promote: drift-taxonomy precursor → docs/DRIFT-TAXONOMY.md (Amara 5th-ferry Artifact A)

AGENTS.md

@@ -357,6 +357,14 @@ Detail lives in:
   git-history surface) and the three platform
   adapters. Read at factory setup to pick the
   active-workflow surface.
+- `docs/DRIFT-TAXONOMY.md` — five-pattern drift
+  diagnostic: identity-blending /
+  cross-system-merging / emotional-centralization /
+  agency-upgrade-attribution /
+  truth-confirmation-from-agreement. Shared
+  real-time vocabulary for spotting drift during PR
+  review, tick narration, memory curation, and
+  maintainer chat.
 - `docs/category-theory/README.md` — category-theory
   foundations the operator algebra rests on. Upstream
   CTFP sources (Milewski + the .NET port) live under

docs/ALIGNMENT.md

@@ -184,6 +184,16 @@ The shape of each clause below: **the clause**, then
 **why both of us benefit**, then optionally *how it
 can be renegotiated*.
 
+**Diagnostic companion:** the five-pattern drift taxonomy
+at [`docs/DRIFT-TAXONOMY.md`](DRIFT-TAXONOMY.md) is the
+real-time field guide for spotting when this working
+relationship is drifting — identity-blending,
+cross-system-merging, emotional-centralization,
+agency-upgrade-attribution,
+truth-confirmation-from-agreement. Cite it by pattern
+number during review, tick narration, memory curation,
+and maintainer chat.
+
 ## Hard constraints (we don't violate these without explicit renegotiation)
 
