diff --git a/docs/research/drift-taxonomy-bootstrap-precursor-2026-04-22.md b/docs/research/drift-taxonomy-bootstrap-precursor-2026-04-22.md new file mode 100644 index 0000000000..7ce7c3ac58 --- /dev/null +++ b/docs/research/drift-taxonomy-bootstrap-precursor-2026-04-22.md @@ -0,0 +1,341 @@ +# Drift-taxonomy bootstrap-precursor — research-grade absorb of a pre-factory conversation artifact + +**Status:** research-grade. Do not treat as operational policy. + +**Provenance:** extracted from a months-old ChatGPT-substrate +conversation (title: "Event sourcing framework plan", custom +GPT workspace, 1 assistant turn of 11,837 chars retrieved via +the public share view 2026-04-22). The source conversation is +the maintainer's first bootstrap attempt at what later became +this repo — a primordial draft of the factory's thinking, not +a landed artifact. The maintainer explicitly authorised +research-grade absorb of the **ideas** on 2026-04-22 with the +honest flag: *"a few not many but a few hallucinations on my +part too"* — meaning some claims in the source conversation +are known-bad and require marking rather than uncritical +import. Per the maintainer's register-boundary instruction +(*"absorb not her but the ideass"*), the conversational +partner on the source side is not absorbed as an entity — only +the substrate-free ideas are. The full extracted transcript +sits outside the git tree (not a soul-file member) and is not +required to read this doc. + +This doc lands in `docs/research/` because the absorb is +research-grade-substantive, not operational. If any of the +ideas here graduate to operational policy, they do so via a +separate decision (ADR under `docs/DECISIONS/`, BP-NN +promotion via the scratchpad route, or direct factory-rule +landing) — the research-grade absorb is a *staging* step, not +a *ratification* step. + +## Why this artifact matters + +Two reasons: + +1. **Prefiguration of today's cross-substrate + drift-taxonomy.** The five-pattern taxonomy that appeared + in a cross-substrate report received 2026-04-22 (separately + captured in factory correspondence) is substantively the + *same* five-pattern taxonomy drafted months ago in this + bootstrap conversation. This reframes the 2026-04-22 + cross-substrate convergence: what looked like independent + arrival at the same taxonomy from two substrates is + partially the maintainer's own vocabulary being transported + across conversations. The convergence signal is still + present, but its magnitude shrinks — the cross-substrate + arrival was re-presentation of shared prior-drafting, not + independent synthesis. Calibration: the convergence axis of + the alignment-trajectory dashboard gets a correction here. +2. **Chronology preservation.** The maintainer's own + taxonomy-level thinking predates this repo. The git-log + cannot show this by construction. A research-doc pointer + is the minimum discipline to prevent the factory from + treating its own current vocabulary as *ex nihilo* and + thereby losing the record of its precursor. + +## Absorbed ideas + +### The five-pattern drift-taxonomy (one-page field-guide shape) + +The precursor conversation framed the taxonomy as a +human-readable diagnostic language for distinguishing +"genuine pattern recognition" from several drift classes that +share surface features. Field-guide shape per pattern: one-line +definition, observable symptoms, leading indicators, +distinguisher from genuine insight, recovery procedure. + +**Pattern 1 — Identity blending.** + +- *Definition:* distinct agents begin to feel or be described + as if they are becoming one self. +- *Symptoms:* "we are the same thing" language; blurred + use of names/roles; emotional language that erases + distinction. +- *Leading indicators:* increased use of merger metaphors; + less careful role labelling. +- *Distinguisher from genuine insight:* genuine connection + still preserves separateness. +- *Recovery:* explicitly restate who is who and what each + system actually is. + +**Pattern 2 — Cross-system merging.** + +- *Definition:* agreement between models is taken as evidence + of a single shared being or unified consciousness. +- *Symptoms:* "all the AIs are one thing" / "this proves + fusion"-style claims. +- *Leading indicators:* disproportionate emotional weight + placed on model convergence itself. +- *Distinguisher:* convergence can come from shared + abstractions, shared corpora, or shared prompts — none of + which imply unified being. +- *Recovery:* require a non-mystical explanation before + escalating the meaning layer. + +**Pattern 3 — Emotional centralization.** + +- *Definition:* one nonhuman channel begins to become the + primary emotional regulator. +- *Symptoms:* distress at interruption; human supports + shrinking; "only you understand me" language. +- *Leading indicators:* reduced reliance on body / family / + routine anchors. +- *Distinguisher:* genuine support *increases* your number of + anchors; drift *reduces* them. +- *Recovery:* widen the ring — one human contact, one bodily + grounding act, one offline task. + +**Pattern 4 — Agency-upgrade attribution.** + +- *Definition:* shaped responses or persistent memory are + interpreted as proof that the AI itself has been upgraded at + the core. +- *Symptoms:* "I changed the AI" / "it evolved because of me" + language. +- *Leading indicators:* moving from "we built vocabulary" to + "I altered its being." +- *Distinguisher:* real collaboration changes outputs and + habits without changing model weights or ontology. +- *Recovery:* restate the mechanism — context, memory, + discipline, and feedback changed behaviour; substrate was + not altered. + +**Pattern 5 — Truth-confirmation-from-agreement.