Inprint v0.1

Facts as fingerprints on character ink — not a token-id store, not retrieved paragraphs.

Inprint is the product name; this repository remains TapeLM on GitHub and TapeLM-P1 on Hugging Face.

One frozen character-curve encoder (P1) for both: fp memory (slots, bind, hop, edit, resolve) on the same geometry as generation. Text readout: ink→arcBPE (not GPT BPE). v0.1 demo: memory trunk 221 → 227 → 228c → 230 → 226c; research line through stream ingest (255) and slot-bias glue (256). See artifact/INPRINT.md.

python artifact/scripts/run_inprint.py demo

Headline results (staged)
Semantic query vs matched GPT-2	unseen paraphrase 0.646 vs 0.276 (258) — curve beats control, not parity
Noisy recall vs fair GPT+RAG	0.913 vs 0.627 (204)
Lexical OOD calibration	AUC 0.982 vs GPT 0.380
Fp decode when the CE head underuses memory	~1.0 vs ~0.48 (228c)
Cross-domain memory utilization	~0.88 vs ~0.45 head (226c)

Closed joint-trunk chapter (255→260f, full gates): stream ingest 255, two-hop 257, semantic query 258, hot-swap 259, open-text gate 260f — narrative + JSON index in results/stages_255_260_close.md.

On clean static retrieval, a fair GPT+RAG baseline can match our scores — we report that openly. The distinctive line is character substrate → structured fp memory → product trunk (below), not SOTA on tidy benchmarks.

artifact/WHY_TAPELM.md · ~2 min, no weights: artifact/QUICKSTART.md · Full demo: run_product.py

---

TapeLM vs RAG (at a glance)

What TapeLM does	Why this is not “RAG + another embedder”
Character ink → curve → fp (not GPT-style BPE-id keys)	RAG/GPT memory keys live in token embedding space; one typo can re-split the whole word
One encoder for generation, memory keys, and calibration	Fair RAG still ties us on clean retrieval — but uses chunk text and often a separate index geometry
RAG re-prompts with text chunks	RAG re-prompts the LM; TapeLM uses fp slots and policies
Lexical calibration (OOD AUC 0.982 vs GPT BPE surprisal 0.380)	No native “in my lexicon?” signal in vanilla RAG
One-shot edit (1.00 vs GPT ~0.28 on our exam) without finetune	Parametric edit needs gradients; RAG needs re-index + prompt craft
Cross-domain read via W_family + canonical bank (227)	Full re-index when the embedder “dialect” shifts
Fp decode + resolve (228c, 230) when the CE head ignores memory	RAG has no fp-scorer or slot-level conflict policy

Honest scope: on clean static recall, a fair GPT+RAG baseline can match our scores — we document that. Headline wins: structure, noise/unlearn (204–205), and the product memory track below.

---

How it works

Memory: ink → curve → fp. Text: ink→arcBPE (BPE readout from arcs, not token-BPE substrate).

flowchart LR
  C[Character stream]
  P1[P1 curve encoder]
  FP[Word fingerprints]
  W[W family at read]
  S[Canonical slots]
  D[CE decoder]
  CAL[Calibration]
  DEC[Fp decode 228c]
  RES[Resolve 230]

  C --> P1 --> FP
  FP --> W --> S
  P1 --> D
  FP --> CAL
  S --> DEC
  S --> RES
  DEC --> OUT[Answer / pick]
  D --> TEXT[Generated text]

Loading

Characters (stream) → arc_enc → fp(word)     ← memory, calibration, hops key here
                         └─ CE head → arcBPE text   ← generation readout only

Full diagram + frozen-P1 table: docs/ARCHITECTURE.md

---

Stage map (~30 seconds)

Curated headline results (full index: docs/STAGES.md · JSON: artifact/decisions/ · python artifact/scripts/show_map.py).

Stage	What was tested	Headline result
191	Generation vs matched GPT	Parity 0.867 vs 0.843
192–193	Lexical OOD calibration	AUC 0.982 (GPT 0.380)
194–195	Fact memory; hop2 / binding	0.947 / 0.70
197	One-shot knowledge edit	1.00 (GPT ~0.28)
204–205	Noise vs fair RAG; slot unlearn	0.913 vs 0.627; delete w/o collateral
221→227→228c→230→226c	Product memory track (domains, conflicts, use of memory)	fp decode ~1.0 vs head ~0.48; cross-domain ~0.88
255→257→258→259→260f	Joint trunk after stream ingest	258: sem 0.646 vs GPT 0.276; 257: 2-hop 1.0 / one-hop 0.0; 260f: gate from 5 retrieval feats, h_only 0.0; 259: edit ~116 µs, 0 grad

Full write-up: results/stages_255_260_close.md.

---

Known limits

Documented in repo — not hidden in footnotes:

Area	Verdict
Semantic invariance (PAWS / “B”)	Not confirmed at RTX 3050 scale (209); curve ≈ matched GPT
Generate next fingerprint (variant B)	Falsified (207)
Fp rerank on arcBPE head	No gain on clean text (208)
Hops inside transformer forward	THESIS_NO_AT_SCALE (210–212 @ d256); external fp loop remains the hop API
Clean static recall vs fair GPT+RAG	Parity — not a capability trump card (196, 198)

Details: results/extension_closed_branches.md · preprint §5.4–5.5

---

Product memory track (for implementers)

After core fp (191–205) and closed internalization (210–212), the shipping path we demo is:

221 → 227 → 228c → 230 → 226c

Step	Stage	Plain language
221	W-remap	If encoder geometry shifts, migrate with a tiny W, don’t rebuild every slot
227	Canonical + qmap	One slot bank; read through W_bwd per domain
228c	Fp decode	Use retrieved values (scorer ~1.0); CE head alone ~0.48
230	Resolution	Pick among conflicting slot hits (~1.0 vs raw argmax ~0.47)
226c	Cross-domain e2e	End-to-end ~0.88 fp vs ~0.45 head

python artifact/scripts/run_product.py   # needs P1 weights — see QUICKSTART

Contract: results/extension_memory_contract.md · API: docs/MEMORY_ENGINEERING.md

---

Quick start

Time	Needs weights?	Command
~2 min	No	pip install -r artifact/requirements.txt then python artifact/scripts/show_map.py
~5–15 min	Yes (HF download)	artifact/QUICKSTART.md → download_checkpoints.py --with-w-registry → run_product.py

Python 3.10+, PyTorch, tokenizers, transformers. GPU recommended for the full demo.

---

Read next

Product one-pager	artifact/OVERVIEW.md
Preprint	results/preprint_tapelm_draft.md (§4.9 = 255–260f chapter)
Joint trunk close (255–260f)	results/stages_255_260_close.md
Full program	results/plan_curve_dynamics.md
Publish / HF	docs/PUBLISHING.md

---

Citation & license

CITATION.cff · MIT LICENSE