadonlyDec", init_ReadonlyDec, op_ReadonlyDec },
 	{ "ReadonlyInc", init_ReadonlyInc, op_ReadonlyInc },
diff --git a/harmony_model_checker/charm/value.c b/harmony_model_checker/charm/value.c
index 649316e0..48f4e6e2 100644
--- a/harmony_model_checker/charm/value.c
+++ b/harmony_model_checker/charm/value.c
@@ -1608,7 +1608,7 @@ hvalue_t value_bag_add(struct engine *engine, hvalue_t bag, hvalue_t v, int mult
         return value_dict_store(engine, bag, v, count);
     }
     else {
-        return value_dict_store(engine, bag, v, VALUE_TO_INT(multiplicity));
+        return value_dict_store(engine, bag, v, VALUE_TO_INT(1));
     }
 }
 
@@ -1741,31 +1741,25 @@ void context_remove(struct state *state, hvalue_t ctx){
             }
             else {
                 state->bagsize--;
-                memmove(
-                    &state_contexts(state)[i],
-                    &state_contexts(state)[i+1],
-                    (state->bagsize - i) * sizeof(hvalue_t) + i);
-                memmove(
-                    (char *) &state_contexts(state)[state->bagsize] + i,
-                    (char *) &state_contexts(state)[state->bagsize + 1] + i + 1,
-                    state->bagsize - i);
+                memmove(&state_contexts(state)[i], &state_contexts(state)[i+1],
+                        (state->bagsize - i) * sizeof(hvalue_t) + i);
+                memmove((char *) &state_contexts(state)[state->bagsize] + i,
+                        (char *) &state_contexts(state)[state->bagsize + 1] + i + 1,
+                        state->bagsize - i);
             }
-            return;
+            break;
         }
     }
-    printf("--> %p\n", (void *) ctx);
-    panic("context_remove: no such context");
 }
 
-// Add context 'ctx' to the state's context bag.  Return the index into the
-// context bag.  Return -1 if there are too  many different contexts
-// (i.e., >= MAX_CONTEXT_BAG).
-int context_add(struct state *state, hvalue_t ctx){
-    int i;
+// Add context 'ctx' to the state's context bag.  May fail if there are too
+// many different contexts (i.e., >= MAX_CONTEXT_BAG).
+bool context_add(struct state *state, hvalue_t ctx){
+    unsigned int i;
     for (i = 0; i < state->bagsize; i++) {
         if (state_contexts(state)[i] == ctx) {
             multiplicities(state)[i]++;
-            return i;
+            return true;
         }
         if (state_contexts(state)[i] > ctx) {
             break;
@@ -1773,7 +1767,7 @@ int context_add(struct state *state, hvalue_t ctx){
     }
 
     if (state->bagsize >= MAX_CONTEXT_BAG) {
-        return -1;
+        return false;
     }
  
     // Move the last multiplicities
@@ -1788,5 +1782,5 @@ int context_add(struct state *state, hvalue_t ctx){
     state->bagsize++;
     state_contexts(state)[i] = ctx;
     multiplicities(state)[i] = 1;
-    return i;
+    return true;
 }
diff --git a/harmony_model_checker/charm/value.h b/harmony_model_checker/charm/value.h
index 8e19b72c..89dfb225 100644
--- a/harmony_model_checker/charm/value.h
+++ b/harmony_model_checker/charm/value.h
@@ -7,25 +7,26 @@
 #include "charm.h"
 
 #define MAX_CONTEXT_STACK   250        // maximum size of context stack
-#define MAX_CONTEXT_BAG      32        // maximum number of distinct contexts
+#define MAX_CONTEXT_BAG       32        // maximum number of distinct contexts
 
 // This contains the state in a Harmony execution.
 //
 // TODO.  State can be reduced in size in various ways;
 //  - pre and stopbag are not always used
+//  - choosing could be an index into the context bag
 //  - the entire thing could be replaced with a collision-resistant hash
 typedef struct state {
     hvalue_t vars;        // shared variables
     hvalue_t pre;         // "pre" state (same as vars in non-choosing states)
+    hvalue_t choosing;    // context that is choosing if non-zero
     hvalue_t stopbag;     // bag of stopped contexts (to detect deadlock)
     uint32_t dfa_state;   // state of input dfa
     uint16_t tid_gen;     // thread id generator
-    int8_t chooser;       // -1 if not a choosing state, otherwise indexes into contexts
 
