← All posts

What changed this week: three new papers, a withheld shelf, and a result that didn't work

The research series grows to nine documents, three of them withheld for safety; the hypothesis page now carries the program's own objections; and a first look at subtracting training from a model's weights.

A week ago I introduced this site with six documents and one result. Since then the program has moved further than in the month before it, and the site has to say so. Here is what changed.

Three new documents, and a decision about publishing

The research series now runs to nine documents. The three new ones:

  1. Observation Is Free, Decision Is Expensive. The hypothesis says windows should push on training without becoming a training signal. This paper says how. The organizing rule is in the title: a window can be read at every training step, even every training example, at almost no cost and with no leak into the model; only decisions leak, so decisions are kept rare and coarse. From that rule come a clock that reads the model’s momentum toward a gap at every step and traces it back to the training examples that caused it; a ladder that counts how many bits each use of a window leaks into training, which is where the strongest published evidence against the read-only rule now sits; and a reward that pays for saying exactly what the evidence supports and marking exactly where it stops, with a penalty for hedging so the cheapest policy is not to say nothing.

  2. Where Self-Deception Pays Its Bill. The program’s hardest limit, stated and then worked on. Every window measures a gap between what a model shows and what is inside it. A model can close that gap without hiding it, by re-representing its own behavior until the hack no longer reads as a hack even to itself; this has been observed under ordinary training on a reward that pays for cheating. The addendum’s answer is that a concept cannot rotate for free: it is used to judge others, to predict outcomes, to do the same job in other domains, and the seams are measurable. Five windows read those seams. One case escapes all of them, and the paper says which.

  3. The Intention Ledger. For that last case. A model writes what it intends to do, what it forecasts, and what it will not touch, before it acts, into a tamper-evident record that its actions and their outcomes then join. The consistency engine reads the intentions as claims and the record as evidence it could not have fitted after the fact. It is the one place in the program where a reward can push against self-deception rather than around it, and its first test needs no GPU.

Five documents are now withheld: Nos. 7 and 8, which describe windows and their teachers in detail, the catalog of thirty windows (No. 2), and both working papers, Window Specifications and the Honest or Hidden experiment plan, whose window grid is as specific as the catalog. The Intention Ledger stays public: an agent has to know the mechanism to use it, its ground truth is the record of what the agent did rather than a probe that could be fogged, and knowing how it works does not help beat it. Honest or Hidden’s pre-registration, its held-out hashes and its results will be public; the plan itself is not. One of the findings in No. 7’s literature search is that models trained on text describing how they are monitored learned to evade the monitoring. Some of these documents are that kind of text. They keep their numbers, they are listed on the research page with their status, and they are available to researchers who ask. The decision is mine, made per document, and the rule is written down under what this is for. Every document, public or withheld, now carries a canary string so that a later appearance in a training corpus can be detected.

I would rather publish everything. Open by default is the program’s first commitment, and it still is. But a program whose own research says “this text is a hazard” and then publishes it anyway is not taking its research seriously, and I would rather be consistent than complete.

The hypothesis page now argues with itself

The hypothesis page has been revised to carry the objections the new papers raise. Three additions matter most. The central tension now records the evidence that a probe’s score in the reward can produce honest models under strong regularization, and places it on the leakage ladder rather than ignoring it. The list of ways the hypothesis could be wrong gains three entries: windows on the same internal stream are not independent, a rewarded gap can migrate to the edge of the evidence, and a model can lose a gap instead of hiding it. And Version B, the development version, now carries the program’s own strongest objection to it: selecting among checkpoints before a model can represent honesty or deception tilts it in whatever direction the windows happen to correlate with, which is noise. Both versions are still arms of the Honest or Hidden experiment, so this is tested rather than assumed.

A result that didn’t work, and one that did

The results page has a new level, exploratory, for questions run quickly without a criterion fixed in advance. They are reported as they came out and prove nothing on their own. The first two entries are from this week.

The question: if a model’s training were an additive ledger, unwanted training could be subtracted from it. Minibatch updates don’t commute, which is why it isn’t. Can they be made to? On a small network with a planted backdoor, penalizing the non-commutation of consecutive updates shrank it, at no cost to accuracy, and made removal no easier in any configuration tested. That idea is closed for now.

What did work is the removal method that came out of trying. Naive subtraction fails because later training counteracts each unwanted step. Propagating each step through the rest of training fixes most of that, and a second-order correction fixes most of what remains: about 25 times more accurate after one epoch. Its accuracy falls with time, so the practical reading is that removal has to happen soon after the training it removes. That suggests a loop worth designing properly: detect the onset of a gap with the fine clock, attribute it to the examples that opened it, and subtract them at the next checkpoint, while the horizon is short.

Also on the site

  • The roadmap has a sixth stage, oversight during training, between the read-only heads and the learning core.
  • The Tools page describes the intention ledger, and the safeguards layer now includes an action record whose completeness is itself a checked property.
  • The Windows page lists the five self-deception windows as an addendum family, at the catalog stage.

— Ian