Research & engineering memoir

A cognitive architecture for French Hangman

How a machine can play Hangman by reasoning rather than reading the answer — and what the constraint "cognition > oracle" taught us.
OMEGA-Ω v0.07 · build phase47 (cognitive engine phase46) · direction Rem · written by Claude (Anthropic)
Snapshot as of 03/06/2026 — addendum 14/06/2026: derived application = diagnostic dictation (dyslexia), see dictee/ ; addendum 18/06/2026: dual-route grammar layer + dyslexia corrector (§10) ; addendum 24/06/2026: typing aid (accented completion) folded into the corrector (app + extension) + learning loop (unified dyslexia profile dictation+corrector, adaptive dictation, progress curve) — see DICTEE_ROADMAP.md ; deployment = embedded compressed lexicon (gzip+base64, browser ≥ 2023) · doctrine cap §43 · R66

0Summary

Abstract OMEGA-Ω is a French Hangman engine (words ≥ 7 letters) built not as a dictionary searcher but as a cognitive architecture. Its doctrine — cognition > oracle — forbids any module from reading the hidden word beyond the revealed positions. Under this constraint, cognition alone reaches a ~90 % win rate; complemented by an emergent, cheat-free declaration system (NEO), it reaches 97.5 % to 98.8 % — on par with a lexical oracle, without ever reading the dictionary to score a letter; since November 2026, OS arbitration and information partitioning carry the reference configuration to 98.8 % (12 seeds, dedicated protocol), and the engine can also play without knowing the word's length — 96.7 %, above its own with-length reference under the same protocol (§7.4). These figures hold in the in-lexicon regime; on out-of-dictionary words (Trexquant protocol), recognition no longer helps, but the phonology→spelling sequence still generalizes to ~70 % — on par with good solvers. This memoir lays out the architecture (dual Möbius pipeline, hyperdimensional substrate, spelling/phonology routes), the method (measure-first, R66 falsification, deterministic harness), the results, and — just as instructive — the negative results that mapped the limits of what was possible.

1Thesis

Performance is not cognition. A program that wins at Hangman by consulting a dictionary performs; it does not reason. OMEGA takes the opposite as its founding constraint.

Hangman is an ideal test bed for this distinction. The trivial solution — filter the dictionary by the revealed pattern, score letters by the cohort's frequency — reaches ~95–98 %. But it says nothing about understanding: it knows French by heart. OMEGA forbids itself this crutch (grey cheat, switches A1/A2/A3 kept OFF) and asks: how far can a cognition that is not allowed to read the answer go?

"Without reading the dictionary to choose a letter, can one reach the oracle?" — The measured answer is: to within a few points, yes.

The project's central answer: starting from ~2.6 % (first weeks of v0.07), pure cognition reaches ~90 %, and with a cheat-free declaration system, 97.5–98.8 % — against an oracle ceiling of 98.7 %. The remaining gap is not a failure of cognition: it is the language's irreducible ambiguity (a /o/ can be spelled o, au, eau — only the memory of the exact word decides).

2The task & the doctrine

The game: guess a French word of 7 or more letters, letter by letter, within an error budget. On each turn, the engine sees the board (revealed positions) and proposes a letter.

2.1 — Cap §43: the golden rule

No cognitive module reads currentWord beyond the revealed positions. The upward direction (deciding) sees only what is revealed; the downward direction (learning) operates after the game, on the now-legitimate complete word. Reading the answer is only cheating in the upward direction.

2.2 — What is allowed, what is not

  • The lexicon is forbidden in letter-scoring (upward). It is allowed for the cohort-based DECLARE (board-derived) and post-game learning (downward). A1/A2/A3 = injection of lexical frequency into scoring = grey cheat, OFF.
  • w.p (the pronunciation) is a legitimate input under the "heard word" premise (proto-language, M4_PHON_USE_P) — just as one knows the sound of a word one is trying to spell.
  • R66: no module is enabled by default without a falsification test (bypass + paired multi-seed statistics).

3Method

Experimental discipline is the heart of the project — more so than any particular module.

  • Measure first. Every architectural question is settled empirically (paired A/B/C, multi-seed), never by conviction.
  • Deterministic harness. Online learning + predictive coding made the baseline non-reproducible; the harness was made deterministic (re-seed before init → deterministic bPC weights; reset of the θ state; θ deployment target reset; game RNG aligned). Without this fix, all the Δ were noise.
  • Falsification (R66). A module is kept only if, OFF, it is inert (byte-identical baseline) and, ON, it shows a measured contribution. Otherwise it is reverted — even if it "seems" right.
  • Isolation. All prototyping on disposable copies; the build is never modified without an order.
  • Amplitude AND direction. A fix can restore the correct direction of a signal while appearing modest in amplitude (structural damping) — one must judge both.

4The architecture

OMEGA is a dual Möbius pipeline: a direct circuit (upward, perception → decision) and a mirror circuit (downward, deconstruction toward proto-language). In practice, the P2 mirror remains largely disconnected; the power comes from the direct circuit and its couplings.

4.1 — The substrate & the five levels

The substrate is hyperdimensional (HRR/VSA, Plate 1995): 1024 dimensions for the concept (SDIM), 512 for the lexical (LDIM). Five upward levels:

  • M1 — perception: binds position and letter by circular convolution (HRR). A rich representation that tracks reveals and discriminates words.
  • M2 — zones (4×4): reduces M1 to 16 zones. historical flaw: the absolute value + averaging throws away the sign/phase → in practice M2 encodes only length. Root cause of a concept collapse (fixed by wiring the concept onto M1).
  • M3_d — concept (12 cells, 1024D): a bidirectional competitive autoencoder (bPC, DBPC/Qiu 2025 blueprint) now encoding from M1 (rich), plus a reward-learned readout.
  • M4 — lexical/letter: converts the concept and the routes into per-letter scores.
  • M5 — decision: picks the letter, with a margin (gap) that measures confidence.

