diff --git a/dev/sim/coco/A2L2.py b/dev/sim/coco/A2L2.py
index dedd779..8226373 100755
--- a/dev/sim/coco/A2L2.py
+++ b/dev/sim/coco/A2L2.py
@@ -9,6 +9,8 @@ from cocotb.handle import Release
 from dotmap import DotMap
 import itertools
 
+from OPEnv import *
+
 # ------------------------------------------------------------------------------------------------
 # Classes
 
@@ -100,7 +102,8 @@ class A2L2Trans(DotMap):
       if self.load:
          self.addr = self.addr & 0xFFFFFFF0
       elif self.store:
-         self.addr = self.addr & 0xFFFFFFE0
+         #self.addr = self.addr & 0xFFFFFFE0 #wtf definitely 16B-aligned occurring
+         self.addr = self.addr & 0xFFFFFFF0
          if self.be == None or self.data == None:
             raise Exception('A2L2Trans: store must have BE and data')
          else:
@@ -196,16 +199,6 @@ class A2L2Trans(DotMap):
       else:
             raise Exception(f'A2L2Trans: unsupported store len={self.len}')
 
-
-# ------------------------------------------------------------------------------------------------
-# Functions
-
-def hex(n, pad=0):
-   if pad:
-      return f'000000000000000000000000{n.value.hex()[2:].upper()}'[-pad:]
-   else:
-      return n.value.hex()[2:].upper()
-
 # ------------------------------------------------------------------------------------------------
 # Tasks
 
