OMEGA-Ω

0Methods & conventions

0.1 — Status of claims

Every load-bearing fact is labelled. The scientific value of the project rests on this strict separation:

• measured reproducible empirical result (paired headless run, fixed seeds), dated source.
• design settled architectural choice, not re-argued in this version.
• hypothesis interpretation or conjecture not yet settled by measurement.
• excluded deliberately outside the decision chain (gray cheating or crutch).

0.2 — Three metrics, not to be confused

0.3 — Measurement protocol

• Seeded and paired comparisons (same words, same seeds); A/B/C/D conditions.
• Typical N: 150–2000 games; 8 paired seeds for fine deltas (z statistic).
• Chained games, never reset between plays: learning is inter-game (principle L04). Measuring with a reset invalidates any cumulative mechanism measured (the "30 games × resets" measurement had skewed the evaluation of F177).
• Inertia OFF: a toggle is cleanly disconnectable if, switched off, the output is byte-identical to the baseline.

0.4 — Reproducibility caveat

Fixed in phase46 (Fix#1): the word picker and the bench helpers now go through a seeded omegaRand() → deterministic headless measurements (bit-for-bit repeatable). Only the intentional fallback to Math.random() remains when no seed is set (free interactive play). The "robust" figures come from the seeded paired headless protocols.

0.5 — Version note (read once)

1Object & framing

OMEGA-Ω is a cognitive architecture that plays Hangman on a hyperdimensional-computing (HDC) substrate. The goal is not to maximise wins but to make the decision emerge (the choice of a letter) from brain-inspired mechanisms.

design Category distinction (internal ref. F169): a solver maximises wins by any means (n-grams, full dictionary); a cognitive architecture that plays Hangman demonstrates learning under impoverished constraints. OMEGA is the latter. Per F169, 30–40 % wins without a lexicon would already be a good result (the human cognitive limit on word prediction).

2Overview & performance

A self-contained web application (a single HTML file) bundling the cognitive engine, the French dictionary and the measurement dashboard. The v0.07 rebuild was motivated by the separation of the spelling/phonology substrates: their cross-contamination made the routes degenerate — an unintended reproduction of surface-dyslexia errors (Plaut & Shallice 1993).

2.1 — Performance: sourced figures

The rate depends on difficulty, configuration and horizon. Always state the regime.

2.2 — Founding principles (L01–L04)

• L01 — double Möbius design: two antagonistic circles (direct = perception→decision; mirror = result→correction) linked by the OS.
• L02 — functional invariance: concepts emerge above the substrate; the 12 concept cells of M3_d are learned (Hebbian), not hard-wired.
• L03 — OS as dialectical governor: keeps the tension between the two circles without letting either dominate (a construction specific to the project; neighbours: GAN, adversarial RL).
• L04 — inter-game learning: the mirror and the sound↔letter map persist across games; no reset.

2.3 — Four-level learning cascade (F139)

1. Phonetic primitive (hard-wired): 14 articulatory features × 26 letters + an HRR ortho component (0.85 / 0.15).
2. Emergence of a proto-language: 5 sub-brains of M1_d produce vocabulary.
3. Sound↔letter map (phonGraphMap[26][SDIM]): learned by playing (observation + consolidation).
4. Hangman: the testbed; decision = comparison of the internal state against the map (and/or the M5_d softmax).

3The task: Hangman (specification)

• Words: length ≥ 7 letters; lexicon 88,939 words (Lexique 4, ≥ 7 subset, upper-cased/deaccented).
• Start: 0 letters revealed ("hardened" Hangman).
• Budget: 6 errors.
• Legitimate information: revealed letters, length, tried letters, knowledge gained through experience.
• Illegitimate information: the hidden word — directly, or by enumerating the dictionary down to the single compatible word.

Hangman is chosen for its context poverty: the player must actively generate hypotheses (like "sounding out" an unknown word). A demanding and revealing testbed.

4Architecture

Two parallel columns, a conceptual bridge, an arbitration module.