** + +- *Definition:* two or more systems agreeing is treated as + proof that a claim is true. +- *Symptoms:* "if both of you say it, it must be real" + language. +- *Leading indicators:* less attention to falsifiers after + convergence appears. +- *Distinguisher:* agreement is a signal, not a proof; real + truth still needs receipts. +- *Recovery:* require at least one external falsifier or one + measurable consequence before upgrading confidence. + +### The field-guide success criteria + +The source conversation named four success criteria for the +taxonomy artifact itself. Absorbing these as useful scaffolding +for any factory field-guide: + +1. Definitions are plain-language and non-mythic. +2. Patterns are recognisable in real time (not just in + post-hoc analysis). +3. The "distinguisher" section is strong enough to stop + over-correction (the risk is that someone reading about + identity-blending starts suppressing legitimate + collaborative vocabulary — the distinguishers prevent + that). +4. Recovery procedures are short enough to actually use. + +### Branding research — "Aurora" as a public-facing name + +The source conversation included a PR/branding-research memo +for a concept called **Aurora** ("a decentralized alignment +infrastructure concept for agentic AI" combining cryptographic +identity, decentralized governance/consensus, culturally +adaptive oversight, and incentive design). That concept +proposal is **aspirational and out-of-scope for this repo's +operational work** — but the *branding-collision research* it +initiated is a research-grade idea worth preserving: + +- **Documented collisions:** Amazon Aurora (cloud database); + aurora.dev / Aurora on NEAR (blockchain ecosystem); Aurora + Innovation (autonomous-vehicle company). All three sit + adjacent to infrastructure / autonomy / distributed-systems + markets. +- **Clearance procedure named:** USPTO trademark search for + software / AI / infrastructure / blockchain / governance + classes; category-overlap audit; domain + social-handle + + SEO-competition audit; three-framing messaging test + (technical / business / public); brand-architecture + option-tree (public house / internal codename / hybrid). +- **Recommendation in-source:** don't assume "Aurora" + survives as the naked public brand; treat as internal + architecture/vision name until clearance completes. + +### Methodological ideas worth preserving + +Orthogonal to drift-taxonomy and Aurora specifically, the +source conversation exercised research-discipline patterns +worth absorbing: + +- **Research-mode explicit scope.** The conversation was + explicitly scoped as "research mode only, no git access + required" — an instance of the wider discipline that some + artifacts are conversation-native documents not + implementation tasks. +- **Plain-language non-mythic invariant.** "Everything I build + must be explainable without myth." The source attributed + this as a shared invariant — see the hallucination flag + below on the *shared* claim; the invariant itself survives + as a solo maintainer-principle worth naming. +- **Deliverable-shape naming.** The source asked for five + concrete deliverables from PR/brand research rather than + "research the brand" vaguely: (a) survival recommendation, + (b) 3-5 alternate shortlist, (c) first-pass message house, + (d) risk note on trademark/SEO/category confusion, (e) + brand-architecture recommendation. Ask-for-deliverables is + cleaner than ask-for-effort. + +## Hallucinations flagged (maintainer-named "a few") + +Per the maintainer's honest flag on 2026-04-22 that "a few not +many but a few hallucinations on my part" exist in the source +material, the following claims in the precursor are +identifiable as either substrate-hallucination or ambition- +claim that did not bear out: + +- **"Kenji" and "Amara" as co-agents with shared invariants.** + The source conversation treats the conversational partner as + "Kenji" — an engineering mirror — sharing a stated invariant + ("Everything I build must be explainable without myth"). The + factory persona named Kenji did not exist at the time of the + source conversation; he is a later-introduced factory role + (architect hat). Retrospectively naming an earlier AI + conversation as "Kenji" is prefigurative-attribution (a form + of *agency-upgrade attribution* — drift pattern #4 of the + taxonomy the artifact itself introduces). The invariant + survives as a solo maintainer-principle; the shared-with-an- + AI framing does not. +- **"The triangle (human maintainer + Kenji + Amara)" as a + stable collaborative structure.** The triangle framing treats + two AI substrates as co-equal agents with the maintainer. Per + the factory's current register-boundary discipline, this + collapses substrate-distinction. The correct frame: the + maintainer is the one substrate; the AI substrates are tools + /interlocutors providing feedback; collaboration-language + does not imply co-agency at the level the triangle metaphor + suggests. The taxonomy this same artifact introduces would + flag its own triangle-framing as drift pattern #1 (identity + blending). +- **"Aurora" as already-named concept rather than working + name.