leaseNextJob hands back any pending replay_article job, it has no idea about
runs, and the worker was attributing whatever came back to whichever run was
active. One recovered dead letter from run 1 would have been stamped with run
2's id, given run 2's feedback brief, and dragged run 2's cursor to wherever
that old article sits in the archive. A pinned run would then decide its set was
finished after a couple of articles. There are 139 dead letters and they are
built to recover, so this was not hypothetical.
The idempotency key already says which run enqueued the job. Ask it.
Also: refuse to inherit the parent's model label when starting a run. Inheriting
is exactly how run 1 came to be labelled qwen for predictions deepseek made.
The split moves to the replay container's actual restart time rather than the
commit timestamp five minutes later. Verified the running container really does
have the instrument rules, the de-anchoring and the enum before trusting it as
the boundary. It makes no difference to the partition, there are no replay
predictions at all between 15:57 and midnight that day, but the boundary should
be the thing that actually changed the prompt.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WnNxwxfXSbeNtjvtz5gayb
Nothing in the pipeline has ever fed an outcome back to the thing that makes
predictions. Calibration reads autonomy_outcomes, but calibration only gates
whether to act on a prediction, never what the prediction is. So the only thing
that has ever changed this system's output is a human editing the prompt.
score-replay-runs.js asks "compared to what". The answer is not flattering:
over 2,114 scored replay predictions the system is right 50.05% of the time
while answering "negative" to every one of the same bars scores 54.45%. It is
4.4 points below a constant, z=-4.06. The whole deficit is the prior. It says
positive on 65% of calls when 45.5% of bars beat SPY, a 19 point skew. Its
discrimination, P(up|positive) minus P(up|negative), is +3.2 points with
p=0.15, so the direction it picks is weakly informative and completely buried
by how often it defaults to positive.
The first version of that script compared each direction group's accuracy to
"always that direction" on the same rows, which is an identity and tests
nothing. T3 replaces it with the two proportion test that actually asks whether
the choice of direction carries information.
build-feedback-brief.js turns a run's scored outcomes into a memo the next run
reads before predicting. Generated from the data, not written by hand, or it is
just me editing the prompt again with extra steps.
Replay can now be pinned to an explicit article set, which is what makes two
runs comparable at all. Comparing two calendar windows of one run compares two
market regimes: the epochs in run 1 line up exactly with article vintage, E0 is
late 2024 and E2 is 2026, so nothing could be attributed. A new run also
inherits its parent's watermark instead of recomputing it from today, which
silently guaranteed a different archive slice every time.
prompt_version never moved across four material prompt changes, so every
proposal on record claims to come from the first prompt. coordinator-2 and
replay-coordinator-2.
docs/replay-run-2-preregistration.md fixes the bar before the run exists,
including which result counts as learning and which is only calibration.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WnNxwxfXSbeNtjvtz5gayb
The graph has been built for months and fed nothing but a dashboard. Nothing in
the autonomy pipeline has ever read an edge: not the coordinator, not
calibration, not execution. Its only downstream consumer, trade_signals, last
produced anything in April. It was roughly 70% of the llm bill and informed no
prediction, decision or order.
Relationships are the one piece of context a per-event coordinator genuinely
cannot derive from its own articles, because "this company supplies that one" is
knowledge about companies rather than about this event. So the coordinator now
receives the relationships of the companies the event is about, and is told to
name the relationship in causal_channel when it reasons through one.
The cutoff filter is the part that matters. first_seen_at on a relationship is
derived from article dates rather than processing time, so a historical proposal
only sees what the world had actually revealed by its own cutoff. Without that
this feature would quietly reintroduce the lookahead the evidence check exists to
prevent, and it is pinned by a test rather than left to review.
Relationships are explicitly background rather than evidence: predictions still
have to cite the article ids the story came from, and an instrument the articles
give no reason to care about is still not a prediction.
strategy_version moves to autonomy-2, because a prompt change this material
changes what a prediction means and the two populations should be comparable
later rather than silently blended.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WnNxwxfXSbeNtjvtz5gayb
One untradable ticker rejected everything alongside it. In a single day that was
93 proposals discarding 171 predictions, and 76 of those named something we could
trade perfectly well. They were lost because a sibling in the same response said
EURUSD.
Tradability is a filter, so it applies per prediction now. The untradable one is
dropped and logged, its siblings are kept, and the stored payload records what
was removed so the filtering is auditable rather than invisible.
Lookahead deliberately still rejects the entire proposal. Evidence that did not
exist at the proposal's own cutoff means the response is corrupt rather than
merely untradable, and keeping the rest of it would hide the one thing most worth
seeing. Both halves are pinned by tests.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WnNxwxfXSbeNtjvtz5gayb
The event outcome worker re-requested PSTG and GROQ on every poll for as long as
the process lived, because a fetch failure only logged and continued while the
prediction stayed pending. Ten requests every five minutes, indefinitely. Per
ticker backoff now doubles to an hour, so a symbol with no market data costs one
request an hour instead of one a minute. It also translates dotted tickers the
same way the autonomy worker does, which is why that helper moved into the
shared price module rather than being copied.
The gdelt loop had no pause on its error path at all, so once the api started
refusing connections it spun through failures continuously, burning cpu and
filling the log with the same stack. It has been doing that for days. Backs off
to half an hour now and resets on success.
isTransientCoordinatorFailure matched 408, 429 and 5xx but not a budget 402/403,
so the 380 jobs that dead-lettered during the exhausted quota window could never
come back on their own, including 55 live events. Budget failures are transient
in a way an ordinary auth failure is not, and a wrong key still dies permanently
because it says invalid or unauthorized rather than naming credits.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WnNxwxfXSbeNtjvtz5gayb
Six live predictions were marked unresolvable, including MSFT twice, WMT and
ITW. Re-running the calculation against yahoo resolves all six, so they were
never unresolvable, they were scored before the market data existed and then
thrown away permanently.