• Orthographic route (M*_d): revealed letters → feature → spatial pattern → concept → lexical candidates → decision.
• Phonological route (M*_phon): conversion into SAMPA phonemes (distinguishes é/è/an/on…), a second reading route (Coltheart 2001).
• M_S hub (shared concept): an amodal crossroads, cross-modal auto-encoding (Rogers 2004; Patterson 2007) — enter through one modality, reconstruct the others.
• OS — arbitration module (L03): combines ortho + phon via w(r) = −r/(1+r) (r = phon strength ÷ ortho strength), decides the letter, dilutes any signal pre-imposing an answer. Publishes OS_diss, OS_lect, OS_mod, OS_speed.

4.1 — Structural layers

• Pipeline 0 (statistical priors): in v0.06, three modules M_E (episodic memory), M_A (letter priors + reward shaping), M_P (pattern matching). measured Absent from the v0.07 build but absorbed: priors → the LETTER_FREQ constant; pattern → A2/computeLex4LetterScores; episodic ≈ the game itself. v0.07 consolidated, it did not lose.
• Typed channels: ~32 channels (CONCEPT, VECTOR, EVENT, SCALAR) connect the modules; some cross-route links are dormant.
• Independent substrates: ortho 26D (identity), phon 26D one-hot + 14D features, cross-modal HRR hub — no derivation chain between substrates (a key v0.07 fix).
• Möbius loop (P1/P2) design: a designed closed circuit (P1 direct, P2 mirror, P1→P2→P1). Receiving the word in P2 is by design (simulation, Hesslow 2002), with the OS handling anti-cheat. In practice P2 and the extended mirror modules are disconnected (see §12).

Diagram — two routes, two loops, one arbiter

                       ┌────────────────────────────┐
                       │    OS — arbitration module │
                       │   combines ortho + phon     │
                       │   w(r) = -r / (1 + r)       │
                       └────────┬──────────┬─────────┘
                 ortho decision │          │ phon decision
        ──────────────────────  │          │  ──────────────────────
            ORTHO ROUTE (M*_d)   │          │     PHON ROUTE (M*_phon)

        bottom-up↑  mirror↓                 bottom-up↑  mirror↓
        M5_d  ──►  M5_m  (reward)           M5_phon ─► M5_phon_m (reward)
         ▲           │                       ▲          │
        M4_d ◄┄ cross p ┄┄┄┄┄┄┄┄┄┄┄┄┄┄┄┄┄┄ M4_phon ─► M4_phon_m   effective
         ▲      (M4_PHON_USE_P)              ▲          ┆
        M3_d ◄─ written ─ M3_m (a-Hebb)      M3_phon ┄► M3_phon_m   observational
         ▲     (conceptCells)                ▲
        M2_d  ──►  M2_m                      M2_phon      M2_phon_m  absent
         ▲           │                       ▲
        M1_d ◄─ ≈0.1 ─ M1_m (letScore)       M1_phon      M1_phon_m  absent

  Ortho mirror = complete (5 stages) · phon mirror = truncated (M5→M4; M3 obs; M2/M1 absent).

5Modules M1–M5, loops, constants

5.1 — The five stages

Clarifications: the 16 zones of M2 are dimensional (compartments of the vector), not positional. The M5 temperature tunes exploration (clear gap → decisive choice; small gap → exploration).

5.2 — Two loops per route

• Bottom-up (M1→M5): perceives and decides; reads the word only at revealed positions (cap §43); never accesses hidden letters.
• Top-down (mirror, M5→M1): receives the result, adjusts the weights; teaches, does not decide.

Mirror wiring. Ortho complete: M5_m (reward) → M4_m (letterPenalty) → M3_m (anti-Hebbian, writes conceptCells) → M2_m (zonePenalty) → M1_m (letterScore). M3 shares its weights (the only stage where the mirror writes); M4/M2/M1 have a separate state fed back in; M1_m.letterScore may co-decide M5 with weight M5_D_M1_M_WEIGHT = 0.1 (F198), but this is now gated by the toggle M5_D_M1_M_ENABLED = default OFF (decision 17/06, AUDIT_OMEGA §1.4.2: M1_m does not earn its 0.1) → in baseline M1_m does NOT influence the decision; the assert only pins the weight (0.1) if it is re-enabled. Obsolete: "0.0 baseline" AND "0.1 without toggle". Phon truncated: M5_phon→M4_phon effective (read by bottom-up M4_phon), M3_phon_m observational, M2_phon_m/M1_phon_m built + called but with no consumer (dormant — not "never built"; cf. AUDIT_OMEGA §1.4/D1).