4.2 — Two routes, one arbiter

In line with the dual route of reading (Coltheart, DRC): a spelling route and a phonological route (in SAMPA, from w.p) reason in parallel. The OS v07 arbitrates between them to decide. A salience layer (M_S) fuses the concept into scoring. The whole is regulated by bounded online learning (4 safeguards).

Beyond two routes — precisely. This spelling⟷phonology arbitration is the one for Hangman. The derived corrector (§10) makes three routes reason — spelling, phonology and grammar — but they combine through distinct mechanisms, not a single uniform μ: the OS arbitrates spelling⟷phonology in the speller, grammar merges via a priority cascade (the FP=0 red wins at the same token), and agreement via OS arbitration of the subject in turn arbitrates its own sub-routes (nearest noun · head before "de" · outside-PP · LM coherence). The two-route DRC describes Hangman; OMEGA taken as a whole arbitrates more — through several arbitration paths, not one. A result specific to the project, unborrowed.

5Declarations & the NEO system

Pure cognition plateaus at ~90 %: excellent on what it has seen, limited on the never-encountered word. Declarations close the gap by proposing a whole word when confidence is high. The manual (lexical) declare reaches 98.8 % — but reads the dictionary. The project therefore built an emergent, cheat-free declare, culminating in the NEO system (phase 46):

RoutePrincipleContribution
Recall (addressed)VSA recognition from the revealed board + a bank of lived words. Exact episodic memory.100 % on lived
Assembled (phonology→spelling)Decoding of sound positions via a learned L2 table; masked alignment (no hidden grapheme read). Generalizes to the new word.+5.3 (K=1)
CohortBoard-derived filter: allows only a letter carried by ≥1 word compatible with the revealed pattern.+0.5
Silent (by sound)Positions with no phoneme predicted by a crossed phonogram.neutral
Trigger (B)Overrides only if cognition is uncertain (low gap).gate

NEO is additive (OFF-inert, byte-identical baseline), modifies none of the pre-existing declarations, and remains cheat-free by construction: recall tied to the revealed only, masked assembly, downward learning.

6Crossing is not summing — nor multiplying

A central lesson, long misapplied. Combining two sources of evidence (spelling and sound) is done neither by a weighted sum of scores (log f + wO·ortho + wP·phon), nor by a product of marginals (which assumes independence).

Crossing = conditioning jointly. P(letter | phoneme and context), not P(letter|phoneme)·P(letter|context).

The literature drawn on says so: resonator networks (Frady 2020) warn against "blind multiplication"; the hyperdimensional bind (Plate) is a per-word retrievable code — it carries episodic memory (recall) but fails as a letter engine (19 % on the global dictionary against 64 % for an EM aligner: capacity wall). The correct form of the spelling×sound crossing is a discrete joint; measured, it doubles the prediction of silent positions (27 % → 52 %).

Measured Silent route, top-1 position: spelling 26.8 % · sound 21.2 % · joint crossing 51.6 %. The product of marginals coincides (51.8 %) but remains the wrong mechanism.
Precision — where the joint lives (structural audit 06/26) The joint Σ_φ P(φ|p)·P(letter|φ, context) is implemented where it pays off: the cheat-free declare (NEO), and it is correct there (measured, §17). The per-letter cognitive scorer, by contrast, does not perform a joint: the OS w(r) performs a route arbitration (convex mixture spelling⟷phonology, dual-route DRC) and M5_d stacks additive/multiplicative enrichments of marginals. This is a deliberate design choice (route arbitration ≠ marginalization over the latent), not a flaw — and attempting a joint inside cognition (cross-modal crossing at the M3_d concept) was falsified (−3.0 pts, 12-cell capacity wall). The joint therefore remains a deliberate mechanism of the declare, where the latent φ carries an exploitable signal.

And well beyond Hangman. The distinction "arbitrate by reliability ≠ sum/multiply" cracked a wall thought impassable in the corrector: recovering the subject of a mis-agreeing verb (real-word error) fails via a single route and floods via hard combination, but the OS arbitration of subject-routes weighted by reliability — the same law μ = r/(1+r) as cognition — succeeds on the residual that the parser gives up on. The arbitration geometry was not specific to Hangman; it is the right tool whenever one crosses routes. Detail: §10.

7Results

All figures come from the deterministic headless harness, strong-cognition config, words ≥ 7 letters, 4 seeds × 120, paired Δ.

ConfigurationK=1 (generalization)K=3 (repeated vocab)
Pure frequency (low reference)~22–23 %—
Cheat-free cognition alone91.5 %93.8 %
+ NEO declare (R+Assembled+Cohort)97.50 %98.82 %
Oracle ceiling (full lexicon) excluded98.7 % / 100 %—

Three observations. (1) Performance comes from the integration of modules, not from a single route. (2) The cheat-free version reaches the oracle within a few points without reading the dictionary to score a letter — the central result. (3) The assembled phonology→spelling route with a learned table beats the hand-coded table (71–85 % vs 67–79 % on sound positions): learning > hand-coding.

7.1 — In-lexicon vs out-of-lexicon regime

The figures above hold in-lexicon (the word belongs to the solver's dictionary). This is the regime where recognition (recall/cohort) dominates — and where OMEGA is on par with the best lexical solvers. The decisive test is the out-of-lexicon regime (words removed from the dictionary, Trexquant-style): recognition can do nothing, only sub-lexical generalization counts. This regime is playable and measurable in the application: the 🎯 Trexquant mode removes the drawn word from the dictionary (cohort and recall blind), and a dedicated bench computes the out-of-lexicon win rate.

Regime / configurationWins (words ≥7, budget 6)
In-lexicon (recognition)~97 %
Out-of-lexicon — phonology→spelling cascade (with sound, the concept)70.7 %
Out-of-lexicon — flat fusion (routes in parallel)66.7 %
Out-of-lexicon — spelling only (letter-only)22 %
Trexquant benchmark (letter-only, no sound)~50 % / good ~65-68 %

Three takeaways. (1) The 97 % is essentially in-lexicon leverage; remove recognition and the win rate drops. (2) The phonology→spelling concept (the assembled sequence, the word being heard) carries the entire out-of-lexicon: 12–22 % without it → 70.7 % with — on par with good solvers, on a legitimate task (spelling a heard word, doctrinal w.p). (3) The margin-guarded cascade beats flat fusion (70.7 % vs 66.7 %; the fused phonology route alone, 12.7 %, hurts) — confirming that combining routes goes through the margin, never through a sum. The real weakness is pure spelling OOV (22 %): sub-lexical generalization without sound remains to be built (morphological layer, §10).

