Shay f6c1f01a13 Lane 4: Registry Consolidation (language_packs to packs)

2026-07-04 15:11:28 -07:00

16 KiB

Raw Permalink Blame History

ADR-0073 — Anchor lens: substrate-driven substantive variation

Status: Accepted (umbrella ratified; sub-ADRs L1.1–L1.4 each ratify in isolation) Date: 2026-05-19 Ratified: 2026-05-19 Author: Shay Phase: Plan Phase L1 (anchor lens, inside-out) Builds on: ADR-0048 (pack-grounded surface), ADR-0050 (COMPARISON), ADR-0052 (teaching-grounded surface), ADR-0063 (cross-pack resolver), ADR-0064 (cross-pack teaching), ADR-0068..0072 (register subsystem) Composes against: the register axis (ADR-0068..0072) as an orthogonal axis — register varies surface text, anchor lens varies the proposition itself.

Context

The register subsystem (R1–R5) closed the presentation axis: same truth, different wording. That is cosmetic by design — every register holds grounding_source and trace_hash byte-identical across {neutral, terse, convivial} for the same input (core demo register-tour asserts this turn-by-turn).

The original deterministic-AI critique CORE answers (identity informs doing, truth is coherent) asks for something stronger: the system saying genuinely different things from different conceptual substrates, not the same thing in different clothes. That axis is substantive variation.

The substrate to deliver it already sits half-built on disk.

packs/data/grc_logos_micro_v1 (11 entries, alignment), grc_logos_cognition_v1 (20 entries, no alignment), he_logos_micro_v1 (9 entries, alignment), he_core_cognition_v1 (20 entries, no alignment) — all four mounted by default in RuntimeConfig.input_packs.
Cross-language binding is shared semantic_domains atoms, not transliteration tables: the same load-bearing tag (e.g. logos.aletheia.verity) appears across grc/he/en, and the alignment file binds entry-id pairs with evidence ids. This is the deterministic pivot an anchor-lens composer would traverse.
allow_cross_language_recall=True already; cross-language generation is deliberately gated off.

Three honest gaps separate the substrate from anchor-lens-driven surfaces:

Distinction-bearing breadth absent. Greek's interesting move is splitting one English concept into multiple lemmas (γνῶσις / ἐπιστήμη / σύνεσις; ἀγάπη / φιλία / ἔρως / στοργή; αἰών / χρόνος), but the cognition packs carry one lemma per concept. Hebrew distinction-bearing families (חסד, שלום, צדק) absent. Without these families, anchor lens cannot render "knowing-as-experience grounds knowing-as-system."
No teaching corpus for non-English. teaching/cognition_chains/, teaching/relations_chains/, teaching/cross_pack_chains/ are English-only. Anchor lens needs reviewed chains traversing the substrate's distinctions, not bare lexicon.
No realizer infrastructure for cross-lingual surface composition. Today's composers assume English lemmas. The anchor-lens composer must render Greek/Hebrew distinctions deterministically and ideally in English compound phrasing (e.g. "knowing-as-experience") rather than emitting raw non-English glyphs at a layperson.

The register subsystem solved a similar inside-out arc in five phases (R1 class → R2 wiring → R3 first knob → R4 seeded variation → R5 operator surface). Anchor lens warrants the same discipline: the architecture is sound but the content prerequisites are real, and entangling code work with bulk pack ingestion is the failure mode this ADR exists to prevent.

Decision

L1 ships an anchor lens as a fourth pack-layer sibling (identity / safety / ethics / register / anchor lens), delivered inside-out across four phases mirroring the R1–R5 cadence. Each phase lands its own ADR; this ADR is the umbrella that names the architecture and binds the phases together.

Per feedback-no-timelines: phases are sequenced by prerequisites, not clock-time. No phase starts until its predecessor's invariants are pinned in CI.

Architectural commitments (all phases)

Anchor lens is the substantive axis. Register holds grounding_source and trace_hash byte-identical across choices; anchor lens deliberately moves both — different substrate ⇒ different proposition ⇒ different trace_hash. The R5 register-tour seam claim does not extend to anchor lens. A new tour (core demo anchor-lens-tour) asserts the opposite claim: trace_hashes_distinct_across_lenses and surface_propositions_distinct_across_lenses.
Anchor lens does not emit non-English glyphs. The user-facing surface is composed in English compound phrasing ("knowing-as-experience", "knowing-as-system") derived from the non-English lemma's semantic_domains atoms. Greek/Hebrew glyphs appear only in audit / trace / provenance fields, never in ChatResponse.surface. This protects the layperson surface contract (core demo conversation) and matches existing allow_cross_language_generation=False policy.
The pivot is semantic_domains, not lemma strings. Anchor lens never resolves on "γνῶσις" literally; it resolves on logos.epignosis.experiential_knowledge (or whatever atoms the ratified pack content commits to). This keeps the seam language-neutral so future substrates (Sanskrit, Aramaic) compose without touching anchor-lens code.
No new mutation surface. Anchor lens packs are ratified just like register / identity / safety / ethics packs: JSON + companion .mastery_report.json self-seal. Pack mutation stays proposal-only.
Replay equivalence holds within a lens. Same input × same anchor lens × same packs ⇒ byte-identical surface and trace_hash. Replay equivalence does NOT hold across lenses — the whole point of anchor lens is that it changes the proposition.