5.3 — Main constants

The ratio α_Hebb/α_decay = 8 is an identified debt (unbounded norm, §12).

6Toggle reference (47)

47 flags in total (phase45): 31 exposed as toggle buttons, plus controls in dedicated panels (dictation, bind bridge) and base JS flags. Phase46 adds 6 NEO toggles + 4 parameters → 37 toggles; phase47 adds the M_TREXQUANT_MODE mode → 38 toggles (§17). UPDATE 06/26: +3 cheat-free blue toggles (M_NEO_PHON_COHORT board-derived sound · M_NEO_PHON_COHORT_JOINTE · M_NEO_OS_ARB OS arbitration) → 41 toggles; M_DECLARE_DUAL adopted (preset). UI colour code, 4 categories: green optimal config · blue optional cheat-free · orange reads the SOUND of currentWord (word heard) · red gray cheating (A1/A2/A3). Detail of the 3 new ones: §17.1. Default = state at load. "preset ✔" = switched on by the reference config (§8).

Note. At startup, only M4_PHON_USE_P ("φ · Croisement p") is ON; the reference config (§8) then switches on the "preset ✔" toggles. The heavily-coupled toggles (M_VOIE_PHON, M_OS_V07, M_BPC_M3D) touch many points in the code — their effect is not local.

7Lexique 4 dossier — the A2 toggle

7.1 — What A2 is

L01_A2_M4_LEX4 (L01 ENH.) operates inside M4_ortho. From the cohort (dictionary words compatible with the pattern + exclusions), it scores each letter by its frequency among the candidates and injects that score into letter scoring — the estimator P(letter ∈ word | board), frequency-weighted. Reference model: the cohort model (Marslen-Wilson & Welsh 1978).

7.2 — Statistics

7.3 — Why A2 is excluded from scoring

A2 reaches ~98.7 % (seen) / 100 % (new) because it has the whole lexicon: a "new" word is never new to it (its cohort always contains the true word and converges on it). This is the oracle-shortcut OMEGA refuses. Its redundancy is measured (it absorbs Cobweb, corr 0.845; it makes the bPC coupling counterproductive). It carries the 95 % of the full config without cognition.

7.4 — What is allowed instead

A2 remains usable only for DECLARE (proposing a word, as a "teacher"), never as a letter-scoring engine. The cheat-free equivalent is M_DECLARE_DUAL (cohort of the board, not of the full dictionary): +5 pts in a cognition config, without reading the dictionary beyond what is board-compatible.

8Manual

8.1 — Launch

1. Open OMEGA_v0_07_phase47…html in any browser: a self-contained monolithic application (code + lexicon inlined). No server, no dependency.
2. Backup freeze OMEGA_v0_07_base.html: a stable state for rollback — do not touch it.
3. Golden rule: the browser console is read-only. Every action goes through a UI button/toggle — never a command that would modify state (reproducibility).

8.2 — The interface panels

8.3 — Reference configuration (cheat-free cognition) — UPDATE 14/06/2026

Optimal preset re-measured on the current code, to be activated in L01 ENHANCEMENTS + Declare NEO:

ON (cognition): A4, A5, A6, M_VOIE_PHON, M_OS_V07, M4_PHON_USE_P,
      M_SUBSTRAT_ORTHO_PURE, M_PHON_FEEDBACK, M_BPC_M3D,
      M_BPC_READOUT_COUPLE, M_PHON_READOUT_COUPLE, M_PHON_CONCEPT_BIND,
      M_OS_LEARNING (+ its 4 guards)   [M_OS_LEARNING_ONLINE OFF: measured no gain + drift]
ON (declare NEO): M_DECLARE_NEO, M_NEO_RECALL, M_NEO_ASSEMBLED,
      M_NEO_COHORT, M_NEO_G2P_EXP (pen 0.5), M_DECLARE_DUAL (adopted 06/26)   (mute/trigger OFF; NEO_CONF 0.75, RECALL_MARGIN 0.20)
cheat-free OPTION WITHOUT currentWord (blue): + M_NEO_PHON_COHORT + M_NEO_PHON_COHORT_JOINTE → ~97.3 % (vs 97.5 % "word heard")
OFF: A1, A2, A3                                (gray cheating / crutch)
TO MEASURE: L01_B2_MOBIUS, M_IG_SELECT, M_IG_PSUCCESS ; M_NEO_OS_ARB (a clean alternative to DUAL, ≈ but does not beat it)

Measured win rate (in-lexicon, seed 12345, 120 words, warmup 150): cognition alone (base) 91.46 % (K=1) / 93.75 % (K=3) → +NEO (R+A+Cohort): 97.50 % (K=1) / 98.82 % (K=3) (4 seeds × 120, deterministic harness; cf. NEO notes + §17). Cheat-free, at the level of manual declare.