diff --git a/dev/sim/coco/A2O.py b/dev/sim/coco/A2O.py
new file mode 100755
index 0000000..e54e865
--- /dev/null
+++ b/dev/sim/coco/A2O.py
@@ -0,0 +1,201 @@
+# A2O Core
+
+import cocotb
+from cocotb.triggers import Timer, RisingEdge
+from cocotb.binary import BinaryValue
+from cocotb.handle import Force
+from cocotb.handle import Release
+
+from dotmap import DotMap
+import itertools
+
+from OPEnv import *
+
+# ------------------------------------------------------------------------------------------------
+# Tasks
+
+async def A2ODriver(dut, sim):
+   """A2O Core Driver"""
+
+   transTypes = {
+      '00': 'IFETCH',
+      '08': 'LOAD',
+      '20': 'STORE'
+   }
+
+   ok = True
+   readPending = []
+   countReads = 0
+   mem = {}
+   sim.msg('A2O Driver: nothing to do.')
+
+   #while ok and not sim.done:
+   #   await RisingEdge(dut.clk_1x)
+
+
+# A2O Checker
+# check protocol, etc.
+async def A2OChecker(dut, sim):
+   """A2O Core Checker """
+
+   me = 'A2O Checker'
+   ok = True
+   sim.msg(f'{me}: started.')
+
+   # errors
+   creditsLdErr = dut.c0.lq0.lsq.arb.ld_cred_err_q
+   creditsStErr = dut.c0.lq0.lsq.arb.st_cred_err_q
+   errors = [
+      {'name': 'Load Credits', 'sig': creditsLdErr},
+      {'name': 'Store Credits', 'sig': creditsStErr},
+   ]
+
+   while ok:
+
+      await RisingEdge(dut.clk_1x)
+
+      if not sim.resetDone:
+         continue
+
+      for i in range(len(errors)):
+          assert errors[i]['sig'].value == 0, f'{me} Error: {errors[i]["name"]}'
+
+# A2O Monitor
+# count transactions, etc.
+# fail on bad addresses
+async def A2OMonitor(dut, sim):
+   """A2O Core Monitor"""
+
+   me = 'A2O Monitor'
+   ok = True
+   sim.msg(f'{me}: started.')
+
+   # completions
+   iu0Comp = dut.c0.iu_lq_i0_completed
+   iu0CompIFAR = dut.c0.iuq0.iuq_cpl_top0.iuq_cpl0.cp2_i0_ifar
+   iu1Comp = dut.c0.iu_lq_i1_completed
+   iu1CompIFAR = dut.c0.iuq0.iuq_cpl_top0.iuq_cpl0.cp2_i1_ifar
+   iuCompFlushIFAR = dut.c0.cp_t0_flush_ifar
+   cp3NIA = dut.c0.iuq0.iuq_cpl_top0.iuq_cpl0.iuq_cpl_ctrl.cp3_nia_q           # nia after last cycle's completions
+
+   # GPR ppol and arch map
+   gprCompMap = []
+   lastGprCompMap = []
+   #wtf check what 33:36 are!
+   for i in range(36):
+      gprCompMap.append(dut.c0.iuq0.iuq_slice_top0.slice0.rn_top0.fx_rn0.gpr_rn_map.xhdl3.comp_map0[i].comp_map_latch.dout)
+      lastGprCompMap.append(i)
+
+   gpr = []
+   for i in range(144):
+      gpr.append(dut.c0.xu0.gpr.gpr0.loc[i].dat)
+
+   # CR fields pool and arch map
+   crCompMap = []
+   lastCrCompMap = []
+   for i in range(8):
+      crCompMap.append(dut.c0.iuq0.iuq_slice_top0.slice0.rn_top0.fx_rn0.cr_rn_map.xhdl3.comp_map0[i].comp_map_latch.dout)
+      lastCrCompMap.append(i)
+
+   cr = []
+   for i in range(24):
+      cr.append(dut.c0.xu0.cr.entry[i].reg_latch.dout)
+
+   # XER pool and arch map
+   xerCompMap = []
+   lastXerCompMap = []
+   for i in range(1):
+      xerCompMap.append(dut.c0.iuq0.iuq_slice_top0.slice0.rn_top0.fx_rn0.xer_rn_map.xhdl3.comp_map0[i].comp_map_latch.dout)
+      lastXerCompMap.append(i)
+
+   xer = []
+   for i in range(12):
+      xer.append(dut.c0.xu0.xer.entry[i].reg_latch.dout)
+
+   # CTR pool and arch map
+   ctrCompMap = []
+   lastCtrCompMap = []
+   for i in range(1):
+      ctrCompMap.append(dut.c0.iuq0.iuq_slice_top0.slice0.rn_top0.fx_rn0.ctr_rn_map.xhdl3.comp_map0[i].comp_map_latch.dout)
+      lastCtrCompMap.append(i)
+
+   ctr = []
+   for i in range(8):
+      ctr.append(dut.c0.xu0.ctr.entry[i].reg_latch.dout)
+
+   # LR pool and arch map
+   lrCompMap = []
+   lastLrCompMap = []
+   for i in range(1):
+      lrCompMap.append(dut.c0.iuq0.iuq_slice_top0.slice0.rn_top0.fx_rn0.lr_rn_map.xhdl3.comp_map0[i].comp_map_latch.dout)
+      lastLrCompMap.append(i)
+
+   lr = []
+   for i in range(8):
+      lr.append(dut.c0.xu0.lr.entry[i].reg_latch.dout)
+
+   while ok:
+
+      await RisingEdge(dut.clk_1x)
+
+      if not sim.resetDone:
+         continue
+
+      comp = ''
+      #wtf seeing something weird here
+      # there are cases where x's are in some bits of comp ifar's; maybe ok (predict array?) but why is completed indicated?
+      if iu0Comp.value == 1:
+         comp = f'0:{sim.safeint(iu0CompIFAR.value.binstr + "00", 2):06X} '
+
+      if iu1Comp.value == 1:
+         comp = f'{comp}1:{sim.safeint(iu1CompIFAR.value.binstr + "00", 2):06X} '
+
+      if comp != '':
+         comp = f'{comp}{sim.safeint(iuCompFlushIFAR.value.binstr + "00", 2):016X}'
+         sim.msg(f'C0: CP {comp}')
+
+      if sim.a2o.traceFacUpdates:
+         # renamables
+         for i in range(36):
+            good, arch = sim.safeint(gprCompMap[i].value.binstr, 2, rc=True)
+            if good and arch != lastGprCompMap[i]:
+               sim.msg(f'GPR Update: R{i:02d}={hex(gpr[arch], 16)}')
+               lastGprCompMap[i] = arch
+         for i in range(8):
+            good, arch = sim.safeint(crCompMap[i].value.binstr, 2, rc=True)
+            if good and arch != lastCrCompMap[i]:
+               sim.msg(f'CR Update: F{i:01d}={hex(cr[arch], 1)}')
+               lastCrCompMap[i] = arch
+         for i in range(1):
+            good, arch = sim.safeint(xerCompMap[i].value.binstr, 2, rc=True)
+            if good and arch != lastXerCompMap[i]:
+               v = xer[arch].value.binstr
+               sim.msg(f'XER Update: SO/OV/CA={v[0:3]} LEN={int(v[3:],2):02X}')
+               lastXerCompMap[i] = arch
+         for i in range(1):
+            good, arch = sim.safeint(ctrCompMap[i].value.binstr, 2, rc=True)
+            if good and arch != lastCtrCompMap[i]:
+               sim.msg(f'CTR Update:{hex(ctr[arch], 16)}')
+               lastCtrCompMap[i] = arch
+         for i in range(1):
+            good, arch = sim.safeint(lrCompMap[i].value.binstr, 2, rc=True)
+            if good and arch != lastLrCompMap[i]:
+               sim.msg(f'LR Update:{hex(lr[arch], 16)}')
+               lastLrCompMap[i] = arch
+
+# ------------------------------------------------------------------------------------------------
+# Classes
+
+class A2O:
+   driver = A2ODriver
+   checker = A2OChecker
+   monitor = A2OMonitor
+
+   def __init__(self, sim):
+      pass
+
+class A2OCore(DotMap):
+   def __init__(self, sim):
+      super().__init__()
+      self.sim = sim
+      self.traceFacUpdates = False
\ No newline at end of file
diff --git a/dev/sim/coco/OPEnv.py b/dev/sim/coco/OPEnv.py
index fc1db51..ccbc145 100755
--- a/dev/sim/coco/OPEnv.py
+++ b/dev/sim/coco/OPEnv.py
@@ -18,6 +18,8 @@ class Sim(DotMap):
    def __init__(self, dut, cfg=None):
       super().__init__()
       self.dut = dut
+      self.resetDone = False
+
       # defaults
       self.memFiles = ['../mem/boot_ieq1.bin.hex'] #wtf cmdline parm
       self.threads = None
@@ -30,10 +32,13 @@ class Sim(DotMap):
       self.maxCycles = 500
       self.memFiles = None
       self.config = DotMap()
+      # these should be in a2 core class
       self.config.core = DotMap({
          'creditsLd': None,
          'creditsSt': None,
-         'creditsLdStSingle': False
+         'creditsLdStSingle': None,
+         'lsDataForward' : None,
+         'cpcr4_sq_cnt' : None
       })
       self.config.a2l2 = DotMap({
          'badAddr': [('E0','E0', 'IRW')]
@@ -56,6 +61,25 @@ class Sim(DotMap):
             self.threads = 1
          self.msg(f'Set threads={self.threads}.')
 