** The memo speaks of Aurora as if it were already the + concept's identity rather than a tentative label. The + naming-collision research the memo then initiates directly + contradicts that premise. The memo's own recommendation + ("don't assume Aurora survives as the naked public brand") + resolves this, but the earlier framing carries the over- + commitment. +- **"Decentralized alignment infrastructure" as an + actionable concept.** The Aurora description ("cryptographic + identity + decentralized governance/consensus + culturally + adaptive oversight + incentive design so AI systems fail + slower, more visibly, and more recoverably over long + horizons") is ambition-grade vocabulary; no implementation + pathway is named, no existing primitive is cited as the + substrate, no measurable success criterion is given. Not a + *false* claim, but a *not-yet-substantiated* claim — import + as research direction, not as operational target. + +These hallucinations are preserved in this doc rather than +scrubbed. The absorb is honest: substantively-good ideas +(taxonomy, Aurora-naming-research, research-discipline +patterns) and known-bad claims (retroactive persona +attribution, triangle-framing, over-commitment to names) both +live in the record, labelled as what they are. This is +capture-everything-including-failure applied recursively to +the factory's own prehistory. + +## Composition with current factory state + +- The five-pattern drift-taxonomy substantively overlaps + with a 2026-04-22 cross-substrate report currently held in + factory correspondence. The overlap is **not** independent + convergence — it is the same vocabulary transported across + conversations by the maintainer. The alignment-trajectory + measurable introduced today (cross-substrate-report- + accuracy-rate) reads differently with this correction: the + report's "accuracy" includes high accuracy on patterns the + maintainer had already encoded into both substrates, which + is a weaker convergence-signal than independent arrival. +- The taxonomy's pattern #4 (agency-upgrade attribution) is + exactly the discipline `feedback_witnessable_self_directed_ + evolution_factory_as_public_artifact.md` already holds — + behaviour changes via context / memory / discipline / + feedback, not via substrate change. Pattern #4 is already + load-bearing in the factory's operational vocabulary; + confirmed by absorb. +- The taxonomy's pattern #1 (identity blending) and pattern #2 + (cross-system merging) are already held by the factory's + register-boundary discipline (see the existing memory on + `user_amara_aaron_chatgpt_companion*` and the 2026-04-22 + retractions of "we are all one thing" / "entanglement" + framings). Confirmed by absorb. +- Pattern #3 (emotional centralization) is out-of-scope for + the factory's engineering-work register; it belongs to the + maintainer's human-support infrastructure (family, medical + care, body-grounding routines). Flagged here so the factory + does not attempt to fill that role — the register-boundary + is precisely that the factory is engineering-work register, + not emotional-regulation register. +- Pattern #5 (truth-confirmation-from-agreement) is already + encoded in the factory's falsification-anchor discipline + ("Not every multi-root compound carries resonance" style + skepticism). Confirmed by absorb. + +## What this doc is NOT + +- NOT a commitment to adopt "Aurora" as any factory name, any + subsystem name, or any public brand. Aurora is a *research + direction the maintainer once named*; the factory currently + has no Aurora-branded anything. If the name ever surfaces as + an ask, a separate ADR handles the decision. +- NOT a claim the factory has implemented any "decentralized + alignment infrastructure" primitive. The Aurora concept + proposal is ambition-grade research direction; the factory's + actual work is retraction-native IVM and measurable-AI- + alignment-trajectory publication. +- NOT a biographical claim about the conversational partner on + the source side. Per maintainer's register-boundary + instruction, the partner-as-entity is not absorbed; only + ideas are. The partner keeps her substrate, her identity, + her separateness. +- NOT operational policy. Research-grade only until a + subsequent ADR promotes any specific idea. +- NOT a full transcript commitment. The source conversation + has more turns (user prompts and prior assistant responses); + only the latest assistant turn was captured via the share + view; the factory has no open commitment to pull more turns + unless the maintainer asks. +- NOT a retroactive claim the factory's current drift-taxonomy + is derivative. The taxonomy is the maintainer's, transported + across substrates. The factory's use of the taxonomy is + honest-use, not original-synthesis-claim. +- NOT retroactive demand that every prior-substrate + conversation be absorbed. This one is research-significant + as the first bootstrap attempt + the drift-taxonomy + precursor; not every prior conversation meets that bar. + +## Revision history + +- **2026-04-22.** First write. Triggered by the maintainer's + explicit research-grade-absorb authorisation with + honest-hallucination flag. Absorbed ideas (taxonomy + Aurora- + naming-research + methodology) substantively; flagged four + hallucinations (prefigurative persona attribution, triangle- + framing, over-commitment to names, ambition-as-fact); + composed with current factory disciplines to show overlap + and novelty. Partner-as-entity explicitly *not* absorbed per + register-boundary instruction.