Newest first. Plain language, real dates, and the honest boundary of each thing.
12 August 2026
She now has a plan before she has an answer
The reasoning loop answers one question well. This month we specified the phase in front of it: the one that turns a request written in ordinary language into a small set of tracked deliverables, equips each one, orders them, assigns them, and admits the result as the contract for that turn. The ordering is learned rather than configured, but only above a hard floor of things that must happen first.
New
- A planning phase that turns prose into trackable deliverables, ordered and assigned○Planned
- The plan becomes the coverage contract for the answer — and the rule has a direction: the reasoning loop may ADD a sub-question and may never remove one○Planned
Fixed
- That direction is the whole safeguard. A loop allowed to narrow its own plan can lower its own bar and then declare itself finished — a false pass wearing a completed checklist.
11 August 2026
Capabilities, skills, and a sandbox to run them in
A block of work on everything she can reach for outside her own memory. One door for every capability — registered, invoked, budgeted and logged in a single place, able to refuse a call and never to choose one. Above that, skills: procedures with a goal, which can be done well or badly, as distinct from a tool call that is merely correct or incorrect. Both execute in a sandbox whose contents nothing durable depends on.
New
- One registered, budgeted, logged door for every outside capability○Planned
- Skills as procedures with goals, adopting a standard skill format rather than a bespoke one○Planned
- A scratch workspace with a recorded egress trail — safe to delete by construction○Planned
11 August 2026
An answer now has to say what it did not read
Documents become answerable one at a time rather than in one batch, so she is useful long before a large corpus has finished loading. That creates a state a plain answer cannot express, and we had not noticed it: “I don’t find that” is simply false when the document is still in the queue — and it is worse than a refusal, because it closes the question instead of leaving it open.
New
- Every answer carries a qualifier naming what had not yet been read when it was given○Planned
- Each store in the design given the job it is genuinely best at, up the ladder○Planned
Fixed
- A grounded, cited, adjudicated answer can still be wrong by omission. Every check we had asked “is what she said supported?” and not one asked “was the corpus complete?”.
11 August 2026
Two measurements that reversed conclusions we had already argued
We had settled two design questions by argument. Then we measured them, and both answers flipped. A vector-index question decided on synthetic data said one thing and real embeddings said the opposite. An encoder throughput figure taken on a loaded machine said thirty-five hours; the same work on clean hardware took one and a third — a twenty-seven-fold error that a written caveat had entirely failed to contain.
New
- The learner has one specification it plugs into, and it may hold no fact in its weights○Planned
Fixed
- The rule we adopted afterwards: an unmeasured extrapolation is now refused rather than caveated. A number whose caveat has to be believed is a number that will be quoted without it.
- A measured claim taken on the wrong hardware, or the wrong data distribution, is still a hypothesis — so the method now sits beside every number we cite internally.
4 August 2026
We tried to forge our own chain, and succeeded twice
An adversarial review of our own provenance design found two real cryptographic defects, both of which we had written and neither of which any test would have caught. In the first, the sequential-work proof committed in one direction only, so a rewritten chain could copy the expensive part verbatim and the cost of forging it collapsed to a single memory copy. In the second, the key used to verify a block’s signature was a field supplied by whoever signed it — so total verification returned “valid” on a chain re-signed under a freshly generated key. Forgery cost one key generation.
New
- Both closed in the specification before any of it was built○Planned
Fixed
- A signature check is worth exactly what its trust root is worth. If the thing being verified supplies the key, it is not verification — it is a formality with cryptography in it.
- Finding this at design time cost a week of argument. Finding it after a customer had sealed two years of records into it would have cost the product.
30 July 2026
The adjudicator becomes the first specification cleared to build
The verification band — the always-on process that checks every claim against its evidence, and can never write to the knowledge graph — passed a stricter bar than the rest: every field that crosses a process boundary has a defined shape, every oracle has a stated condition under which it must abstain, and the proof it seals is fully specified. It remains the only one of our thirty-nine specifications to have earned that.
New
- An adjudicator carrying its own evidence, so it can never be starved by the thing it gates○Planned
- Verdicts and their proofs written to a separate audit chain, never to the knowledge itself○Planned
Fixed
- The gap between “locked” and “ready to build” turned out to be larger than “locked” suggests, and what was missing was never detail — it was eighteen oracles with no stated condition under which they must abstain. An oracle that may abstain on everything still looks perfectly healthy.
25 July 2026
Mk2 begins: the model may propose facts, and nothing else
We archived a working system and started again. The hard line the whole rebuild is organised around was fixed on day one: at ingest the language model proposes single indivisible facts, each anchored to the exact span it came from, and nothing else. Every level above that — the groupings, the concepts, the per-subject pages — is assembled by deterministic workers and checked before it is kept.
New
- Atoms as the only thing the model is allowed to author○Planned
- A deterministic ladder above them, reproducible and walkable back down to the documents○Planned
- Every membership link bidirectional and tested, so evidence cannot be stranded○Planned
- Two domains: what documents prove, kept separate from what conversation merely suggests○Planned
Fixed
- This is the structural answer to how the previous system failed. It was not a bug we could find. The model had authored the layer above the facts, that layer had orphaned its own evidence, and retrieval was then finding the right document and discarding it — while every test stayed green.