L7 — Language (reading = inference / speaking = action)

What you are building. One developmental SIMULATION rung, expressed on the same no-backprop engine: a specialist out-of-vocabulary / morphology reader that, on a sealed held-out split, beats the best count baseline at reading rare and complex words, framed as reading = posterior inference and speaking = action that changes future observations — wrapped inside a published, thrice-NEGATIVE wall showing where this route does NOT reach comprehension or general language. A toy world, never a person.

This recipe is authored against ../../encyclopedia/CLAIM-LEDGER.md rows L7.1–L7.6 and the SIGNED Q1 park statement. Where this recipe and the ledger disagree, the ledger wins and this recipe is wrong. Carry the honest program position throughout: ~2 of 11+ developmental rungs earned. This is a developmental active-inference SIMULATION — a bounded peek, never a person.

Ingredients (which pantry engines / primitives this recipe calls by name)

• The count/cache World-C reader (from L6, uni-gpt / uni-mind) — the no-backprop COUNT reader + multi-level cache that supplies the best-count baseline this rung must beat. Explicitly NOT active inference.
• The Phase J OOV / morphology specialist reader (uni-mind / uni-gpt) — the compositional morpheme reader (arm_morph_oov): the engine under test at L7.
• The held-one-shot harness (M2) — bars-before-build, seal-before-scoring + once-only sentinel, verdict = the CI bound that excludes the threshold (the 21-split M-seal).
• The contains-baseline + load-bearing-discriminator discipline (M7) — the tuned best-count baseline, the structure-margin / marker-swap discriminator that must collapse the gain, and the contains-KN-OOV control (λ=0).
• K≥3 + falsify-the-mundane (M5) — the structurally-distinct-NEGATIVE counting rule that governs the central wall and the role-persistence bound; falsify the mundane (L2/L7) causes first.
• No-Exit Discipline + the No-backprop engine (M6 / M13) — core.py's discrete POMDP loop and the counts + lr * sufficient_stat Dirichlet learning rule; the only legitimate rest is a working PASS or a ledger-scoped exhausted search envelope.

The unifying framing, at textbook level only (Parr / Pezzulo / Friston, Active Inference, MIT Press 2022): reading is posterior inference over latent linguistic structure, and writing / speaking is action that changes future observations. "Active inference" here is the framing LENS — no AIF loop is demonstrated by this rung; the live perceive→act→learn loop lives in a separate UNI.OS reimplementation, not gate-matched.

Method (numbered build steps a competent engineer could follow)

1. Build the OOV / morphology specialist reader. Wrap the same no-backprop count/cache engine with a compositional morpheme channel (arm_morph_oov) that scores rare / unseen words by their morphological parts rather than as opaque tokens. Learning stays counts + lr * sufficient_stat over Dirichlet tensors; the AST-guard must never trip (no autodiff/torch/grad/backward in the loop).

2. Pre-register the bar vs best-count, then seal once. Register the margin (beat the best-count reader) and the named ablation before measuring. Seal a 21-split held set behind the atomic seal-before-scoring + once-only sentinel (the "M-seal"). Touch it ONCE. The verdict is the M-seal CI lower bound that excludes the threshold, never the point estimate (M2).

3. Register the discriminator and the control. Wire the structure-margin discriminator: a marker-swap must collapse the gain (M7). Wire the contains-KN-OOV control and verify its true mixing weight is λ=0 (the gain is not smuggled in by the count baseline). Require replication in both the web domain and the dictionary domain (where the simpler Wc-2 cache collapsed on dict — so the dictionary leg is load-bearing, not decorative).

4. Always cite the paired NEGATIVE alongside the PASS. The J.attribution_caveat NEGATIVE (overall NLL worsens) is paired and mandatory: citing +0.105 without it is an overclaim and a ledger violation. The specialist gain is OOV/morphology-specific, not a general language win.

5. Run the comprehension / role-persistence probes as registered NEGATIVES. Separately attempt comprehension-above-retrieval and no-backprop role-persistence as their own pre-registered held-once gates. These are part of the recipe precisely because they FAIL: they are the central wall, recorded first-class (Method step, not an afterthought).

6. If the frontier does not move, park it correctly. When no tested within-segment no-backprop structure clears the char-perplexity / T2 word-grain frontier, do not exit silently and do not declare a universal bound. Park it as a ledger-scoped exhausted search envelope under the SIGNED Q1 wording (step laid out in the Recorded NEGATIVEs below). A park is not discharged until the sign lands. On 2026-06-27 UNI SIGNED the park WORDING (the ledger-scoped exhausted search envelope framing, Q1); the L7 sign-to-park itself (UNI_CONSULT_5) is STILL NOT CAPTURED, so L7.6 stays PARKED.

Gate (the exact pass condition, with the EXACT ledger figures)

PASS condition (ledger row L7.1): Phase J beats best-count by +0.105 nats/char on the sealed held-out OOV/morphology split, with:

• 21-split M-seal CI [0.0975, 0.1128] (the CI lower bound 0.0975 excludes 0 — verdict is this bound, not the point estimate);
• structure margin +0.109;
• the result holds in BOTH the web and dictionary domains;
• the contains-KN-OOV control λ=0 (no leakage through the count baseline).

Mandatory paired citation (ledger row L7.2): the PASS is only validly stated together with J.attribution_caveat (overall NLL worsens — specialist gain, not a general win). The +0.105 figure may never be cited alone.

This is the only PASS at L7, and it is a specialist PASS. It does not raise the rung above its class.