8.4 — Measurement procedure (replicable)

1. Apply the preset, fix the seed.
2. +1 Tick to understand one move; Auto for a game; do not conclude performance from the interactive mode.
3. Measure: seeded protocol, chained games (never reset between plays — L04), N ≥ 150 (≥ 2000 for long drift), 8 seeds for fine deltas.
4. Read coverage and OS_diss as much as wins.
5. Verify a new toggle's inertia OFF (byte-identical output) before judging its contribution.
6. Export the JSON; compare A/B with z.

8.5 — Working discipline

• Edits on a throwaway extraction (/tmp/_c.js); the saved build is never touched until a brick is validated.
• Save = ZIP, on explicit order only; one at a time.
• Inert OFF: every added brick leaves the baseline byte-identical when off.
• Check the doctrine already accumulated before investigating (time has been lost rediscovering documented findings).

9Anti-cheat framework

• Legitimate: revealed letters, length, tried letters, everything learned by playing (bench of seen words, sound↔letter tables).
• Gray cheating: enumerating the full dictionary to isolate the word (A2, full lexical cohort). Tolerated only to declare.

9.1 — The const ph = G.wp.get(currentWord) case

Reads the target word's phonology = a key→answer lookup. Its status depends on two factors: (1) regime — in Hangman (board-only) the word is not "heard" → hard cheating; in dictation the word is the input → legitimate; (2) destination — confined to P2/end-of-game learning/dictation → tolerated; fed back down into M4_d/M5_d → hard cheating. The "scrambled phon → rate unchanged" test: currently a latent, unexploited oracle. Rule: never wire it into Hangman scoring; for a cheat-free phon route, derive ph from the cohort (at the cost of the ~78 %).

9.2 — Cheat-audit procedure (reusable)

To test whether a source X is exploited by scoring: scramble/neutralise X and measure Δ(rate). If Δ ≈ 0, X is not exploited (a tolerated latent oracle); if Δ < 0 significantly, X pulls the decision (to be cut if X is an oracle). Judge a brick in the full system (ablation), never in isolated override.

10Operating regimes

Each mechanism dominates in the regime that supplies its input.

• Poor context (Hangman, isolated word): shape prevails → phon→ortho and recall dominate. The lever for the last points = recall (recognising), not grammar.
• Rich context (running text): grammar dominates (~92 %); the residue (silent endings) is lifted by grammatical category (POS) — on the silent -ent, uncertainty drops from 0.93 to 0.019 bit (−98 %, AQUA-type measurement).
• Dictation (word heard): phonology is a legitimate input; phonemes + grammar surpass shape alone.

Consequence: in single-word Hangman, 100 % on a never-seen word is impossible cheat-free — this is the very definition of cognition vs oracle.

11Results: success & failure

11.1 — Success

• Integration thesis confirmed measured: pure frequency ~22–23 % → cheat-free reference config ~90 % on the same hardened task. No isolated route carries this score (bricks < the whole; scrambled phon → rate unchanged).
• Recall (seen words): 95.3 % / 1.33 err, within ~3 pts of the oracle (98.7 %) — VSA recognition approaches full lexical coverage without reading the dictionary.
• Generalisation (new words): 54.7 % / 4.45 err by phon→ortho assembled decoding, without any recognition — honest, like a human facing an unknown word.
• Effect of the learning fix (M_LEARN_FROM_COGNITION, 8 seeds): wrong phon target 20 %→0 %; mirror cascade MISS-frac 71 %→80 % of the healthy level (z=14.6).