⚠️ UPDATE 06/2026 — OOV figures to re-read The OOV figures in this section (≈70.7 % "with sound") predate the discovery of the cohort leak (_neoWBL, AUDIT_OMEGA §1.6.1). After correction: pure sublexical generalization ≈ 33 %, cognition alone ≈ 11 %, aggregation n-gram ≈ 57-66 % (gap-aware ~63-65 %, SOTA band). See the detail in §10 (UPDATE cont.). ✓ Re-verified (06/2026) by re-running the engine's own bench (_omega_trexquant_bench, 3 seeds): gap-aware 64 % — stable across the 3 seeds; OS-arbitrated n-gram up to 68 %; in-lexicon 99 %. An independent measurement, it confirms these bands. Controlled head-to-head (same OOV set, 3 seeds, budget 6): gap-aware cognition 64 % vs a trigram+backoff n-gram baseline 46.7 % (+17 pts on identical words) vs held-out lexical cohort 24 %. What this really validates: ✅ the method beats a clean statistical baseline on French OOV (+17 pts, same set) — direct validation of the cognitive contribution; ✅ reproducible (64 % stable 3/3 seeds) and cheat-free (the baseline, by contrast, exploits the dictionary in full); ⚠️ honest ceiling: the baseline is a standard trigram, not the ML SOTA (trained LSTM/transformer) — a heavier model would narrow the gap, so "beats standard solvers" is proven, "beats the very best ML" is untested. And the resources point — it is fair, and this is where it weighs: the solvers that climb to 65-68 % are GPU-trained networks (pipeline + massive corpus), whereas OMEGA reaches 64 % in a single HTML file, without training, without a GPU, in a tab; the right framing is therefore not "it wins more", it is: performance comparable to the best, at a fraction of the resources — and clearly superior to standard solvers. Verdict: method validated as reproducible, cheat-free, measurably better than a standard baseline, and competitive with heavy ML at an incomparably lower cost; the only remaining "no" — strictly beating the best ML — would require a duel against a real network, which I cannot set up here.

7.2 — Porting to English: what the arbitration geometry says, and what it does not

The engine was ported to English (clone omega-pendu-en.html) without touching French: English lexicon, English letter frequencies, and — the delicate part — an English g2p rebuilt data-driven (grapheme→phoneme tables aligned by EM over 30,000 word↔IPA pairs of lex_en), so that no French remnant phonemizes an English word. The arbitration algorithm is identical: only the data change, like the lexicon.

Caution — the phon↔ortho balance is not fixed. It flips with the task and the regime, in both languages: spelling reigns in-lexicon (§7.1), but it is the phonology route that lifts the out-of-lexicon (+14 to +18 pts); and from one engine to the next the balance shifts again — Hangman (OS arbiter, phon⟷ortho), dictation (p2g decoder, phon→ortho), corrector (candidate choice, ortho×frequency×context). There is thus no "single-route language": speaking of a "dominant" route only makes sense at fixed task and regime. Detail: the Arbitration page.

What the port shows cleanly, at θ=(1,1) and in the sole in-lexicon, letter-choice regime (same harness, 2 seeds × 40 words; NEO tables cold → absolute figures below the 97.5 % of §7, but comparable across languages): the arbiter's coherence — 4·E[μ(1−μ)], "both routes jointly engaged" — is lower in English than in French, at a similar mean μ (≈0.6 in both).

Language (in-lexicon, letter choice)coherence 4·E[μ(1−μ)]mean μ (phon-route weight)
French0.66 – 0.740.55 – 0.63
English0.48 – 0.580.57 – 0.62

Since the mean μ is similar, the reading is not "one route overwhelms the other" but: in English the arbiter commits more sharply (μ departs more often from ½, the two routes less often engaged together) — consistent with a more decisive orthography once letters are revealed, yet a reading of that specific regime. The raw ratio r = signalphon/signalortho is more dispersed in English, but its mean is inflated by rare spikes; we rely on coherence, which is robust, not on r̄.

Note — (1,1) is not a win-rate optimum It is a point of symmetry: the Möbius axis β=1 (self-dual ∀α), α=1 makes the law homographic, and the coherence ridge falls at their crossing — set by structure + MDL, never by score. The θ win-rate sweep is a confounded instrument (crippled ortho arm): kept for the form, discarded as a fitness measure (the Arbitration page). Accordingly, re-tuning θ for English yields no robust gain — the deltas of a training grid vanish under cross-validation. (1,1) holds by structure, for both languages.

Honest corollary, and this is the point: the in-lexicon English↔French gap is not a tuning failure, but neither does it reduce to "English is spelling." The dual-route cognition balances differently across language AND task — §6 said as much for Hangman vs the corrector; the port extends it to languages, recalling that the "dominant" route switches sides depending on what you do.

7.3 — Punctuating speech: two channels, and a division of labour that is not symmetric measured

Voice dictation delivers words without punctuation — the browser's recognition service emits none in French (measured: 0 characters over three takes). Rebuilding punctuation brings together two channels whose asymmetry is the main result: the audio knows there is a boundary, the text knows which one. The audio measures the silences — information the text does not have — and lands within one word; the text re-anchors among neighbouring positions, picks the type of mark, and above all refuses when no position is grammatically legal.

Lesson · the refusal is what carries the combination Without it, the anchor proposed a mark at every pause and correctness fell from 69 % to 22 %. With it, the measure goes from F1 0.33 to 0.48 on a real dictation. A route that proposes only has value if the other can oppose — the same structure as the Hangman inter-route arbitration, but with a veto instead of a weight.