+   # helpers
+   # wtf: assertFail doesn't work??#@#!!
+   def safeint(self, v, n=10, assertFail=True, rc=False):
+      try:
+         res = int(v, n)
+         ok = True
+      except:
+         self.ok = False
+         self.fail = f'Bad integer conversion: {v},{n}'
+         if assertFail:
+            assert(False, self.fail)
+         res = 0
+         ok = False
+      if rc:
+         return (ok, res)
+      else:
+         return res
+
+
 class TransQ(DotMap):
    def __init__(self):
       super().__init__()
@@ -115,3 +139,12 @@ class Memory(DotMap):
          else:
             self.sim.msg(f'Mem Update: @{addr:08X} {self.data[addr]:08X}->{data:08X}')
       self.data[addr] = data
+
+# ------------------------------------------------------------------------------------------------
+# Functions
+
+def hex(n, pad=0):
+   if pad:
+      return f'000000000000000000000000{n.value.hex()[2:].upper()}'[-pad:]
+   else:
+      return n.value.hex()[2:].upper()
diff --git a/dev/sim/coco/tb.py b/dev/sim/coco/tb.py
index 9e66f9a..33b8896 100755
--- a/dev/sim/coco/tb.py
+++ b/dev/sim/coco/tb.py
@@ -11,6 +11,7 @@ import itertools
 from dotmap import DotMap
 