Phase L1.1 — Content phase (pack enrichment, no code change)

Prerequisite for every later phase. Enrich the substrate so distinction-bearing lemmas exist on disk:

grc cognition pack additions (target: full distinction families for the load-bearing English collapses):
- Knowledge family: γνῶσις (present), ἐπιστήμη (systematic knowledge), σύνεσις (insight), σοφία (wisdom, present).
- Love family: ἀγάπη, φιλία, ἔρως, στοργή.
- Time family: αἰών (age, era), χρόνος (clock time), καιρός (opportune moment).
he cognition pack additions: חסד (covenant-love), שלום (peace-wholeness), צדק (righteousness-rightness).
alignment.jsonl on the cognition-tier packs (currently only the micro packs carry it). Binds new lemmas to their English / cross-lang counterparts where they exist, and records relation: cross_lang.no_english_collapse for families English doesn't split.
Distinguishing semantic_domains atoms per family member — e.g. logos.epignosis.experiential_knowledge, logos.epignosis.systematic_knowledge, logos.epignosis.insight — so the anchor-lens composer has ratified atoms to render against.

No realizer / runtime / composer code in L1.1. The phase deliverable is enriched lexicon + alignment + ratified self-seal. This is the highest-leverage step because it unblocks every later phase without touching code at all.

Phase L1.2 — Pack class + loader (architectural class, no consumer)

Once L1.1 ratifies the substrate, introduce the anchor-lens pack class. Mirrors the register pack class (ADR-0068):

@dataclass(frozen=True, slots=True)
class AnchorLens:
    lens_id: str                         # e.g. "grc_logos_v1"
    primary_substrate: str               # "grc" | "he" | "en"
    semantic_domain_preferences: tuple[str, ...]
                                         # ordered atoms; left-most wins
    cognitive_mode_label: str            # rendered in English compound
                                         # phrasing — "experiential" / "systematic"
                                         # / "covenant-bound" / etc.

@classmethod
def unanchored(cls) -> "AnchorLens": ...
    # In-memory sentinel.  Lens-aware composers MUST treat it as
    # structurally identical to "no lens applied" — byte-identical
    # to pre-L1.2 behaviour.

L1.2 ships the class, the loader, and one ratified pack (default_unanchored_v1 — structurally equivalent to the unanchored sentinel). No composer consumes the lens yet. CI pins:

anchor_lens_byte_identity_null_lift — full lane unchanged whether RuntimeConfig.anchor_lens_id is None or default_unanchored_v1.

Phase L1.3 — First non-trivial lens + composer wiring

L1.3 ships the first lens that actually moves the proposition:

grc_logos_v1 — primary substrate grc; domain preferences pivot knowledge-family queries onto logos.epignosis.experiential_knowledge rather than the en-default cognition.knowledge.
he_logos_v1 — primary substrate he; pivots claim-family queries onto logos.aletheia.covenant_verity rather than cognition.truth.

chat/pack_grounding.py gains an AnchorLens parameter (mirroring the R3 register parameter pattern). Composers consult the lens's semantic_domain_preferences to pick which lemma to render before constructing the surface string. Rendering stays English compound: e.g. "Knowledge-as-experience is grounded in personal acquaintance, not abstract system."

CI pins:

anchor_lens_lifts_proposition — same prompt × {unanchored, grc_logos_v1, he_logos_v1} produces three different propositions (surface differs structurally, not just lexically; trace_hash differs deliberately).
anchor_lens_no_glyph_leak — non-ASCII characters are absent from ChatResponse.surface regardless of lens (audit / provenance fields are unrestricted).

Phase L1.4 — Telemetry + operator surface + tour demo

Mirrors R5. TurnEvent + ChatResponse gain anchor_lens_id. Telemetry serializer surfaces it. core chat --anchor-lens <id> flag. core demo anchor-lens-tour walks the same prompts under {unanchored, grc_logos_v1, he_logos_v1} and asserts:

lens_ids_recorded_per_turn — TurnEvent populates anchor_lens_id on every turn.
trace_hashes_distinct_across_lenses — the opposite of register-tour's claim, and load-bearing here.
surface_propositions_distinct_across_lenses — at least one prompt yields a structurally-different proposition across lenses (not just marker variation).

Exit code 0 iff every claim holds.