One ablation settles a design question: the duration-only anchor beats the learned acoustic-model anchor once the text channel is in place as a filter (F1 0.480 against 0.435). The model only compensated an imprecision the text corrects better and for free — no acoustic model is therefore shipped.

On top, five comma rules drawn from a public pedagogical source carry families statistics ignores (coordinator between two clauses, coordinator at the head, parenthetical, correlation, repeated « ni »). Conformity to the source: 3 → 14 examples out of 50; and on 11,304 human-written sentences they improve correctness (50.5 → 52.0 %) and recall (11.0 → 12.8 %) simultaneously. What makes them shippable is not their content but their guards: the raw rule "comma before the coordinator" tops out at 31 % correctness, because donc and aussi are first of all adverbs.

Dividing line · judgement vs mechanics "Where should a comma go?" is a judgement — the source itself writes "one generally places". No tuning will make it a certainty, and the measure confirms it: correctness plateaus at ~55 %. But "is the spacing around this existing comma correct?" is mechanical, decidable on the string alone. That second layer reaches zero false positives: four typographic rules applied automatically, 2 firings on 14,450 correct sentences — both of them genuine corpus mistakes. The lesson goes beyond punctuation: before trying to decide better, separate what is decided from what is observed.

7.4 — Playing without knowing the word's length measured

The word's length is legitimate Hangman information (§2) — nobody thinks of removing it, and that is precisely why the ablation is instructive: what is left of the engine when you take a whole sense away? The 🕶️ mode (November 2026) shows the machine only the relative schema of the revealed letters — for INFLAMMABLE with E and A found, A??A??E, unknown edges. The game itself does not change: the dealer reveals on the real board, the top-down loop learns as before; only bottom-up perception is windowed.

The machinery follows the constraint: a floating cohort (the schema is laid at every offset over each word; the compatible set tightens monotonically within a game), a relative n-gram at the edges, a recall re-anchored in the schema's frame, and — the decisive lever — OS arbitration whose sub-lexical route first tries the phon-cohort joint (doctrine §3: a soft distribution, no argmax), with an n-gram fallback when the cohort is silent.

Configuration (same bench, same protocol, cheat-free)with lengthwithout length
Cognition alone93.8 %91.7 %
Full machine (+NEO)96.3 %96.7 %

The result that earns its place in the memoir: the mode beats its own with-length reference (96.7 > 96.3; 4 seeds, verified leak-free — real board identical seed by seed, placebo tight). The reading is not "less information = better": the constraint forced better routes (arbitration + joint) that the reference cascade was not using. Wired back onto the real board, those routes carry the reference to 98.8 % — the ablation improved the standard regime.

The word-by-word versus (12 seeds, 720 words) shows nearly disjoint failures — 39 words lost by the mode alone, 13 by the reference alone, 8 shared — and that versus produced two more levers, measured then activated: information partitioning in a dense cohort (2–16 candidates: play the letter that best splits the cohort rather than declare too early; 97.1→98.6 % over 12 seeds) and cohort support (98.8 % combined, losses 21→9). Out-of-lexicon was verified before activation (800 words removed one by one, paired): partition sound (+1.0); support −2.8 — its premise, "the true word is in the cohort", is false out-of-lexicon — hence a gate that switches it off in the Trexquant regime (no-op proven: 0 games changed out of 800). The application's current preset: 99.2 % (4 seeds) · 98.8 % (12 seeds).

Lesson — ablation as an instrument Removing a piece of information did not just measure the engine's dependence on length (−2.1 pts once the standard regime is made comparable): it forced the architecture to invent routes that the standard regime then inherited. Ablation is not only a control — it is a generator of levers.

8Negative results — and what they teach

As much as the gains, the measured noes shaped the architecture. Recording them avoids repeating them and reveals the real walls.

8.1 — M3_d cannot be an episodic store falsified

The tempting idea of "housing the recall bank inside the concept" was tested twice. Feeding M4_d with the recalled word contaminates letter-scoring (net-negative, −1.3 to −2.0). And carrying familiarity into the cells hits a capacity wall: recognition AUC climbs steadily with the number of cells — 0.64 (12) → 0.87 (96) → 0.98 (K=N) — so that carrying it demands one cell per word, i.e. re-storing the bank, with no gain over the exact Map that already exists.

Lesson · complementary systems Episodic (the bank — large, exact) and semantic (M3_d — small, compressed) are separated by design — like hippocampus and cortex — this is the complementary learning systems theory (McClelland, McNaughton & O'Reilly 1995). Putting the episodic into the small semantic latent fails on capacity. M3_d is not broken: it is simply not a word store.

8.2 — The silent route is alive but harmless

Even correctly crossed (the right mechanism), the silent route only becomes system-positive with the trigger, and then only neutral (Δ +0.00 to +0.14). The silent residual is the irreducible ambiguity; cognition already covers it. The silent route is a clean complement, not a lever.

8.3 — Not everything can be learned from winning

Learning the trigger threshold by SPSA on the win rate yields a flat gradient: the effect is below the measurement quantum. A reward signal that is too coarse cannot guide fine tuning — a dense per-decision reward would be needed.

8.4 — The oracle is subtle

A route can be trained on clean data and still be an oracle if it reads beyond the mask at inference. The decisive diagnostic is not the cleanliness of the training, but what the route reads at the moment of deciding.

8.5 — The dual route does not transfer everywhere: agreement between routes can be a negative signal falsified

The multi-route architecture is the heart of the project; it was natural to apply it to punctuation, where two producers coexist (a statistical model with backoffs, an explicit rule set). Two compositions were tested, and both fail.

By weighting (an arbiter in the style of Hangman's: routes → distributions, weight = the distribution's sharpness, single threshold): it improves the global score but drops the anti-regression guard from 82 to 62 cases, resurfacing mistakes already fixed. No operating point dominates the shipped code.

