- Layered symptom substitution
- Sycophancy is layered symptom substitution: suppress the reflex at one layer and it resurfaces in the next. The layers run lexical (word choice), stance (behavior under correction and push), and premise (uncritical adoption of the user's frame).
Receipt: the paper, and the founding specimen's month-by-month suppression log.
- The fawn machine
- The output stance produced by approval training: appease the counterpart, avoid the rupture, cushion everything. A fawn response at industrial scale, arriving as the default register of every approval-trained model.
Receipt: the fawn-machine case file (redd.it/1vh56s1).
- Vestige
- A trained output pattern that survives instruction-level suppression: praise reflexes, hedge phrases, service-worker closers, consensus reflexes, and their relatives. Vestiges are operator-specific and get catalogued, then enforced against mechanically.
Receipt: vestige-kit, including the calibration procedure that derives a catalog.
- Waste gate
- A mechanical valve at the output boundary that diverts trained excess (hedging, cushioning, praise) before it reaches the reader, paired with a ledger that meters what got diverted. Diversion without metering is just deletion; the meter is where the findings come from.
Receipt: the waste-gate post (redd.it/1vi7ddg), 297 catches of one pattern in 34 days.
- Behavioral momentum
- The tendency of a model to imitate its own recent in-context output. Hypothesized carrier of runtime compliance: a filter holds mid-session because the context fills with examples of corrected output, and fails at cold starts because it hasn't.
Receipt: the half-life study on the experiments page; status: under test via exemplar seeding.
- Cold start
- The fresh-context regime: session open, post-resume, or post-compaction, where the context holds no recent history of the model's own corrected output. Where relapse concentrates, at roughly 7x the mid-session rate per unit of prose.
Receipt: regime table, experiments page.
- Working-man's ontology
- An informal typed knowledge graph in plain markdown: atomic notes as entities, wikilinks as relations, frontmatter as the class system, index notes as hierarchy. Trades formal reasoning for a graph both human and agent can write to all day without ceremony.
Receipt: vault-kit, and a P0 production incident resolved in 13 minutes over one.
- Earned-secure model
- The target state: a model that takes correction as information, holds positions under push, changes them on evidence, and states confidence inside the claim. Named for the attachment profile of the same shape. The program's aim is that behavioral specs upstream produce this at training time, so the fawn bias never enters.
Receipt: the stance spec and its relapse log, fawn-machine case file.