11.2 — Failure

• Profile of the losses (phon dyslexia) measured: across 100 losses, 96 % involve a phonetically close pair. Distance 0 (homophones I/Y) 5.8 %; distance 1 (voiced/voiceless P/B, T/D, K/G, F/V, S/Z) 58.3 %; distance 2 29.2 %; distance 3+ 6.7 %. Cause: voiced/voiceless pairs share 13/14 features → near-identical vectors.
• Structural mirror ceiling: the cascade's contribution is capped by (1−winRate); at ~97 % the amountFactor drops to ~26 % (design F178, not a bug).
• Weighted quality capped ~0.80 (max 0.84 OFF / 0.88 ON) vs target 0.90.
• phonGraph inert: ‖Δ‖ ≈ 0 (sign-based learning → zero mean; cos ≥ 0.997 vs substrate).

11.3 — Honest assessment

Cheat-free cognition reaches ~90 % and approaches the oracle on seen words, but: (a) it plateaus below the oracle on new words (expected); (b) the residual losses are dominated by the unresolved voiced/voiceless phon confusion; (c) the M3_d concept cells (shape) are weakly connected to the hub (M3_d.output=0 under bPC) — but its reward-readout cLetterScore is a cheat-free contributor measured at +3.4 (§12; AUDIT_OMEGA §3.1): "M3_d weakly connected" holds only for the cells/hub, not the readout.

12State & limits

Ranked by impact × effort, each point framed as a falsifiable experiment. None degrades the baseline (everything is off/decoupled, byte-identical at rest).

Recommended order: (1) purge non-seed Math.random done (phase46 · Fix#1); (2) hand-coded PHON_TO_LETTERS → learned L2 table done — the NEO/emergent assembled route, L2 learned board-derived (§17); (3) lodge the recall bench in M3_d falsified — measured net-negative (concept→M4 contaminates, −1.3 to −2.0); capacity probe: carrying familiarity requires K≈N cells (= the bench), 12 cells AUC 0.64 vs bench 0.98 → familiarity lives in the bench, M3_d stays the small semantic latent (§17); (4) bound the Hebbian + fix the bijective projection to do.

13Open research questions

• θ → (α=1, β=1)? Does the OS, learning on its own (🧬 Train θ), fall back on the analytic dosing?
• Route ratio r ≈ 5.35? Does the ratio signal_phon / signal_ortho (≥ 1000 games) approach 5.353?
• Fraction of dormant channels ≈ 0.079?
• Ortho/phon asymmetry: intrinsic or learned? Does a symmetric init diverge spontaneously?

To become scientific, each correspondence would need a mechanism that predicts the constant from the cognitive setup — otherwise the result remains an analogy.

14References

Status: structural/active · present but inactive · not implemented.

15Glossary

16Appendices

16.1 — Ablation table (measured effects)

16.2 — Quick reference of constants

5 M1_d sub-brains, phases [0,103,207,311,415], LDIM=512 · 16 M2 zones (4×4), EMA 0.05, σ 1.5/4.0 · 12 M3_d cells, α_Hebb 0.008, α_decay 0.001, threshold 0.005, SDIM=1024 · M4_d 70/30, top-10 · M5_d T 0.05–0.5, gap×20, M1_m weight 0.1 (F198) · reward +1/−0.5/+2/−1 · anti-Hebbian 0.004.

17Declare NEO (phase 46)

A new emergent and crossed declare added additively on top of phase45: recall (addressed) + assembled (phon→ortho) + cohort (board filter) + mute (via sound, crossed) + trigger. OFF-inert (byte-identical baseline), none of the 5 existing declares modified, bPC M3_d allowed but NEO is not M_BPC_DECLARE. Current build OMEGA_v0_07_phase47_trexquant_mode.html (engine = phase46).

17.1 — The NEO toggles (UI)

UPDATE 06/26: 3 cheat-free blue toggles added (board-derived sound · joint · OS arbitration) + M_DECLARE_DUAL adopted in the reference config (§8.3, §17.5). UI colour code: green optimal config · blue optional cheat-free · orange reads the SOUND of currentWord (word heard) · red gray cheating (A1/A2/A3).

* sub-flags active only if the master is ON. Numeric parameters (NEO_CONF 0.75 · MUTE_CONF 0.85 · TRIGGER_GAP 0.005 · RECALL_MARGIN 0.20) at their measured defaults, not yet tunable in the UI; M_NEO_G2P_EXP_PEN (0.5) is tunable (the "pén" [pen] field of the revealed-g2p toggle).