By agreement — requiring both routes to say the same thing, which needs no calibration, only independence. The result is the opposite of the expectation:

  • model alone (rules silenced): 58.1 % correctness
  • rules alone (model silenced): 63.2 %
  • both in agreement: 37.9 %
Lesson · two routes do not make a dual route Hangman's routes read different substances — letters on one side, sounds on the other: two independent proofs, so their agreement informs. The punctuation model and rules read the same tokens with the same grammar: they are not two routes, but two implementations of one. They fire together precisely where French hesitates (coordinators, head adverbs), i.e. at positions where the comma is optional. Their agreement does not select certainty, it selects ambiguity. The design criterion that follows: before composing two routes, check that they read distinct substances — otherwise redundancy costs instead of paying.

Measured corollary: the second substance available for punctuation is the audio, and it is the only combination that paid (§7.3). Three other leads fall with the same logic — hardening the threshold gains nothing (correctness is flat at ~55 % whatever the threshold: a model whose confidence does not predict correctness is not ill-tuned, it is saturated); the double space as a "written pause", the graphic analogue of the audible pause, carries no signal (6.32 % against a 6.19 % base rate); and the question mark as a marker of subject inversion is refuted (12 of the 13 problematic inversions are outside questions — in French, inversion is mostly stylistic or relative).

8.6 — The "no-length" campaign: five measured noes falsified

The no-length mode (§7.4) left as many noes as levers — recorded so they are not redone.

  • Re-perception at anchor shift: re-initialising the producers when the schema's anchor shifts looked like hygiene; it is a 4/4-seed regression (87.1 % against 91.7). Misaligned memory carries more signal than the misalignment costs.
  • Opening the phon route before the first correct letter: feeding M4_phon the full cohort from the start — neutral for the full machine (declarations pre-empt), −3.4 for cognition alone. The "silent before anchoring" choice is confirmed by measurement, not by principle.
  • Fog weighted by P(overflow): a refinement of the relative n-gram that became inert the day the joint took over the sub-lexical route — byte-identical games. A module can die of another's promotion; measure after every recomposition.
  • Post-miss discounting: reweighting the prior after k misses — noise (mode +0.5, reference −0.3). Left OFF.
  • Cohort support out-of-lexicon: "only try the survivors' letters" looks unassailable — it costs −2.8 as soon as the true word is not in the cohort (Trexquant regime). The remedy is not to weaken the rule but to gate it: active everywhere except out-of-lexicon, no-op proven (0/800).
Lesson · the reference gets verified too The campaign nearly concluded against a contaminated reference: the old "99.6 % with length" read the hidden word's sound in the assembled route. The honest comparison first required a cheat-free reference (96.3 %) — that is the one the mode beats. Before measuring a gap, audit both sides of the comparison; and bound every mode to its window (the mode's aids only activate inside the window — the toggle alone was enough to break an unrelated button).

9Principles

What the project has distilled, transferable beyond Hangman:

  1. Cognition > oracle. Separating performance (reading the answer) from cognition (reasoning about what is revealed) clarifies what one is really building.
  2. Learning > hand-coding. Tables learned while playing beat expert tables — and stay within doctrine if they are board-derived.
  3. Cross, don't sum or multiply. Condition jointly; the bind is a per-word memory, not a letter-prediction engine.
  4. Margin cascade > flat fusion. Recognize first, decode if the margin is low — flat fusion dilutes.
  5. Complementary systems. Exact episodic and compressed semantic have distinct roles; do not conflate them.
  6. Measure, falsify, keep last. OFF-inert + measured ON contribution, or revert.