     // The state includes a variable-size bag of contexts.  This is represented
     // by an array of contexts of type hvalue_t, which is followed by an array
     // of multiplicities (of type uint8_t) with the same number of elements.
-    uint8_t bagsize;
+    uint16_t bagsize;
     // hvalue_t contexts[VAR_SIZE];   // context/multiplicity pairs
 } state;
 #define state_contexts(s)   ((hvalue_t *) ((s) + 1))
@@ -141,9 +142,8 @@ void value_ctx_extend(struct context *ctx);
 hvalue_t value_ctx_failure(struct context *ctx, struct engine *engine, char *fmt, ...);
 bool value_ctx_all_eternal(hvalue_t ctxbag);
 bool value_state_all_eternal(struct state *state);
-void context_remove_by_index(struct state *state, int ctx_index);
 void context_remove(struct state *state, hvalue_t ctx);
-int context_add(struct state *state, hvalue_t ctx);
+bool context_add(struct state *state, hvalue_t ctx);
 char *json_escape_value(hvalue_t v);
 void value_trace(struct global *global, FILE *file, struct callstack *cs, unsigned int pc, hvalue_t vars, char *prefix);
 void print_vars(struct global *global, FILE *file, hvalue_t v);
diff --git a/harmony_model_checker/compile.py b/harmony_model_checker/compile.py
index 02fba1c7..4e7f4277 100644
--- a/harmony_model_checker/compile.py
+++ b/harmony_model_checker/compile.py
@@ -5,8 +5,8 @@
 
 import harmony_model_checker.harmony.harmony as legacy_harmony
 from harmony_model_checker.harmony.harmony import Scope, Code, State
-from harmony_model_checker.harmony.ops import FrameOp, ReturnOp, PauseOp, JumpOp
-from harmony_model_checker.harmony.value import AddressValue, ContextValue, LabelValue
+from harmony_model_checker.harmony.ops import FrameOp, ReturnOp
+from harmony_model_checker.harmony.value import AddressValue, ContextValue
 from harmony_model_checker.parser.antlr_rule_visitor import HarmonyVisitorImpl
 from harmony_model_checker.parser.HarmonyParser import HarmonyParser
 from harmony_model_checker.parser.HarmonyErrorListener import HarmonyLexerErrorListener, HarmonyParserErrorListener
@@ -117,10 +117,6 @@ def _load_file(filename: str, scope: Scope, code: Code, init: bool):
     ast.compile(scope, code, None)
     if init:
         _, file, line, column = ast.token
-        invlabel = LabelValue(None, "invariant")
-        code.nextLabel(invlabel)
-        code.append(PauseOp(), ast.token, ast.endtoken, stmt=(line1, column1, line2, column2))
-        code.append(JumpOp(invlabel, reason="check invariants again"), ast.token, ast.token, stmt=(line1, column1, line2, column2))
         code.append(ReturnOp(("result", file, line, column), AddressValue(None, [])), ast.token, ast.endtoken, stmt=(line1, column1, line2, column2))  # to terminate "__init__" process
     legacy_harmony.namestack.pop()
 
diff --git a/harmony_model_checker/harmony/ops.py b/harmony_model_checker/harmony/ops.py
index 2ef15f9b..144b1ba7 100644
--- a/harmony_model_checker/harmony/ops.py
+++ b/harmony_model_checker/harmony/ops.py
@@ -668,23 +668,6 @@ def tladump(self):
     def eval(self, state, context):
         context.pc += 1
 
-# TODO. Maybe can be the same as ContinueOp
-class PauseOp(Op):
-    def __repr__(self):
-        return "Pause"
-
-    def explain(self):
-        return "yield to another thread"
-
-    def jdump(self):
-        return '{ "op": "Pause" }'
-
-    def tladump(self):
-        return 'OpPause(self)'
-
-    def eval(self, state, context):
-        context.pc += 1
-
 class StoreVarOp(Op):
     def __init__(self, v, lvar=None, reason=None):
         self.v = v
@@ -968,7 +951,7 @@ def __repr__(self):
         (lexeme, file, line, column) = self.result
         if self.default is None:
             return "ReturnOp(%s)"%lexeme
-        return "ReturnOp(%s, %s)"%(lexeme, strValue(self.default))
+        return "ReturnOp(%s. %s)"%(lexeme, strValue(self.default))
 
     def jdump(self):
         if self.result is None: