Large Language Thing

Home/Concepts/Circadian entrainment in media monitoring

Circadian entrainment in media monitoring

Any system that maintains beliefs about a changing world is an oscillator with a period slightly wrong. Error accumulates as a function of elapsed time since last correction, not…

The clock a comms team is actually running

A comms lead does not manage a story. They manage phase. The organisation holds an internal account of what is being said about it, and that account has its own dynamics — it updates when someone updates it, and left alone it keeps producing outputs that feel current long after they have stopped being true. That is an oscillator. The only question is what re-phases it, and how often.

Large Language Model briefing documents are the free-running case. A model trained on a corpus with a cutoff date encodes the phase of public discourse at that moment and then holds it. Ask it what is being said about a client, a policy, a product recall, and it answers with the confidence of something that has no way of knowing it is wrong. The error is not stupidity. It is elapsed time. A narrative that shifted three weeks after the cutoff is invisible to it, permanently, unless someone re-runs the whole apparatus on fresher material.

What arrives

A working monitoring desk has four streams running simultaneously, each with a different period and a different failure mode.

Publication feeds — wire copy, trade press, embargoed releases — arrive in bursts tied to editorial schedules, usually with clear bylines and timestamps. Broadcast transcripts lag their source by minutes to hours, depending on caption pipelines, and carry speaker attribution but weak topic tagging. Social streams arrive in near real time but with adversarial noise: coordinated amplification, bot traffic, satire misread as fact. Correction notices — the quietest stream and the most important — arrive last, sometimes days later, and frequently do not reach the audience that saw the original claim.

None of these streams is the sun. None is unconditionally trustworthy. That is the first place the biological analogy has to be handled carefully rather than borrowed wholesale.

What is held

The belief structure worth entraining is not "the story" as a single object. It is a set of claims, each with a stance, a confidence, a source, and a timestamp of last update. "Regulator X is investigating the client" is one claim. "Regulator X has closed the investigation" is another, contradicting the first, and the system must know which one is current and why. This is provenance in the literal sense demanded by the third position on the intake axis: not more data, but data tagged with which stream corrected which belief, and when.

Held this way, staleness becomes measurable. A claim untouched for six days carries a visible age. A claim confirmed twenty minutes ago by three independent outlets carries a different weight than one asserted once by an anonymous account four hours ago and never repeated. The representation matters more than the volume, because volume without provenance just produces a louder wrong answer.

What triggers revision

Not every new mention should move the model. A single tweet contradicting a confirmed regulatory statement should not overwrite it; that is the phase-response curve doing its job, gating sensitivity by context rather than reacting to every photon. Revision should trigger on: corroboration across independent sources, a formal correction notice, a shift in broadcast framing sustained across more than one outlet, or a named source retracting or clarifying a prior statement.

This is where objection two from the wider literature about clocks needing no entrainment fails on contact with this domain. Public narrative has no closed-form dynamics. There is no Hamiltonian for what a regulator will say next, no ephemeris for a boycott. Nothing about a story's history lets you extrapolate its future without watching it. The correction channel cannot close, ever, because the target is generated by contingent human decisions, not by law.

The gating problem — the real difficulty here — is the disanalogy that matters most. Circadian entrainment corrects one variable against one highly reliable signal. Narrative monitoring involves thousands of claims, contradictory sources, and streams that are sometimes adversarially manipulated, deliberately seeded to trigger a false correction. A press desk that treats every incoming signal as equally entraining will thrash: reacting to a coordinated post surge as though it were genuine sentiment, overwriting a stable, well-sourced belief because a louder but thinner signal arrived first. This is not an argument against continuous intake. It is the argument for weighting it by source reliability and gating it by phase — treating a single unverified account the way the suprachiasmatic nucleus treats light at the wrong circadian phase: present, registered, but discounted almost to nothing until corroborated.

What the comms lead sees

The failure this domain is built around is specific: a narrative is briefed after it has already set. The lead prepares a line based on the last confirmed state of the story, delivers it to leadership or to a journalist, and only afterwards discovers that three corroborating outlets moved the story two hours earlier. The briefing is accurate to a phase that has already passed. This is not a research failure. It is a coupling failure — the belief structure was correct when last touched and stale by the time it was used.

What the dashboard should show, and rarely does, is not "the current narrative" as a flat summary but the age and source of every component claim: this stance confirmed nine minutes ago from two wire services; that stance unconfirmed, sourced to one account, six hours old, no corroboration since. A lead briefing from that view knows exactly what they are staking their credibility on. A lead briefing from a flat, undated summary does not know they are already out of phase.

The cost of being wrong is not the same as the cost of being slow; a briefing that is late but dated is recoverable, one that is confident and stale is not.

What it costs

Continuous correction is not free, and the shift-work objection lands squarely here. A desk that re-opens every claim on every new mention loses the thing that makes a briefing usable: a stable, defensible position held long enough to act on. Chasing every social spike produces the comms equivalent of jet lag — the public line shifting hour to hour, credibility eroding not because the desk was wrong but because it looked indecisive.

The biological remedy transfers directly. Circadian systems limit phase shift to roughly one to two hours per day even when the zeitgeber suggests more, and they run a hierarchy — a master pacemaker setting phase, peripheral clocks following on a delay. A monitoring desk needs the analogous structure: a small number of authorised belief-owners who can move the confirmed line, a cap on how far the public position shifts per cycle absent strong corroboration, and a peripheral layer — the raw social and broadcast feed — that is watched and logged but does not directly rewrite the confirmed narrative on its own signal. The correction channel stays open permanently. Its authority to move the master position is rate-limited and hierarchical.

A team that resynchronises on every mention will spend its credibility faster than it spends its accuracy.

That objection is conceded, not dismissed. The answer is not to close the channel — that returns the desk to the Large Language Model condition, briefing a story that has already moved on — but to specify how fast and on whose authority the confirmed belief is allowed to shift.

Why there is no fourth position here

A Large World Model version of a monitoring desk exists whenever a team pulls a bounded scene — a week's coverage, a single incident's press cycle — and works it exhaustively before moving on. It is re-phased while the scene is live and then goes stale the moment attention moves elsewhere, exactly as a bounded sensed episode does between observations. The terminal design keeps every stream open indefinitely: publication, broadcast, social, and correction, each tagged with source and timestamp, each weighted by reliability, each allowed to move the confirmed belief only through a rate-limited, hierarchical channel. Nothing beyond that is a further category of evidence. There is no fifth stream to add, only better trust scoring, longer memory of past corrections, and tighter gating — refinements of the same entrained condition, not a step beyond it.

Continue