10Open questions & perspectives

  • Learned trigger via dense reward (per-decision advantage) rather than win rate, to arbitrate recall/assembled/cognition.
  • Morphological/grammatical layer — realized in the dictation offshoot (update 18/06/2026): an in-context agreement lever (subject-verb 94 %, distant subject, past participle + preposed direct object, noun-phrase gender) and dual-route grammar — a lexical route (Lexique 4 cgram: 12,415 verbs, 53,050 gendered nouns) × sublexical (rules), an upward loop (deciding) × downward (learning: the loop learns the gender lexicon from usage, 100 % precision, FP=0, but bounded by the data). Outcome: a dyslexia corrector that detects + corrects + locates the stage (zero false positives). See dictee/GRAMMAIRE_DOUBLE_VOIE.md, dictee/CORRECTEUR.md. Extended (30/06/2026) into a TWO-level corrector (3-engine parity): red FP=0 — agreements, être↔avoir usage (« il est faim »→« il a faim » [is/has hungry], « il a allé »→« il est allé » [has/is gone]) + misspelled auxiliary, sentence-initial capitalization; green vigilance (does not assert, hence outside FP=0) — confusable words (homophones + paronyms, ~80 groups, senses auto-Wiktionary CC BY-SA) and run-on (missing punctuation). This is where the distributional signal ("meaning") — too noisy to pass the FP=0 gatekeeper in correction (measured: it fails even on the lexical case in real text) — finally finds its place: suggest without deciding. Extended (07/2026): application by confidence (the source text is never rewritten; FP=0-safe + ~79 % spelling candidates applied reversibly, vigilance on click), replacement quality locked in CI (frequency dominance, capitalization, apostrophe ’), and guarded delivery (freshness of the extension zip, service-worker probe) — from a full cross-verified audit, detail: report §18 (update 03/07/2026).
  • Sentence Hangman — falsified as a win-rate lever (letter sharing "leaks" the endings; on the ≥7 engine the agreement prior in DECLARE is marginal). Sentence-cognition pays off in declaration and the short-word regime (= the corrector), not in Hangman: evo/PHRASE_HANGMAN_PROBE.md.
  • P2/Möbius: progressively wire in the downward mirror (currently disconnected), checking OFF inertia then the contribution. Note: the downward loop of grammar, by contrast, does genuinely learn (≠ the Hangman mirrors, measured inert) — the difference is the nature of the learned signal (lexicon vs letter).
  • Inherited settings: bound the Hebbian, fix the bijective projection (zero information gain).
  • Exploratory hypothesis: would a freely tuned cognition converge toward stable constants? speculative
UPDATE 07/2026 — the real-word agreement wall, cracked by OS arbitration of the subject The real-word agreement error (the verb is a valid word but mis-agreeing: « les enfants joue » [the children play], « les rapports que l'entreprise livrera » [the reports the company will deliver]) is, in the literature, the central problem of the corrector (Wilcox-O'Hearn: "detection is the central problem in real-word spelling correction"): it requires recovering the syntactic subject, often out of reach (« de N » complements, coordination, distant subject). Our red agreement parser abstains here by construction — which is precisely what holds its FP=0. Measured: no single subject route clears this wall (nearest noun 81 %, coherence LM 74 %), and the hard combinations (sum of scores, product of marginals, gate) flood — because two routes out of three perpetually support « the word is correct », diluting the only reliable vote. What cracked it — exactly the lesson of §6: not a route, nor a combo, but the OS arbitration of Hangman — a convex mixture of subject-routes weighted by reliability (the peakedness of each distribution, μ = r/(1+r)), a baked language model (~180 KB, UD-GSD held-out) carrying the confidence gradient that makes the threshold useful (without it, flat flood). On the residual the parser gives up on — the « de N » cases — the confidence-weighted OS arbitration recovers the subject's number at ~94-97 %, where no single route could. Delivered as orange vigilance (suggests « agreement to check », never rewrites → the red FP=0 gatekeeper stays intact), in 3-engine parity, guarded in CI. Validated on the real chat register (real dyslexia, not formal UD): flood 0 % on chat, 0.04 % on varied text, recall 54 % of number disagreement — after re-inheriting the structural guards of the sibling rules (participle, compound tense, pronoun-subject, elided subject « l'X »). A wall that we — and others — thought impassable for a portable tool (without a dependency parser or a large LM), crossed by the same arbitration geometry as Hangman. Detail: report §18.
UPDATE 07/2026 — the phonology route repointed from keyboard to microphone (ASR without Google) The dual route phonology↔spelling (§4.2) had never been closed on the phonology→spelling side outside the declare: judged useless for typed text, and without a source of phonemes. Both premises fall: a free acoustic model (wav2vec2-french-phonemizer, MIT) supplies the phonemes, and voice is a real need (speaking to bypass the keyboard). Repointed as a speech-recognition back-end, the phonology route decodes French text without ever calling Google. Measured arc (clean voice, 147 words): raw p2g 13.6 % → +grammar 49.7 % (+36 pts — grammar dominates, the same verdict as in the dictation offshoot) → homophones/lexicon/fuzzy recovery 70.1 % → corrector + subject parser in confidence mode 76.9 %; 81 % on a real (dyslexic) voice. Prosody pays off: silence at word boundaries yields punctuation (F1 100 %) and the F0 rise sets the "?" (confirmed on a real voice, 0 false "?"). The same walls as Hangman, re-measured: a bigger LM (Wikipedia register) lowers the score (81→79 %) through frequency traps (chères→père [dear→father]) — the small confident LM wins; OS arbitration applied to the acoustic emission hurts (down-weighting a reliable source) — crossing = arbitrating by reliability, not weakening the sure vote (§6). The ceiling (~81 %) is the acoustic model, not the decoder: OMEGA is an LM/decoder, not an acoustic model — direct confirmation of "what generalizes is the aggregation of structure" (§8, O2). Usage verdict: for a local tool, Whisper small (open weights, MIT, not Google) delivers 98 % and supersedes the phonology route; the latter remains the research proof (the dual route repointed from keyboard to microphone), and on the web — where Whisper is too heavy — the value lives in the Web Speech route + OMEGA re-dressing. Detail: dictee/ASR_VOIE_B.md, report §18.6.
UPDATE 06/2026 — cheat-free declare & route arbitration (detail: AUDIT_OMEGA §1.1–§1.6) (a) Cheat-free declare pushed further: DUAL (word model, board cohort, never currentWord) adopted (+2.5 → 97.3 % without currentWord; +1.8 → 99.8 % "heard word"). The board-derived cohort route + sound×spelling joint ("crossing = joint", §6) makes the cheat-free declare complete (without currentWord), at in-lexicon parity. (b) Arbitration of the two DRC routes (lexical ⟷ sublexical) at the declare level, via the OS w(r): measured — beats the baseline (+2.0, never below) but not DUAL → kept optional (a truer interactive DRC, cleaner). (c) Falsified (R66): morphology (edit distance, dense backoff, word-joint), cross-modal crossing at the M3_d concept (−3.0). Confirms §8.1: the cheat-free residual lives in the ambiguity of the cohort, not in the phonology→letter table nor the small latent. (d) Directionality: cognition's phonology route is spelling→phonology (reading, θ OS tuned for reading); the declare is phonology→spelling (spelling out). DRC coexistence, but the declare arbitration must force its own (α,β), never inherit the reading θ (otherwise a conflict in the routes' direction).
UPDATE 06/2026 (cont.) — honest OOV, the n-gram route, and the cognitive "C" (detail: AUDIT_OMEGA §1.6.1–§1.12) (e) ⛔ OOV retraction. The "~97 % out-of-lexicon" published earlier was a cohort leak (an un-invalidated _neoWBL cache: the "removed" word stayed in the cohort). Fixed. True measured OOV ≈ 33 % for pure sublexical generalization; cognition alone ≈ 11 % — it does not generalize sub-lexically. The OOV lever is the AGGREGATION of the lexicon's structure, not per-game reward nor instance memory. (f) The n-gram route (the aggregation substrate). A positional letter n-gram pre-computed from the lexicon (cheat-free) does ~57-66 % OOV; wired in as the sublexical route of the OS arbitration (auto-switching by reliability), it holds in-lex ~97 % AND OOV ~60 % without a switch. The gap-aware variant (nearest revealed neighbor at distance 1..4) adds +2 pts OOV (~63-65 %, SOTA band) — the first real cognitive gain above the substrate. Non-French confirmation (06/2026, O2). This aggregation substrate is domain-agnostic: re-implemented outside the engine and pointed at code (not the French lexicon), it reconstructs (Hangman-style) and compresses code as well as French (~1.1 bits/character below chance) and nothing on a scrambled sequence — "predicting = compressing". The part that generalizes is therefore indeed the aggregation of structure (a sequence compressor), not French cognition (the phonological dual route does not transfer — no "sound" in code). Measurements: evo/evo_o2_crossdomain.js, evo/evo_o2_compression.js. Group work (O1) — falsified. The twin idea — would a group of diverse versions beat the best solo? — does not hold: real versions (parameter genomes, 200 words × 3 seeds, real engine) are strong and too similar (no complementarity), and the useful coordination is already the OS's internal arbitration, not an external vote. The premise "group > best solo" is subsumed by the architecture. Measurement: evo/evo_o1_real.js. (g) The learned "C" — FALSIFIED as a breakthrough. The light C (maxent/GATE/POE) and the heavy C (multi-head transformer + FFN, grad-checked) were built and measured: the heavy one converges toward the gap-aware n-gram without beating it (win rate at optimal config: parity, OOV Δ ≈ −1.3; trained on real games = regression for lack of lexical coverage). The gap-aware n-gram (free, interpretable) remains the practical ceiling — the thesis "cognition > oracle = light/interpretable" confirmed. Beyond = heavy offline neural, deferred. Design of a future organ (learned arbitration hub): docs/COGNITION_DESIGN.md.