 from OPEnv import *
+from A2O import *
 from A2L2 import *
 
 # ------------------------------------------------------------------------------------------------
@@ -92,9 +93,17 @@ async def config(dut, sim):
    creditsLdMax = dut.c0.lq0.lsq.arb.ld_cred_max            # hdw check
    creditsSt = dut.c0.lq0.lsq.arb.store_cred_cnt_d          # 32 max
    creditsStMax = dut.c0.lq0.lsq.arb.st_cred_max            # hdw check
-   creditsLdStSingle = dut.c0.lq0.lsq.arb.spr_xucr0_cred_d.value   # 1 total credit
+   creditsLdStSingle = dut.c0.lq0.lsq.arb.spr_xucr0_cred_d  # 1 total credit
    #wtf this affects A2L2 - default=1
-   #creditsLdStSingle = dut.c0.lq0.lsq.arb.spr_lsucr0_b2b_q.value   # 0=crit first, every other 1=crit first, b2b **the a2l2 spec does not say crit must be first**
+   #  dut.c0.lq0.lsq.arb.spr_lsucr0_b2b_q   # 0=crit first, every other 1=crit first, b2b **the a2l2 spec does not say crit must be first**
+   lsucr0_d = dut.c0.lq0.ctl.spr.lq_spr_cspr.lsucr0_d
+   lsucr0_q = dut.c0.lq0.ctl.spr.lq_spr_cspr.lsucr0_q
+   cpcr2_d = dut.c0.iuq0.iuq_ifetch0.iuq_spr0.cpcr2_d
+   cpcr2_q = dut.c0.iuq0.iuq_ifetch0.iuq_spr0.cpcr2_l2
+   cpcr2_act = dut.c0.iuq0.iuq_ifetch0.iuq_spr0.cpcr2_wren
+   cpcr4_d = dut.c0.iuq0.iuq_ifetch0.iuq_spr0.cpcr4_d
+   cpcr4_q = dut.c0.iuq0.iuq_ifetch0.iuq_spr0.cpcr4_l2
+   cpcr4_act = dut.c0.iuq0.iuq_ifetch0.iuq_spr0.cpcr4_wren
 
    await RisingEdge(dut.clk_1x)
 
@@ -112,58 +121,41 @@ async def config(dut, sim):
       await RisingEdge(dut.clk_1x)
       creditsSt.value = Release()
 
-   if sim.config.core.creditsLdStSingle:
-      creditsLdStSingle = Force(1)
+   if sim.config.core.creditsLdStSingle is not None:
+      v = 1 if sim.config.core.creditsLdStSingle else 0
+      creditsLdStSingle.value = Force(v)
       sim.msg(f'A2L2: only one load OR store allowed when credits=1/1.')
       await RisingEdge(dut.clk_1x)
       creditsLdStSingle.value = Release()
 
-   await RisingEdge(dut.clk_1x)
-
-async def coreMonitor(dut, sim):
-   """Watch for core events. """
-
-   me = 'a2oMonitor'
-
-   # errors
-   creditsLdErr = dut.c0.lq0.lsq.arb.ld_cred_err_q
-   creditsStErr = dut.c0.lq0.lsq.arb.st_cred_err_q
-
-   # watches
-   iu0Comp = dut.c0.iu_lq_i0_completed
-   iu0CompIFAR = dut.c0.iuq0.iuq_cpl_top0.iuq_cpl0.cp2_i0_ifar
-   iu1Comp = dut.c0.iu_lq_i1_completed
-   iu1CompIFAR = dut.c0.iuq0.iuq_cpl_top0.iuq_cpl0.cp2_i1_ifar
-   iuCompFlushIFAR = dut.c0.cp_t0_flush_ifar
-   cp3NIA = dut.c0.iuq0.iuq_cpl_top0.iuq_cpl0.iuq_cpl_ctrl.cp3_nia_q           # nia after last cycle's completions
-
-   # queue depths, etc.
-
-   errors = [
-      {'name': 'Load Credits', 'sig': creditsLdErr},
-      {'name': 'Store Credits', 'sig': creditsStErr},
-   ]
-
-   done = False
-
-   while not done:
-
+   #wtf make a function - needs mask,thread
+   if sim.config.core.lsDataForward is not None:
+      v = 1 if sim.config.core.lsDataForward else 0
+      sim.msg(f'LSUCR0 = {hex(lsucr0_q.value), 8}')
+      sim.msg(f'Setting LSUCR0[DFWD] = {v}.')
+      v = v << 2
+      v = (lsucr0_q.value.integer & ~0x4) | v
+      lsucr0_d.value = Force(v)
       await RisingEdge(dut.clk_1x)
+      lsucr0_d.value = Release()
+      sim.msg(f'LSUCR0 = {hex(lsucr0_q.value), 8}')
+
+   if sim.config.core.cpcr4_sq_cnt is not None:
+      v = sim.config.core.cpcr4_sq_cnt
+      sim.msg(f'CPCR4 = {hex(cpcr4_q[0], 8)}')
+      sim.msg(f'Setting CPCR4[SQ_CNT] = {v}.')
+      v = v << 0
+      v = (cpcr4_q[0].value.integer & ~0x1F) | v
+      await RisingEdge(dut.clk_1x)  # need cuz of act?
+      cpcr4_d[0].value = Force(v)
+      cpcr4_act.value = Force(1)
+      await RisingEdge(dut.clk_1x)
+      await RisingEdge(dut.clk_1x)  # need cuz of act?
+      cpcr4_d[0].value = Release()
+      cpcr4_act.value = Release()
+      sim.msg(f'CPCR4 = {hex(cpcr4_q[0], 8)}')
 