17.2 — Cheat-free (verified line by line)

The only access to the "complete" hidden word: w.p (pronunciation) = a legitimate input (proto-language premise, M4_PHON_USE_P); the masked align blocks any leak of an unrevealed grapheme. A1/A2/A3 remain gray cheating OFF.

17.3 — Crossing ≠ multiplying

The mute crosses sound and ortho — it neither adds them (dual log f + wO·ortho + wP·phon) nor multiplies them. Crossing = joint P(letter | head-phoneme, offset, revealed neighbours) with backoff.

• Resonator networks (Frady 2020): "do not multiply blindly" (decoding a bound product = hard combinatorial search).
• bind = a code retrievable by word (the recall's role: 0 dead cells, ×6 generalisation, 100 % retrievable). It is not the letter engine: on the global dictionary, bind = 19 % vs EM 64 % (frozen spec, capacity wall).

17.4 — Results measured

Deterministic headless harness (reseed before init → bpcW; _omega_OSL_reset; M_OS_v07.α/β=1), strong-cognition cheat-free config, 4 seeds × 120, paired Δ + z.

Win-rate lever = Recall + Assembled + Cohort (manual-declare level, cheat-free). The trigger takes the mute from −0.6/−0.8 to ≈ 0: it no longer drags.

Brick-by-brick contributions (Δ vs base): recall +1.76 (K=3; 100 % on lived, inert on a new word) · assembled +5.28 (K=1) / +3.43 (K=3) · cohort +0.50 (K=1) / +0.17 (K=3; hit 96.7→98.3 %).

17.5 — UPDATE 06/26 — cohort without currentWord, DUAL adopted, OS arbitration measured

Full detail + reproductions: AUDIT_OMEGA.md §1.1–§1.6. Summary:

Regime honesty (orange). The 97.5 % cheat-free rests on M_NEO_ASSEMBLED reading wp.get(currentWord) = the sound of the word ("word heard" premise). This is legitimate in dictation, but in Hangman it reads a property of the answer → flagged orange in the UI. Benchmarks without this premise: OOV phon→ortho ~70.7 %, pure ortho OOV ~22 %.

Board-derived cohort + joint (full cheat-free, without currentWord). M_NEO_PHON_COHORT derives the sound from the board-cohort consensus (no more wp.get(currentWord)). Measured cost: ~0 in-lexicon, ~7 pts OOV (purity guard 0.5). The joint sound×ortho (_neoCRS, crossing = joint) beats the argmax by +2.2 pts on every seed (K=1, 4 seeds).

DUAL adopted (§8.3): a word-level declare (a naive-Bayes word model, not the per-letter crossing governed by §6/§3.1 → no breach), cheat-free (board + length, never currentWord), +1.8 → 99.8 % word-heard / +2.5 → 97.3 % without currentWord, stable. Caveat: in-lexicon recognition → neutral OOV. OS arbitration: the only *principled* mechanism that beats the base (+2.0, a true interactive DRC via M_OS_v07_step), but does not beat DUAL → kept OFF-inert as a clean alternative.

Directionality (safeguard). The cognition's phon route is ortho→phon (sounds out the revealed letters — reading), the OS arbiter has a reading-tuned θ; the declare is phon→ortho (spelling, Hangman's strength). A bidirectional DRC coexistence; but the declare-level OS arbitration must force its own α=β=1 (done) so as not to inherit the reading θ — otherwise the routes' senses conflict.

Falsified (do not redo, R66): end-distance morphology (−1.0) · dense-backoff morphology (+0.3 noise) · cross-modal crossing at the M3_d concept (−3.0, capacity wall) · word-joint (−2.3) · frequency crossed at the phoneme (−4.3). The cheat-free residue lives in the cohort ambiguity, not in the phon→letter table nor the concept.

17.6 — Reservations hypothesis / to do

• Mute neutral with the trigger, not a lever: an irreducible silent residue, already covered by cognition. Alive and harmless.
• Trigger: SPSA on the gap = flat (constant win rate → no gradient; effect below the measurement quantum). A dense per-decision reward (advantage [mute∈word] − [cognition∈word]) would be needed to learn it online.
• NEO numeric parameters not yet tunable in the UI (only the 6 toggles are).
• Power 4 seeds × 120; R66 suggests ≥ 200 × 4 to tighten (stable directions).

17.7 — UPDATE 06/2026 — honest OOV, the n-gram route, and the cognitive "C" measured