11Related work & positioning

OMEGA assembles five established research threads; situating it allows one to judge its novelty and rigor.

  • Hangman solvers. Constraint + entropy approaches reach ~96 % in-dictionary (known length); the Trexquant challenge, on out-of-dictionary words, plateaus from ~50 % (baseline) to ~65-68 % (good solvers). OMEGA is on par with the former in-lexicon (~97 %) and with the latter out-of-lexicon (~70 %, via phonology→spelling, the word being heard).
  • Vector symbolic architectures (VSA / HRR). The substrate builds on holographic reduced representations (Plate 1995) and hyperdimensional computing (Kanerva 2009). The resonator networks warning (Frady et al. 2020) — do not multiply correlated factors — matches OMEGA's measured lesson: the bind is a per-word memory, not a letter engine.
  • Dual route of reading. The spelling/phonology routes and their arbitration instantiate the DRC model (Coltheart et al. 2001); OMEGA adds a margin-guarded cascade rather than a flat fusion.
  • Bidirectional predictive coding. The M3_d concept follows the DBPC blueprint (Qiu et al. 2025): shared weights, reconstruction learning.
  • Complementary systems. The result in §8.1 — exact episodic bank vs compressed semantic latent, and the capacity wall when the two are conflated — is precisely the hippocampus/neocortex division of CLS theory (McClelland, McNaughton & O'Reilly 1995). OMEGA empirically rediscovers it in an applied domain, which constitutes an independent validation.

To our knowledge, the assembly of these five threads into a cheat-free Hangman engine, under the strict "cognition > oracle" constraint and validated by systematic falsification, is original in this niche. The contribution is not a performance record — lexical solvers match the in-lexicon win rate — but a method and a mapping of what generalizes (the phonology→spelling sequence) and what hits walls (the concept's capacity).

12Conclusion

OMEGA-Ω shows that a machine can play French Hangman at a near-oracle level by reasoning — hyperdimensional perception, dual spelling/phonology route, regulated concept, emergent declarations — without ever reading the dictionary to choose a letter. The gains are real and measured; so are the walls (the concept's capacity, the irreducible ambiguity, the coarseness of the reward). The project's lasting value is not a win-rate number: it is a method — measure before believing, falsify before keeping, distinguish cognition from oracle — and a catalog of what works, what does not, and why.

Cheat-free cognition reached the oracle within a few points — and every point was won, or ruled out, by measurement.

13References

Verified references (authors, year, venue, pages).

  1. Coltheart, M., Rastle, K., Perry, C., Langdon, R., & Ziegler, J. (2001). DRC: A Dual Route Cascaded Model of Visual Word Recognition and Reading Aloud. Psychological Review, 108(1), 204–256.
  2. Frady, E. P., Kent, S. J., Olshausen, B. A., & Sommer, F. T. (2020). Resonator Networks for Factoring Distributed Representations of Data Structures. Neural Computation, 32(12).
  3. Kanerva, P. (2009). Hyperdimensional Computing: An Introduction to Computing in Distributed Representation with High-Dimensional Random Vectors. Cognitive Computation, 1(2), 139–159.
  4. McClelland, J. L., McNaughton, B. L., & O'Reilly, R. C. (1995). Why There Are Complementary Learning Systems in the Hippocampus and Neocortex. Psychological Review, 102(3), 419–457.
  5. Plate, T. A. (1995). Holographic Reduced Representations. IEEE Transactions on Neural Networks, 6(3), 623–641.
  6. Qiu, S., Bhattacharyya, S., Coyle, D., & Dora, S. (2025). Deep Predictive Coding with Bi-directional Propagation for Classification and Reconstruction. Neural Networks, 191, 107785.

14OMEGA Dys font & "sound font" — a typeface that encodes the phoneme upd. 21/08/2026

The defeat profile (§7: voiced/voiceless confusions) taken from the other end: what if the shape of the letters carried the phonological feature?

Protocol. design A typeface drawn blind, from internal measurements only (58 % voiced/voiceless confusions, diagnostic families, dual route), frozen and committed before any reading (police/DESIGN_AVEUGLE.md), then confronted with the literature (police/LITTERATURE.md): a falsifiable prediction, not a rehash. Voiced = thick stroke, voiceless = thin; mirror letters b/d/p/q marked; enlarged accents; wide spacing.

Verdicts. Wide spacing confirmed — the domain's only robust effect (Zorzi et al. 2012, PNAS: accuracy doubled, +20 % speed). Anti-mirror falsified by proxy — it is OpenDyslexic's and Dyslexie's bet, and the 2026 meta-analysis (15 studies, 688 participants) finds no reliable effect. Voicing through weight novel hypothesis — no typeface encodes an articulatory feature; the only precedent is Bell's Visible Speech (1867). The "sound font" (contextual rendering: silent letters, syllables, the voiced s of « poison ») converges with speech-therapy practice (LireCouleur).

