He hears, thinks, then responds.
Bishoplistening
- Hearsspeech becomes text, per friend
- Thinksthe AI decides: is this a moment for me?stay quiettalk
- Speakstext becomes his voice
“Tess wins. Rowan, sit down.”
He has a whole day of his own.
- 09:50walk to library, refill water
- 10:15half-reading notes, phone winning
- 11:45chicken wrap and cold brew
- 12:20drive home
- 12:45quick nap
- 13:10laundry, tidy kitchen
- 13:40gas station run
a new day planned for every date, on the real clock · 17 Sep 2026
He remembers you. Weeks later.
“the exam thing”
a few weeks ago · afternoon
The MCAT she wouldn’t stop worrying about
Tess “if I bomb this I’m dropping out of pre-med lol”
- Bishop → Tess · 6 days laterhey wait, did you ever hear back on the MCAT?
His memory comes in nine parts.
Systems
- Self-canon: who he is, to himself
- Hard facts: a locked biography, owner-approved
- Nightly digest: the day, written up as stories
- Live transcript: the last 15 minutes, word for word
Reading people, 5 layers
- Favor: a number for how he feels about you
- Impressions: things he noticed, each with a quote
- Traits: claims that firm up with evidence
- Dyads: how two OTHER people get along
- Profiles: his summary of you, in his own words
While you sleep, he writes the night down.
- Tess if I bomb this I’m dropping out of pre-med lol
- Rowan you’ve said that before every exam
- Bishop when is it?
The MCAT she wouldn’t stop worrying about
Tess “if I bomb this I’m dropping out of pre-med lol”
at 4am, while everyone is asleep · 726 conversations written up across 105 days · 17 Sep 2026 · the lines are invented
He reads how you two click.
- he’d expect: trash talk about last night’s gamepredictable
- would surprise him: Juniper asking for real adviceunlikely
Mood, energy, and how much he shares.
Something real happened, so this one holds for hours instead of the usual 55 minutes.
with Rowan · close
“ngl that lag spike pissed me off”
with someone he just met
keeps it to himselfillustrative. How close you are decides what he shares, never whether he helps you
One person built all of it.
56,092lines of Python code
105days of his own simulated life
73.4%of the times he could act on his own, he picks: do nothing
counted September 2026
Want the details?
the numbers, the code, and what I don’t claimBishop’s real admin panel shows private conversations, so it sits behind a login. This is its interface rebuilt for this page: the same sky, layout and type, with stand-in names (Rowan, Juniper, Tess, Marcus, Devin). The call, the memory moment, his opinions of the crew and the mood moment are invented; his day, the numbers and the code are real.
Every number on this page, with the date it was counted
- 825 → 792 → 33 moments past every clock throttle → blocked by reading the room → he spoke
- 17 Sep 2026
- 13 named stages in the engage gate
- 17 Sep 2026
- 2–5 a day times he stops to think, in the gaps of his day
- 17 Sep 2026
- 73.4% of those thoughts end in “nothing”, out of 18 decisions he can make
- 17 Sep 2026
- 105 days of generated daily life, one plan per date
- 17 Sep 2026
- 9 memory stores (4 systems + 5 layers of people)
- 17 Sep 2026
- 14 · 12 command verbs · the ones that have actually fired
- 17 Sep 2026
- 15 behavioural metrics on control charts
- 26 Aug 2026
- 62 · 56,092 Python modules · lines (wc -l)
- 22 Sep 2026
- 2,235 / 2,235 self-test checks passing at boot, 67 suites, ~10.3 s, local run
- 22 Sep 2026
- 11 / 17 background loops / slash commands
- 22 Sep 2026
- ~90 named prompt layers, ~50,000 characters
- 22 Sep 2026
- 14 tabs in the real admin panel
- 22 Sep 2026
- 726 · 13 · 105 conversations · people · days, one box
- 17 Sep 2026
The self-test runs as a plain Python process at boot. Line counts are raw wc -l.
