Library

An AI podcast about token economics, a small TKOS demo, a v0.2.2 experiment, and a clearer direction for TopicSpace — all in one day.

Today started with The AI Daily Brief on token economics and ended with belief observability as the category we work in. Two CLI commands against one SQLite substrate proved the dual-consumer claim at the fixture level (tkos state for humans, tkos overlay for models). v0.2.2 on 75 workflow-state questions showed the smallest overlay performed best: A=10.7%, B100=5.3%, B250=5.3%, B500=8.0%. The AI-facing surface is an attention compressor, not a database dump. Models need compact state. Humans need inspectable traces. Same substrate, different surfaces, different budgets.

Long-running assistants can get workflow state wrong even when they have the recent messages. A small experiment in helping LLMs remember what is still true.

A recent log shows what happened. It does not necessarily tell the model what is still true. In a paired comparison over 73 Claude assistant-session questions, adding an operational belief overlay — active beliefs like validation_pending, fix_attempted, action_blocked — cut workflow-state errors from 11.0% to 5.5%. The largest single-metric move was on false-completion claims: 27% under raw log alone, 7% with the overlay. Long-running assistants need a maintained representation of what they currently believe is still true.

Long-running coding assistants don't only need better memory. They need maintained operational state — what is running, what failed, what was fixed, what still needs validation, and what the user has actually approved right now.

Coding assistants don't just make mistakes because they misunderstand code. They make mistakes because they lose track of what is currently true in the workflow. The deploy-approval bug is a clean example: the same user signal that grants approval also triggers the deploy, so the "approval is still required" belief gets retired by the very signal the deploy is responding to. The failure isn't in the model — it's in the lifecycle. Measured across 164 Claude sessions and 20,190 evaluation turns.

An anatomy of runtime epistemic failure. A small clock glitch from a real conversation, dissected as a complete sweep of the architectural gaps we don't yet have.

An AI that wishes you goodnight at 9:40 AM isn't hallucinating — it's failing across the whole belief-lifecycle stack at once. A prior was born under strong warrant; the warrant aged but never expired; fresh contradicting evidence arrived but no arbitration ran; an intervention deployed against the obsolete state. Every layer of runtime epistemic infrastructure was relevant. Every one was missing.

Why runtime governance should be empirically certified before deployment. A working architecture refuses to deploy one of its own governance interventions when the measured behavioral delta against baseline isn't large enough to justify the cost.

Most runtime LLM governance is declared, not earned: an instruction is written, then shipped to 100% of traffic without measuring whether it changes anything. An earned-governance gate runs A/B against the un-intervened baseline and refuses deployment when the delta is too small. The system rejected one of its own plausible interventions today — that rejection is the architectural contribution.

A Temporal Knowledge Operating System for LLMs. The core limitation of large language models is not parameter count or context length — it is epistemological. To move beyond static mimicry we need a programmatic, self-correcting state layer between raw input streams and inference: one that manages how beliefs evolve under uncertainty.

Hallucination is a calibration failure, not a generation failure. A TKOS lets the LLM consult its own reliability map mid-inference (<50ms via an MCP tool) and either express calibrated doubt or fall back honestly when the query is uncalibrated. The substrate exists today as a batch pipeline; the consultation interface is the engineering path that remains.

Messy events → hidden structure → uncertain expectations → changing beliefs → feedback from reality. The same five-layer stack keeps showing up across markets and AI governance. An attempt to name the pattern, say where it fits, and say where it doesn't.

Repeatable events + meaningful patterns + measurable outcomes + enough history → the stack pays for itself. One-off decisions, low-sample work, pure taste, no-feedback domains → it doesn't. The shape holds across domains; only the parameter surfaces (action vocab, risk classes) change.

Most AI evaluation collapses behavior into a single pass/fail score. A more useful question is where a model is reliable, where it's unstable, where risk priors are wrong, and where guardrails will over-fire or under-fire. A first attempt at region-based calibration on real conversation logs.

Aggregate behavior prediction did not beat baseline (−2.8pp). The value is regional. Some regions calibrate cleanly (88% top-1); others miss badly (12%). Six (region × risk) cells diverge by ≥20pp between prior and actual — guardrail-mis-fire candidates the architecture surfaces before the guardrail ships.

The core idea behind topicspace is to treat a changing information environment not as a feed, but as a field. Once that representation exists, additional fields can be derived from it — each capturing a different layer of understanding.

A stream tells you what arrived. A field tells you where structure exists. An expectation field tells you what that structure implies. A revision field tells you how implications change. A performance field tells you which deserve trust. Each layer adds a class of question the one below cannot answer.

The original idea was not to generate more questions. It was to maintain continuity of understanding in a changing field. The expectation-field replay is the first version of that idea that actually feels like the original concept.

Broad expectations are durable. Narrow ones are mostly noise. Weakening precedes death by two to three trading weeks. The patterns held across two families and two starting dates.

Not stock picks. Not recommendations. Five forward tests of what the current topicspace field model implies should happen next, each tied to a specific pattern in the data and stated in a way that can be falsified.

A model that never risks being wrong isn't doing much. The most useful research isn't the kind that sounds confident — it's the kind that can be checked.

Across 5+ months of backtest data, topicspace measures a +4.92pp gap in 10-day forward excess returns between AI hardware names and AI software/platform names — with three counterintuitive sub-findings.

Hardware +3.29pp, software −1.63pp, gap +4.92pp (Welch t=13.5). Software CONFIRMED is the worst software setup. CRWV behaves like hardware. The AI-celebrity hardware names are the cluster's laggards.

A new layer that turns a claim, basket, or ticker question into a grounded brief built from topicspace fields. Not a chatbot, not a recommendation engine — a layer for interrogating the field.

A dashboard shows what's happening. intel helps answer a harder question: what does that actually mean here?

Strong narrative-price gap is not the same as a credible setup. Meta is the cleanest example of why the validity layer matters — and why hesitation can be the insight.

A strong raw signal gets your attention. Validity tells you whether to trust it. Transitions tell you what happened next.

We tested whether narrative-price state could improve stock selection in the AI ecosystem. Raw dislocation was the ranking core. State and sector context made the signal investable.

S4 Sharpe 2.10. Cap S4+B4 Sharpe 2.44. ML ranking failed at this sample size. Pure NDS ranking inside the eligibility layer remains the rule.

After backtesting the model, topicspace is becoming more specific: where narrative pressure is strongest, where that signal is actually valid, and how to turn it into a disciplined workflow.

How topicspace measures the gap between what analysts are saying and what the market is pricing — and what that gap has historically meant.

Narrative-leading states outperformed price-led states. Transitions carried more signal than static presence.

Why narrative intelligence requires field-level state modeling rather than document retrieval — and how topicspace is built around that distinction.