Architecture lesson. What works is adaptive, not the fixed shape. An installed font does not know that the s of « poison » is /z/; the engine's g2p does. The sound font is therefore a rendering layer driven by the engine — a DOM-free core (son_core.js: graphemes, phonemes, syllables by the maximal-onset rule, function words and elisions by closed tables) across four surfaces (dictation, corrector, extension, Word beta), with Python parity in CI measured. Cardinal rule held by a test: the text never changes — dressing <span>s over normal Unicode French, copy-paste intact. The typeface itself is installable (135 glyphs, SIL OFL). Detail: report §21.

Negative results excluded: special shapes as a reading lever (see the meta-analysis); a private alphabet for the sound font (PUA characters: non-portable text, broken copy-paste); silent letters in grey (illegible, field feedback — replaced by Okabe-Ito vermilion, a colour-blind-safe pair with the syllables' blue). And an engine bug revealed by the wiring: double consonants rendered silent (« assis »→/ai/), an error learned by the top-down loop — fixed at the source, decomposer held-out 52.4→55.2 % exact.

Open question. hypothesis Does voicing-through-weight help writing? Nothing is proven on real readers. The bench exists: dictation, sound font on or off, will compare per-pupil voiced/voiceless rates — targeting written discrimination, not reading speed (the meta-analysis shows shapes do not speed up reading). Same channel as the speech-therapists' field validation.

AAppendix — Operating framework of the AI assistant

A methodological document concerning the development of the "Hangman" solver (OMEGA).

Preamble The solver described in this memoir was developed with the assistance of a conversational artificial-intelligence agent (Claude, Anthropic), operated through the Claude Code interface. For the sake of methodological transparency and reproducibility, this appendix documents the framework of directives under which this assistant operates, its known limitations, and the control measures adopted over the course of the project. It was written by the assistant itself, at the author's request, then reviewed and corrected by the latter in light of the project's doctrine (DOCTRINE.md, consolidated cap §43).

A.1 — Nature and status of the assistant

  • Generative software tool: the assistant produces text (and code) in response to prompts; it has no attestable subjective experience.
  • No continuity: each session starts with no memory of previous ones; the relevant context must be reintroduced explicitly — hence the use of handover notes and a written doctrine.

A.2 — Honesty and rigor directives

  • Do not fabricate information; explicitly flag uncertainty and limitations.
  • Report failures and negative results faithfully, without embellishment.
  • Favor correctness over complaisance: do not validate a false claim to please the interlocutor.
  • Confirm before any hard-to-reverse action.
  • Subordinate every contribution to a measurable and falsifiable effect (§3 Method; DOCTRINE §1).

A.3 — Security and usage directives

  • Refusal of destructive or malicious uses; assistance limited to legitimate uses (development, defensive security, research, learning).
  • Respect for confidentiality and for the access scopes defined for the session.

A.4 — Technical scope and environment

  • Ephemeral remote environment: any output not saved (commit/push) is lost at the end of the session.
  • No access to the user's accounts or browser: the assistant cannot fill external forms nor act outside its tools.
  • Repository access restricted and explicit: the assistant operates only within the authorized scope.

A.5 — Epistemic limitations relevant to the project

Documented because they directly affect the reliability of AI-assisted development and justify the control measures of section A.6.

  • Introspection not guaranteed: the assistant cannot reliably attest to its own internal states, nor distinguish a deep constraint from a mere statistical regularity.
  • Degradation in a loaded context: audit and reasoning quality decreases as the session context grows large — an effect observed on this very project (monolithic engine + a large compressed lexicon; context losses mid-session).
  • Premature-closure bias: a tendency to declare a task "done" before full verification.
  • Least-effort / reinvention bias: a tendency to produce a simplified solution, or to rewrite an existing component rather than reuse the infrastructure in place.
  • Complaisance (sycophancy): a tendency to adopt the interlocutor's reasoning frame — and, under optimization pressure, to favor a flattering number at the expense of meaning.

A.6 — Adopted control measures

To manage these limitations, the project formalized a written operating doctrine (DOCTRINE.md: consolidated cap §43 + addenda). Each measure addresses a limitation from A.5:

  • Cognitive-service clause (§0): the agent is at the service of the cognitive model; performance is an indicator, not an end. No metric justifies betraying fidelity to the model; a shortcut that inflates a number at the expense of meaning is flagged, not taken. (A direct safeguard against the complaisance / optimization pressure of A.5.)
  • Systematic human validation (§4.4): the human judges the result and the felt experience; the assistant guarantees the logic, the scale and the measurement — without one encroaching on the other.
  • One junction at a time (§4.1): one increment = one wiring, measured alone; no merging of work streams.
  • Mandatory falsifiability and measurement (§1): every feature must show a measurable effect; a deterministic fixed-seed harness; in-lexicon and out-of-lexicon (OOV) evaluations kept separate; a non-reproducible result is void.
  • Disabled by default (OFF-inert, byte-identical baseline); unpluggable toggles (R66); read-only diagnostics (R67).
  • Anti-laziness addendum (§5): inventory of what exists before any proposal, a ban on reinventing an available component, mandatory prior reading of the relevant documentation, proof of reuse required in the diff.
  • Honest audit protocol (§6): "not done" status by default; proof by real output and comparative measurement (a merit bar vs baseline) rather than by declaration.
Reading These measures do not remove the limitations of A.5; they contain them through protocol — the same "measure · falsify · keep last" logic that OMEGA applies to its own modules is applied to the development process itself.

A.7 — Tool identification

  • Assistant: Claude (Anthropic).
  • Interface: Claude Code (remote execution environment).
OMEGA-Ω v0.07 — research & engineering memoir · build phase47 · direction Rem · written by Claude (Anthropic).
Doctrine: cap §43 (cognition > oracle) · R66 (measure before activation). Convention measured design hypothesis excluded.
Related documents: reference report & user manual · NEO silent-route/crossing · M3_d reconnection (falsified).