Consequences

Capability unlocked at L1.4

CORE produces structurally different propositions from different conceptual substrates, deterministically. A Greek-anchored answer to "what is knowledge?" does not rephrase the English answer; it answers from γνῶσις-vs-ἐπιστήμη as compound English phrasing and produces a proposition that English-default would have collapsed. Combined with register (ADR-0068..0072), the matrix is {substantive lens} × {presentation register} — orthogonal axes.

Trust boundaries

Lens packs ratified like every other pack class. Mastery reports self-seal; loader rejects unratified ids in production mode (per ADR-0027 pattern).
Glyph escape is not a silent regression. L1.3 invariant anchor_lens_no_glyph_leak is a hard gate: a non-ASCII character in ChatResponse.surface fails the lane.
No new mutation surface. Anchor lens consumption stays proposal-only at the pack-mutation layer.

Cognition lane

L1.1 (content only) and L1.2 (null-lift unanchored sentinel) are byte-identical. L1.3 deliberately moves cognition-eval numbers when a non-default lens is loaded; the public-split byte-identical contract holds for the unanchored default.

Backwards compatibility

RuntimeConfig.anchor_lens_id: str | None = None — defaults preserve pre-L1 behaviour.
TurnEvent / ChatResponse fields default to "".
Existing register / identity / safety / ethics interactions unchanged — anchor lens composes upstream of those layers (the composer chooses which proposition to construct; register then varies the surface text of that proposition).

Performance

Negligible. Lens lookup is one tuple-iteration over semantic_domain_preferences; no SHA-256 hash per turn (variant_id discipline is register-specific; anchor lens identity is the lens_id itself).

Verification (L1.4 acceptance gate)

tests/test_anchor_lens_pack_loader.py            N passed
tests/test_anchor_lens_null_lift.py              N passed
tests/test_anchor_lens_runtime_threading.py      N passed
tests/test_anchor_lens_proposition_lift.py       N passed
tests/test_anchor_lens_no_glyph_leak.py          N passed
tests/test_anchor_lens_telemetry.py              N passed
tests/test_anchor_lens_tour_demo.py              N passed

Curated lanes (must remain green):
  smoke / cognition / teaching / packs / runtime / algebra

Cognition eval byte-identical under unanchored:
  public 100 / 100 / 91.7 / 100
  holdout 100 / 100 / 83.3 / 100

core demo register-tour                          exit 0 (still holds)
core demo anchor-lens-tour                       exit 0

The register-tour claim must continue to hold under any anchor lens — within a fixed lens, register choice still preserves grounding_source and trace_hash. The two axes are genuinely orthogonal, not just architecturally claimed.

Open questions deferred

Composing anchor lens × register simultaneously in the tour. Two-axis demo (lens × register × prompts) is a natural follow-on but inflates the grid 9×; L1.4 ships single-axis lens tour first, composition tour later.
Lens × identity interaction. Does a precision_first_v1 identity bias propositions differently under grc_logos_v1 than under he_logos_v1? Probably yes (axis-hedge phrasing composes through), but documenting it needs L1.3 evidence first.
Cross-language teaching corpus. L1.3 lemma-tier lensing is a prerequisite for grc/he teaching chains; the chains themselves are a separate phase (L2+, ADR not yet drafted).
More substrates beyond grc/he. Sanskrit (jñāna/vidyā/prajñā) and Aramaic distinctions sit naturally on the language-neutral-pivot architecture but are deferred to post-L1 ADRs after the two-substrate case has shipped.

Sequencing summary

L1.1  content phase — pack enrichment + alignment           ADR-0073a (this scope)
L1.2  AnchorLens class + loader + unanchored sentinel       ADR-0073b
L1.3  first non-trivial lenses + composer wiring            ADR-0073c
L1.4  telemetry + operator CLI + anchor-lens-tour demo      ADR-0073d

Each sub-ADR ratifies its phase in isolation, in the R1–R5 style. This umbrella ADR-0073 names the architecture; the sub-ADRs commit to specific lemma lists, atom names, and pack ids. Substrate enrichment (L1.1) is the prerequisite that unblocks everything; nothing in L1.2–L1.4 starts until L1.1 ratifies.

Why this composes against register, not replaces it

Register operates after the proposition is fixed — it only varies the surface string the realizer emitted (post-composer decoration, ADR-0071 R4).
Anchor lens operates before the proposition is constructed — it picks which lemma's semantic_domains the composer should render, which changes the proposition itself.

This means register-tour and anchor-lens-tour test opposite invariants and both must pass:

register-tour   asserts:  per prompt, fixing lens, varying register   ⇒ trace_hash CONSTANT
anchor-lens-tour asserts: per prompt, fixing register, varying lens   ⇒ trace_hash DISTINCT

If either claim ever fails the seam is broken — the orthogonality is the load-bearing architectural commitment.

16 KiB Raw Permalink Blame History Unescape Escape