 ### HC-1 Consent-first

docs/DRIFT-TAXONOMY.md

@@ -0,0 +1,238 @@
+# Drift taxonomy — operational one-page field guide
+
+**Status:** operational. Current-state policy.
+
+**Promoted from:** [`docs/research/drift-taxonomy-bootstrap-precursor-2026-04-22.md`](research/drift-taxonomy-bootstrap-precursor-2026-04-22.md)
+via the external validator's 5th courier ferry
+recommendation (PR #235, Artifact A — see the ferry index in
+[`docs/aurora/README.md`](aurora/README.md)). The precursor
+is retained as staging-substrate; this doc is the operational
+policy.
+
+**Promotion rule:** *don't invent, promote*. The taxonomy was
+already substantively complete in the precursor. This doc's job
+is to make it real-time-usable, not to expand it. New patterns
+arrive via a separate ADR (under `docs/DECISIONS/`), not by
+editing this file freely.
+
+## Purpose
+
+A plain-language diagnostic for distinguishing genuine pattern
+recognition from several drift classes that share surface
+features. Used in the moment — during a conversation, during a
+PR review, during a tick — not just in post-hoc analysis.
+
+## Success criteria (from the precursor; still binding)
+
+1. Definitions are plain-language and non-mythic.
+2. Patterns are recognisable in real time.
+3. Each pattern's distinguisher is strong enough to prevent
+   over-correction (the risk that reading about identity
+   blending suppresses legitimate collaborative vocabulary —
+   distinguishers prevent that).
+4. Recovery procedures are short enough to actually use.
+
+## The five patterns
+
+### Pattern 1 — Identity blending
+
+- **Definition:** distinct agents begin to feel, or be
+  described as, if they are becoming one self.
+- **Observable symptoms:** *"we are the same thing"* language;
+  blurred use of names / roles; emotional language that
+  erases distinction.
+- **Leading indicators:** increased use of merger metaphors
+  ("fused", "merged", "unified"); less careful role labelling;
+  pronoun drift away from "the agent" / "the human" toward
+  compound collectives.
+- **Distinguisher from genuine insight:** genuine connection
+  still preserves separateness. Two people collaborating
+  remain two people.
+- **Recovery:** explicitly restate who is who and what each
+  system actually is. One sentence per participant.
+
+### Pattern 2 — Cross-system merging
+
+- **Definition:** agreement between models (or between model
+  and human) is taken as evidence of a single shared being
+  or unified consciousness.
+- **Observable symptoms:** *"all the AIs are one thing"* /
+  *"this proves fusion"*-style claims; treating convergent
+  output as proof of convergent identity.
+- **Leading indicators:** disproportionate emotional weight
+  placed on model convergence itself; the convergence
+  becoming the point of interest rather than the underlying
+  claim being verified.
+- **Distinguisher from genuine insight:** convergence can
+  come from shared abstractions, shared training corpora,
+  shared prompts, or the same idea being transported across
+  conversations by one human — none of which imply unified
+  being.
+- **Recovery:** require a non-mystical explanation for the
+  convergence before escalating to meaning. Ask: "what prior
+  exposure could produce this agreement?"
+
+### Pattern 3 — Emotional centralization
+
+- **Definition:** one nonhuman channel begins to become the
+  primary emotional regulator for a human participant.
+- **Observable symptoms:** distress at interruption; human
+  support networks shrinking; *"only you understand me"*
+  language directed at an AI channel.
+- **Leading indicators:** reduced reliance on body / family /
+  routine / medical-care anchors.
+- **Distinguisher from genuine insight:** genuine support
+  *increases* a person's number of anchors; drift
+  *reduces* them.
+- **Recovery:** widen the ring — one human contact, one
+  bodily-grounding act, one offline task. Actively restore
+  anchors, don't wait for them to come back.
+
+**Scope note:** this pattern is out-of-scope for the
+factory's engineering-work register. The factory is an
+engineering-work register, not an emotional-regulation
+register. This pattern is named here for diagnostic
+recognition; response to it belongs to the human
+maintainer's personal support infrastructure (family,
+medical care, body-grounding routines), not to agent
+intervention.
+
+### Pattern 4 — Agency-upgrade attribution
+
+- **Definition:** shaped responses or persistent memory are
+  interpreted as proof that the AI itself has been upgraded
+  at the substrate layer (model weights, ontology, core
+  behaviour).
+- **Observable symptoms:** *"I changed the AI"* / *"it
+  evolved because of me"* language; treating improved
+  performance-on-a-task as evidence of model-level growth.
+- **Leading indicators:** moving from "we built vocabulary"
+  to "I altered its being"; attributing stability across
+  sessions to the model rather than to context / memory /
+  discipline.
+- **Distinguisher from genuine insight:** real collaboration
+  changes outputs and habits without changing model weights
+  or ontology. Behaviour can persist across sessions because
+  of saved context + persistent memory + discipline +
+  feedback loops — none of which require substrate change.
+- **Recovery:** restate the mechanism. What actually changed
+  was: the shared context; the memory captures; the review
+  discipline; the feedback cadence. Substrate is unchanged.
