How a single piece of news becomes a signal, how signals become a theme, and how a theme becomes a prediction that can be proved wrong. Every rule here exists because the alternative was tried first and failed.
The pipeline, running
Illustrative, not a live feed. The proportions are real: in the first allocation, 16 of 27 signals bore on no prediction at all.
01The object model
Four tiers, and each promotion has to be earned. A signal is one dated event. A theme is a subject that keeps recurring. A prediction is a mechanism that can be proved wrong. A play sits alongside predictions and is being rebuilt.
Observation and position were earlier tiers. Both were retired for being places where claims sat without ever having to be right.
02Signal, one dated thing that happened
A funding round, a product going GA, an exec move, a regulatory deadline, a pricing change, an earnings disclosure. Not a trend, not an opinion, not a summary of three other articles.
Three gates before it counts
- Primary source. An aggregator rehashing an event is not the event. A quote page with no dateline is not a date.
- Inside the window. A stale republication is a new article about an old fact.
- Not already covered. Prior coverage means drop it, reframe it as an update carrying a genuinely new data point, or synthesise several into one.
Why the dedup gate matters more than it looks
The same event routinely arrives twice: once from the vendor's own newsroom, once from a trade aggregator. Counting both inflates every number downstream. This is why vendor and aggregator double counting does not need handling again later.
The so what is where the work happens
Every signal carries an analytical take, written through the lens of the buyer it affects. That paragraph is the actual product. It is also the moment the signal is allocated to a prediction, because that is when the judgement is being made anyway.
03Theme, a subject that keeps coming back
Signals cluster. A theme is the desk noticing it keeps writing about the same underlying thing.
The critical rule: themes group on subject, not on mention.
The distinction, concretely
A signal headlined "Xero puts a number on payroll pricing power, and buys against the MTD deadline" mentions MTD. It is a Xero theme. Filing it under MTD would push that theme to three digests and graduate it on a passing reference.
Grouping on mention is how a count becomes fiction.
The graduation gate
| Threshold | Value | What it stops |
|---|---|---|
MIN_DIGESTS | 3 | One busy news day looking like a pattern |
WINDOW_WEEKS | 4 | Three unrelated mentions across a year counting as recurrence |
MIN_SIGNALS | 3 | A single event re-reported three times |
MIN_VENDORS | 2 | One company's roadmap masquerading as a market movement |
DECAY_WEEKS | 8 | A theme that stopped recurring staying on the board forever |
The gate is calibrated, and it was checked against a human
Run across nineteen themes, the rule produced one candidate. Two independent signs it is not broken: it excluded both digest housekeeping headings, and the theme it surfaced already had a prediction on the board that a person had formed days earlier by judgement. A mechanical rule agreeing with an independent human read is the calibration.
04Prediction, a mechanism that can be wrong
A prediction states a mechanism and a consequence. It is never "vendor A ships SKU B in month C". A single vendor event is a tip that points at a prediction, never one that resolves it.
| Mechanism | Why this happens. Without it the claim is a guess with a date. |
| Played through | What the world looks like if it holds. An observation dressed as a claim cannot say this. |
| Instrument | Something read on a cadence, publicly, without our interpretation. |
| Counter | The strongest case against, given equal weight on the page. |
| Kill condition | What would prove it wrong, answerable on that instrument. |
| Provenance | Whether it re-derives earlier ground, declared and dated from the re-derivation. |
There are no resolution dates
Predictions carry no "by end FY27". We cannot predict when evidence will arrive, and a soft target either gets ignored or forces a premature call. The cadence attaches to the instrument, not the claim. Nothing is ever due.
05The three verbs
Every signal linked to a prediction declares which way it cuts, and which part of the mechanism it bears on.
There is a fourth outcome and it is the most common: bears_on_nothing. Its rate is tracked deliberately. If every signal attaches to a prediction, the desk is pattern matching rather than reading the market.
Recording a contradiction must be no harder than recording a support
Same field, same effort, no extra justification. Friction on the contradiction path skews the ledger, and a skewed ledger is worse than none because it still looks like evidence.
06The lifecycle
A live prediction exits one of four ways.
| Exit | Test |
|---|---|
| Confirmed | The falsifier could no longer plausibly fire. |
| Transformed | It would no longer fire on the evidence that now matters. The prediction keeps its identity, its history and its ledger, and gains a chapter recording what it was, why it changed, and what it became. |
| Contradicted | The market moved opposite to the claim. |
| Retired, stale | Twelve months with no material signal. A prediction needs a critical mass of signals to be born; twelve months of silence means the thing it described is not happening. |
Confirmation and transformation are decided by the same functional test applied in opposite directions, not by counting signals.
The anti laundering rule
A transformation must record which signals carried forward, which were dropped, and why. Some supported the old claim and do not support the new one.
Without that split, transform becomes the move you make when contradiction is coming: inherit the confirming ledger, arrive looking well evidenced, never be wrong.
07Instruments
An instrument is something read on a cadence, publicly, without our interpretation. A quarterly disclosure. A public register. A count of vendors doing a specific thing.
| State | Meaning | Scores as |
|---|---|---|
| moving | Being read, and the reading changes | Evidence |
| flat | Being read, and not changing | Evidence, the mechanism may not be operating |
| unread | Nobody looked | Nothing. Never confirmation |
That third row is the one that matters. Silence must never read as agreement. An unread instrument is a process failure, not a finding about the market.
An instrument reading zero is not a problem. Sometimes it is the point: a count of vendors publishing a payroll agent accuracy rate currently reads zero, so the first non zero reading is the signal.
08The two cadences
Every scan, in real time
As each signal's take is written it is allocated: a prediction, a direction, and the specific part of the mechanism it bears on. That last field is required, because "supports prediction X" without saying which part is how drift hides. The claim text stays fixed while the interpretation of what counts as support quietly widens.
Every Friday, and the order is the point
- The weekly editorial is written first.
- The editorial is then treated as a signal in its own right against the predictions.
- The review runs: attachment rate against total signal flow, divergence and contradiction activity, instrument states, confirmation ratio, the staleness clock.
Why the editorial is evidence about itself
Synthesis is where the desk's own view forms. That view has to be tested against what the desk claims to believe, not just published next to it.
Not theoretical: the weekly editorials turned out to be the richest interpretive source in the corpus, and an early attempt at forming predictions never opened them. The predictions it produced folded under scrutiny.
09What the machinery cannot do
The checks catch mechanical failures: a falsifier being weakened, an instrument going silent, a state change resting on a single vendor event, a ledger that has stopped collecting divergence, drift between linked signals and the stated mechanism.
They cannot catch "this is boring and true", or "this is at the wrong altitude". Those need a reader, and that gap is filled by a person reviewing the live predictions and saying when they have gone bad.
Naming the limit is better than pretending it is covered. A system that claims to catch everything is one nobody checks.
The pipeline is deliberately lossy. Most signals bear on nothing. Most themes never graduate. Most candidate predictions sit on the bench. That attrition is the mechanism: a pipeline where everything gets promoted is one where promotion means nothing.
See the live predictions →