The due-check counted calendar days while calculateOutcome finds the exit bar by
trading days. A friday horizon-1 prediction therefore looked due on saturday,
when monday's close cannot exist. calculateOutcome returned null and the worker
treated null as permanently dead. This hit short horizons hardest, which is
exactly the cohort that produces the first live evidence.
The sql filter stays loose because it cannot know about weekends, and trading day
arithmetic now decides what is genuinely ready. A null result waits for the
horizon to be properly past before anything is retired, and says so when it
finally gives up.
Separately, at horizon 1 the entry and exit lookups could land on the same bar
and produce an excess return of exactly zero, which was recorded as a real
outcome and scored as a directional miss. ITW and WMT both did this. A horizon
that has not elapsed is no longer a measurement.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WnNxwxfXSbeNtjvtz5gayb
Three separate things had the pipeline frozen for 27 hours.
browserCrawler leaked page slots. context.newPage() sat outside the try, so a
throw or a hang there took the slot with it, and after maxConcurrentPages of
those every caller parked in acquirePageSlot forever. That is what it looked
like from outside: content workers alive, no logs, no progress, 13 chromium
renderers still up 10 hours after start. newPage is inside the try now, waiting
for a slot times out instead of blocking forever, and page.close() is raced so a
wedged renderer cant strand the slot on the way out either.
graphWorker had no backoff on quota failures. A blown OpenRouter monthly limit
returns an instant 403, so it retried as fast as the network allowed: 2356
failures in 20 minutes, drowning every other line in the log. Quota and auth
errors now pause resolution for 15 minutes and log once per window rather than
once per attempt.
The outcome worker retried unresolvable predictions forever. Yahoo writes class
shares with a dash, so BRK.B 404s every time, and a failed prediction stays open
and comes straight back on the next poll. Dots are translated to dashes, which
matters beyond this one name because the allowlist is full of dotted symbols,
and a prediction that fails five times is marked unresolvable instead of
spinning.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WnNxwxfXSbeNtjvtz5gayb
The JSON shape in both prompts used real values as placeholders, and the model
was reading them as the answer:
instrument: 'NVDA' -> 397 of 611 predictions are NVDA (65%),
second place is LMT with 8
horizon_days: 10 -> 606 of 611 are horizon 10 (99.2%), out of
seven allowed horizons
direction: 'positive|negative' -> 502 of 611 are positive (82.2%)
event_type: 'stable_enum' -> the enum was never listed, so the model
invented one label per event, 201 distinct
values across 611 predictions
replayWorker had its own copy of the same prompt with the same values, which is
why both lanes show the identical skew (replay is 147/147 horizon 10, 138/147
NVDA).
Every placeholder is now a description of the field rather than a usable value,
with an explicit line saying not to copy them. event_type is validated against
the same closed family list the cohort key uses, so a label cannot mean one
thing in the prompt and another in calibration. Off-enum labels are salvaged
through the existing mapper when they are placeable and rejected when they are
not, so 'other' does not quietly become the bin again.
This does not by itself create edge. It means the next batch of predictions
measures the model's judgement instead of its willingness to copy an example.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WnNxwxfXSbeNtjvtz5gayb
Offline evidence can no longer authorise anything. createDecisions used to
prefer a live snapshot and fall back to the pooled historical one, so once the
live lane woke up a live prediction could have drawn a BUY off backfill data.
Backfill and replay are fine evidence that the pipeline works, they are not a
live track record.
No live snapshot now means ABSTAIN. The abstain says whether offline evidence
existed for that cohort, so "we have 60 offline samples but no live ones" stays
distinguishable from "we know nothing about this cohort".
Also split the health counter. It counted qualifying cohorts across every
source, which overstated how close we are to being able to trade now that only
live cohorts can authorise. It reports qualifying_live_cohorts and
qualifying_offline_cohorts separately, and applies the concentration cap it was
previously ignoring.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WnNxwxfXSbeNtjvtz5gayb
Archive ingestion had been dead since 2026-08-02 because nothing in the
compose stack actually ran it. Everything downstream starved from there.
- add ingest + enrichment services. server.js only starts the scheduler when
DURIIN_RUN_SCHEDULER is not "false", and workers/index.js was not running at
all, so articles never got event_id/content/has_embedding and the coordinator
had nothing to lease.
- pass an explicit origin from coordinatorWorker. it was never passed, so
acceptProposal defaulted to 'live' and 464 historical backfill predictions
were recorded as live. that also meant verifyEvidence got a null cutoff and
skipped its date check entirely.
- coarsen cohortKey to event families + horizon buckets. 201 free text event
types produced 221 cohorts averaging 2.76 samples, so the n>=30 gate could
never be reached and everything abstained for the wrong reason.
- gate on cohort diversity, not just sample count. one ticker was roughly half
of all resolved outcomes, so a pure count gate was measuring one company.
unknown diversity abstains rather than passing.
- resolve the admin archive db explicitly and probe it. it relied on a
Dockerfile symlink, and without it better-sqlite3 quietly creates an empty
file and serves a phantom archive.
- clamp implausible future publication dates at ingest.
- pin the db backend to sqlite by default. compose hardcoded postgres "true",
which would have overridden the operator's own .env on the next redeploy and
pointed everything at a stale snapshot.
scripts/repair-autonomy-labels.js relabels the affected rows. it is dry run by
default and has not been applied.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WnNxwxfXSbeNtjvtz5gayb