-      for i in range(len(errors)):
-          assert errors[i]['sig'].value == 0, f'{me} Error: {errors[i]["name"]}'
-
-      comp = ''
-      if iu0Comp.value == 1:
-         comp = f'0:{int(iu0CompIFAR.value.binstr + "00", 2):06X} '
-
-      if iu1Comp.value == 1:
-         comp = f'{comp}1:{int(iu1CompIFAR.value.binstr + "00", 2):06X} '
-
-      if comp != '':
-         comp = f'{comp}{int(iuCompFlushIFAR.value.binstr + "00", 2):016X}'
-         sim.msg(f'C0: CP {comp}')
-
+   await RisingEdge(dut.clk_1x)
 
 # trilib/tri.vh:`define  NCLK_WIDTH  6   // 0  1xClk, 1  Reset, 2  2xClk, 3  4xClk,  4  Even .5xClk,  5 Odd .5xClk
 async def genReset(dut, sim):
@@ -183,6 +175,7 @@ async def genReset(dut, sim):
          dut._log.info(f'[{sim.cycle:08d}] Releasing reset.')
          dut.nclk[1].value = 0
          done = True
+         sim.resetDone = True
 
 async def genClocks(dut, sim):
    """Generate 1x, 2x, 4x clock pulses, depending on parms. """
@@ -237,15 +230,22 @@ async def tb(dut):
 
    sim = Sim(dut)
    sim.mem = Memory(sim)
-   sim.maxCycles = 2000
+   sim.maxCycles = 9000
+   # original fpga design needed 4 cred, no fwd (set in logic currently)
+   #sim.config.core.creditsSt = 32
+   #sim.config.core.lsDataForward = 0   # disable=1
+   #sim.config.core.cpcr4_sq_cnt = 0     # default=6
 
    '''
+   # rom
    sim.memFiles = ['../mem/boot.bin.hex'] #wtf cmdline parm
 
    for i in range(len(sim.memFiles)):  #wtf el should be object with name, format, etc.
       sim.mem.loadFile(sim.memFiles[i])
    '''
 
+   '''
+   # rom+test; should end at 700
    sim.memFiles = [
       {
       'addr': 0x00000000,
@@ -255,7 +255,26 @@ async def tb(dut):
       'addr': 0x10000000,
       'file' : '../mem/test1/test.init'
       }
-      ]
+   ]
+   '''
+   '''
+   # rom+bios; should end at 7FC
+   sim.memFiles = [
+      {
+      'addr': 0x00000000,
+      'file' : '../mem/test2/rom.init'
+      }
+   ]
+   '''
+
+
+   # rom+bios+arcitst
+   sim.memFiles = [
+      {
+      'addr': 0x00000000,
+      'file' : '../mem/test3/rom.init'
+      }
+   ]
 
    for i in range(len(sim.memFiles)):  #wtf el should be object with name, format, etc.
       sim.mem.loadFile(sim.memFiles[i]['file'], addr=sim.memFiles[i]['addr'])
@@ -274,10 +293,12 @@ async def tb(dut):
    # start interfaces
    await cocotb.start(scom(dut, sim))
 
-   #wtf don't have to instantiate A2L2 first?
-   #await cocotb.start(A2L2Driver(dut, sim))
-   #await cocotb.start(A2L2Checker(dut, sim))
-   #await cocotb.start(A2L2Monitor(dut, sim))
+   sim.a2o = A2OCore(sim)
+   sim.a2o.traceFacUpdates =  True
+   await cocotb.start(A2O.driver(dut, sim))
+   await cocotb.start(A2O.checker(dut, sim))
+   await cocotb.start(A2O.monitor(dut, sim))
+
    await cocotb.start(A2L2.driver(dut, sim))
    await cocotb.start(A2L2.checker(dut, sim))
    await cocotb.start(A2L2.monitor(dut, sim))
@@ -291,7 +312,7 @@ async def tb(dut):
    await config(dut, sim)
 
    # monitor stuff
-   await cocotb.start(coreMonitor(dut, sim))
+   #await cocotb.start(coreMonitor(dut, sim))
 
    # release thread(s)
    dut.an_ac_pm_thread_stop.value = 0