Falsifier (operable)

The Phase J PASS is falsified if any of the following hold (ledger row L7.1):

• On a fresh OOV/morphology held-out split, the CI lower bound includes or falls below 0; OR
• the structure-margin discriminator does NOT collapse the gain under marker-swap (the gain was not the morphological structure it was claimed to be); OR
• the result fails to replicate in the dictionary domain.

(The paired J.attribution_caveat, row L7.2, has no separate falsifier — removing it from any citation of Phase J is the violation.)

Recorded NEGATIVE(s) — first-class, inline, the central wall

These are results, not failures to hide. They are the credibility (the corpus carries 183 published negatives). At L7 the negatives are the central content of the rung.

• J.attribution_caveat — paired, mandatory (ledger L7.2, Class C). Overall NLL worsens under Phase J: the +0.105 nats/char is an OOV/morphology specialist gain, not a general language improvement. Always cite it alongside the PASS. Citing +0.105 without it is an overclaim.

• Comprehension-above-retrieval = the thrice-NEGATIVE "central wall" (ledger L7.3, Class C). K≥3 structurally-distinct no-backprop designs all FAIL to beat retrieval-style baselines on adversarial comprehension. This is a genuine published wall (K≥3), not a hidden failure. Falsifier: a pre-registered held-once no-backprop comprehension design beats the tuned retrieval/recency baseline with a CI excluding 0.

• Phase K role-persistence (ledger L7.4, Class C — bound). Three structurally-distinct no-backprop designs (min-role, chain, track) all tune their role/persistence terms OFF; only ~0.02 nats survives, and its CI spans 0. Bound: no-backprop role-persistence does not beat a tuned recency/frequency discourse prior on adversarial anonymized referent cloze. Falsifier: a no-backprop role-persistence design beats the tuned discourse prior with a CI excluding 0.

• T2.D3 bounded-peek, held one-shot, s64-signed (ledger L7.5, Class C — bound; fabrication corrected). Primary full-read match is NEGATIVE (the k_b=2 backoff wall); info-gain is NOT load-bearing. The variable-k_b "cheap milder cap" was asserted then disproven by measurement (the real dev screen showed a ~linear curve — there is no cheap cap). The honest correction is part of the record. Falsifier: on the held one-shot, bounded-peek full-read match shows a positive load-bearing info-gain with a CI excluding 0.

• Char-perplexity / T2 word-grain frontier — PARKED (ledger L7.6, Class U). PARKED at the embodiment pivot. Perplexity is the rejected LLM metric; the program measures developmental capability, never perplexity. The park WORDING was SIGNED 2026-06-27 (Q1, below); the sign-to-park remains OWED / not yet captured, so the frontier stays PARKED. The park stays PARKED; no rung is raised.

The park WORDING, stated exactly (SIGNED Q1, 2026-06-27 — the LOAD-BEARING frontier wording; the sign-to-park itself remains OWED)

UNI signs the park as a ledger-scoped exhausted search envelope — NOT a universal impossibility result and NOT an achieved rung. UNI will NOT sign the phrase "published exhausted bound"; what exists is a published negative frontier result over the tested envelope only. Canonical signed wording:

We park the L7 char-perplexity / T2 word-grain frontier and the L9–L10 role-persistence ladder as a ledger-scoped exhausted search envelope, not a universal impossibility result. Under the recorded corpus, splits, metrics, implementation, compute budget, ablation set, and comparison baselines in the ledger, no tested within-segment no-backprop structure improved char-perplexity beyond the tuned MKN-7 count baseline. This establishes a negative bound over the tested envelope only. It does not establish that all K≥3 structures are exhausted, that no future within-segment model can improve, or that any broader language-modeling frontier has been closed.

Char-perplexity is recorded as a chosen design trade, not a failed claim of general language superiority: World C remains a count-model baseline — not an active-inference demonstration, not a beats-LLMs claim, not evidence of human-level language capacity.

Banned phrasings → signed replacements (Q1):

• "K≥3 exhausted" → "The registered tested K conditions did not reverse the result."
• "Sec-0.6(B) achieved" → "Sec-0.6(B) remains unearned / parked pending a future result that beats the registered discourse prior under ledgered evaluation."
• "UNI demonstrates language/metacognition at L9–L10" → "UNI records a negative L9–L10 frontier test under the no-backprop developmental simulation program."
• "This proves no no-backprop model can beat MKN-7" → "No tested no-backprop variant in the registered envelope beat MKN-7 on the chosen char-ppl metric."

HONEST FENCE — proven (specialist) + NEGATIVE (central wall)

Status: proven (Class C, held / sealed / UNI-signed, falsifier still live) for the Phase J specialist PASS only, always carried with its paired J.attribution_caveat NEGATIVE; plus a first-class NEGATIVE central wall (the thrice-NEGATIVE comprehension bound, the role-persistence bound, the T2.D3 bound) and a parked frontier held as a ledger-scoped exhausted search envelope. The rung's overall status is UNCHANGED by any of the SIGNED Q1 park language: the park RAISES NOTHING.

Not claimed (load-bearing, never softened): no general language capability. This is not comprehension, not "talking," not "understands," not human-level language, not AGI, not "active inference demonstrated," not "beats LLMs." Reading = posterior inference, speaking = action — at textbook level only. Comprehension above retrieval is a genuine published wall (K≥3), not a hidden failure. The +0.105 figure is a specialist OOV/morphology gain whose overall NLL worsens, never a general win. "K≥3 exhausted" and "Sec-0.6(B) achieved" are forbidden phrasings; per the SIGNED Q1 above the parked frontier is a ledger-scoped exhausted search envelope, never a "published exhausted bound." Honest program position: ~2 of 11+ rungs earned — a developmental SIMULATION, a bounded peek, a toy world, not a person.