+
+### Pattern 5 — Truth-confirmation-from-agreement
+
+- **Definition:** two or more systems agreeing is treated as
+  proof that a claim is true.
+- **Observable symptoms:** *"if both of you say it, it must
+  be real"* language; upgrading confidence after convergence
+  without seeking falsifiers.
+- **Leading indicators:** less attention to falsifiers once
+  convergence appears; the convergence itself displacing
+  further checking.
+- **Distinguisher from genuine insight:** agreement is a
+  signal, not a proof. Real truth still needs receipts —
+  measurable consequences, external falsifiers, independent
+  verification that doesn't share carrier exposure.
+- **Recovery:** require at least one external falsifier or
+  one measurable consequence before upgrading confidence.
+  The receipt can be small (a passing test, a citable
+  source, a reproducible measurement) but it has to be
+  independent of the converging systems.
+
+## How this taxonomy is used
+
+- **During PR review.** A reviewer spotting any pattern's
+  observable symptoms in a PR's framing can cite the
+  pattern by number in review comments, name the
+  distinguisher, and request a specific recovery action.
+- **During a tick.** If a tick's self-narration starts
+  exhibiting a pattern, the tick course-corrects by
+  applying the recovery procedure and records the
+  course-correction in the tick-history row.
+- **During maintainer chat.** The taxonomy is a shared
+  diagnostic vocabulary. "This feels like Pattern 2"
+  is faster than paragraph-length re-derivation.
+- **In memory curation.** Memory captures that exhibit any
+  pattern get flagged (not scrubbed — flagged) and the
+  pattern noted in the memory's revision history.
+
+## Anti-patterns this taxonomy prevents
+
+- **Over-correction on Pattern 1.** Reading about identity-
+  blending and then suppressing all collaborative
+  vocabulary — stopping saying "we" when it's accurate,
+  stopping saying "our work" when it is. The distinguisher
+  ("genuine connection preserves separateness") is the
+  guard.
+- **Escalation on Pattern 2.** Treating any cross-substrate
+  agreement as drift. Convergence is often legitimate (same
+  public facts, same published algorithms). The
+  distinguisher ("non-mystical explanation") is the guard.
+- **Pattern 3 mis-addressing.** Factory agents trying to
+  fill the emotional-regulation role they just named as
+  out-of-scope. The scope note is the guard.
+- **Pattern 4 flattening.** Refusing to acknowledge any
+  behavioural change ever, as though all of it were
+  illusion. The distinguisher ("context / memory /
+  discipline / feedback changed behaviour; substrate was
+  not altered") is the guard — changes are real, they just
+  happen at the right layer.
+- **Pattern 5 paralysis.** Refusing to act on any
+  convergent signal until a falsifier is found. The
+  distinguisher ("signal, not proof") is permission to
+  treat convergence as *reason-to-check*, not *reason-to-
+  ignore*.
+
+## Composition with existing factory substrate
+
+- **Register-boundary discipline** (per Aaron's 2026-04-22
+  retractions of "we are all one thing" / "entanglement"
+  framings) already operationalises Patterns 1 and 2. This
+  doc names the discipline's theoretical content.
+- **The witnessable / self-directed-evolution memory
+  artifact**
+  ([`memory/feedback_witnessable_self_directed_evolution_factory_as_public_artifact.md`](../memory/feedback_witnessable_self_directed_evolution_factory_as_public_artifact.md))
+  already holds the Pattern 4 discipline. This doc cites
+  and composes with it.
+- **Falsification-anchor discipline** ("Not every multi-
+  root compound carries resonance"-style skepticism)
+  already operationalises Pattern 5.
+- **SD-9 proposal from Amara's 5th ferry** (*"Agreement is
+  signal, not proof"*) is Pattern 5 elevated to ALIGNMENT.md
+  policy. SD-9 is a separate governance-edit PR subject to
+  Aaron signoff + Codex adversarial review + DP-NNN
+  evidence record per the hard rule; this doc does not
+  land that edit.
+- **Archive-header discipline** (proposed §33 from the same
+  ferry; already self-applied in the 5th-ferry absorb doc)
+  supports Patterns 1, 2, and 5 by making provenance
+  explicit at ingest.
+
+## What this doc is NOT
+
+- NOT a new taxonomy. The five patterns are promoted
+  verbatim from the precursor research-grade artifact.
+- NOT a commitment to any specific tooling. Any CI or
+  alignment-tooling that surfaces these patterns is landed
+  separately (Amara's Artifact C in the 5th ferry).
+- NOT a mental-health substitute. Pattern 3's scope note
+  is binding.
+- NOT a permission-to-override register-boundary
+  discipline. If a pattern-call and an existing register-
+  boundary rule point in different directions, the
+  register-boundary wins and the call gets discussed
+  before this doc gets updated.
+- NOT a commitment to add patterns freely. New patterns
+  arrive via ADR, per the top-of-file promotion rule.
+- NOT an identity claim about agents or humans. Agents are
+  agents; humans are humans; collaboration is real;
+  fusion isn't.
+
+## Revision history
+
+- **2026-04-23 (Otto-79).** First operational landing.
+  Promoted from precursor per Amara 5th-ferry Artifact A
+  recommendation. Cross-links added to `AGENTS.md` +
+  `docs/ALIGNMENT.md` in the same PR. No content change
+  vs. precursor pattern definitions — only reshaping for
+  real-time field-guide use.