What measuring found: seven problems, none of which crashed
- 2,606 ms
A 1.7 µs lookup froze him for 2,606 ms: one shared SQLite connection stalled the event loop. 4,000 reads under write load: 12.813 s blocked, 0.004 s with a connection per thread (~3,200× less). Its link to a routing gap is a hypothesis with good evidence, not a proof. BISHOP_AUDIT_2.md:4749-4776 · 17 Sep 2026
- 82%
of what he recalled was about his own name: 289 of 351 lines. The search matched his name, not the topic. Fixed on the query side. BISHOP_AUDIT.md:1440-1441 · 26 Aug 2026
- 12×
A text-similarity check got an exact upper bound: 10,461 of 10,464 pairs rejected before any real work, 945 → 79 ms p50. Measured in isolation; the effect in production is unmeasured. BISHOP_AUDIT_2.md:4742-4745 · 16 Sep 2026
- 3 → 8
A yes/no gate’s token budget. The next model returned nothing at 3, three draws of three, and nothing would have errored. router.py · 22 Sep 2026
- def, def
The weekly audit defined one check twice; Python silently kept the second. A self-test now forbids it. selftest.py
- 0 / 15
“0 entries” was fifteen failed queries: 67 rows still named a removed person. Every scan now reports its success count. BISHOP_AUDIT_2.md:4791-4795 · 17 Sep 2026
- 17%
of live turns misfiled by an unregistered prompt layer, live from 31 Jul, found 21 Sep 2026. trace_bishop.py:1343-1358
One sentence, from separate microphones to a reply
- Voice packets
- A buffer per speaker
- A recognizer per speaker
- Turn detection
- Coalescer
- Engage gate
- Planner
- Feeders (6 s budget)
- Prompt assembly
- Reply + guards
- Speech
Most sentences stop at the engage gate: 13 named stages, cheapest first. BISHOP_AUDIT_2.md:126-153
# A ONE-WORD GATE STILL NEEDS HEADROOM, AND 3 IS NOT IT ON EVERY MODEL.
# These are YES/NO gates, so `max_tokens=3` looked obviously sufficient and was, on
# gemini-2.5-flash-lite. Measured Sep 11 against its named successor, 3 draws per budget:
#
# gemini-3.1-flash-lite mt=3 ['<EMPTY>','<EMPTY>','<EMPTY>'] mt=5 ['YES','YES','YES']
# gemini-2.5-flash-lite mt=3 ['NO','NO','NO'] mt=5 ['NO','NO','NO']
#
# It is NOT reasoning tokens eating the budget (`reasoning_tokens` is 0 on both) -- 3.1 simply
# returns nothing that small. Every one of these gates falls back to a default on an empty
# reply, so the whole failure would have been three gates silently answering "default" forever,
# with no error anywhere. `gemini-2.5-flash-lite` retires Oct 16 2026, so this WILL be swapped.
# 8 is the same price in practice (a gate emits 1 token) and survives the next model too.
_GATE_TOKENS = int(os.getenv("BISHOP_GATE_TOKENS", "8"))def _sigma(vals):
"""Spread from the mean MOVING RANGE, not the standard deviation.
This is the individuals-chart estimator, and here it earns its keep twice over. The stated risk
for this whole feature is a baseline drawn from an already-drifted stretch, which bakes the
drift in -- and a standard deviation absorbs any trend inside the baseline into the spread, so
the limits get widest exactly when behaviour was already sliding. Point-to-point variation
doesn't see a slow slide, so it stays honest about how much this number normally jumps."""
vals = [v for v in vals if v is not None]
if len(vals) < 2:
return None
mr = [abs(b - a) for a, b in zip(vals, vals[1:])]
s = (sum(mr) / len(mr)) / 1.128 # d2 for n=2
if s > 0:
return s
try:
return _stat.pstdev(vals) or None # identical neighbours but a spread overall
except Exception:
return NoneEXECUTABLE_COMMANDS = ("KICK", "TEXT", "SAYCALL", "JOINCALL", "LEAVE", "DRAG",
"TIMER", "REMIND", "CHOOSE", "MUSIC", "GAME", "IMAGE", "PLAN", "QUEUE")What I don’t claim
- Scale: 726 conversations, 13 people, 105 days, one box. Friends, not users.
- Four model roles in code; two models in the live container.
- No latency figure, because none is current.
- Psychology papers are design sources, not validation.
Measured, and not built
- A sarcasm detector: its best signal also describes ordinary speech.
- A memory fix: it passed offline, then got worse against the real index. Reverted.
BISHOP_AUDIT_2.md:4517-4572 · 17 Sep 2026
Stack
One Python process, in Docker, on one Linux box.
py-cordDeepgram FluxOpenRoutermem0 + embedded QdrantElevenLabsTavilySQLite (WAL)Dockerasyncio
The panel itself
FastAPI + React 19 + Tailwind 4 + Vite, 14 tabs, read-only mounts of his data. Private, behind Cloudflare Access.
Proves real-time multi-speaker voice pipeline, nine-store memory, observability for a box with no interactive debugging (résumé → Bishop).
