Correct halting claims: schedule is input-independent in this release

README.md

@@ -20,7 +20,8 @@ pipeline_tag: text-generation
 Hyper thinks in **continuous latent vectors** — "neuralese" — before it writes
 anything. Prior to emitting a single token it runs a recurrent loop over its
 own hidden state, and a trained halt head decides how many of those steps to
-take for each question.
+take. (In this release that decision has collapsed to a fixed schedule — see
+*Known issues*.)
 
 This is a **research preview**, not a finished product. It is early, it is
 small (3B), and one important capability is still developing. All of that is
@@ -50,7 +51,7 @@ automatically: **DirectML** (AMD/Intel on Windows) → **CUDA** (NVIDIA) → **C
 | command | what it does |
 |---|---|
 | `/steps 4` | fixed thinking depth — **4 is the measured best**, and the default |
-| `/steps dynamic` | let the halt head choose (threshold 0.7, usually ~4 steps) |
+| `/steps dynamic` | let the halt head choose — in practice a fixed 5 steps; see *Known issues* |
 | `/halt 0.7` | halting threshold; higher = thinks longer |
 | `/temp 0.7` | sampling temperature — **default 0 (greedy)**, see *Known issues* |
 | `/penalty 1.1` | repetition penalty |
@@ -106,11 +107,6 @@ The mechanism is real and measurable. The reasoning inside it is early.
 
 - **Arithmetic and word problems** when allowed to show working
 - **Knows what it is** — identity is stable and consistent
-- **Runs the loop at a sensible depth on its own** — in dynamic mode the halt
-  head stops at around 4 steps, which is the depth we measured as best. Note
-  we have *not* shown it varies that depth by difficulty: it stops at roughly
-  the same point on easy and hard questions alike. It has learned a good
-  default, not yet adaptivity.
 
 Spot-check on 12 varied questions: 9 correct. Failures were a 3-way ordering
 puzzle, a syllogism, and one long multi-operation expression.
@@ -135,9 +131,17 @@ reply itself is separate and dominates a typical turn.
    reason.
 5. **Weaker than the base model with thinking off.** It has specialised on
    having its latent steps — run it at depth 4.
-6. **`<think>` tags occasionally leak** into output.
-7. **English only**, tested on short reasoning prompts.
-8. **No safety training, and refusal is inconsistent.** `SYSTEM_PROMPT` (top
+6. **Dynamic halting is not adaptive in this release.** The halt head emits a
+   near input-independent schedule — λ ≈ 0.31 on the first step, ≈ 0.24 on
+   every later one, an equilibrium of its training regularisers rather than a
+   read of the question — so cumulative halt mass crosses the 0.7 threshold
+   at step 5 for every prompt, easy or hard. `/steps dynamic` is therefore a
+   fixed depth in practice (and at 5 steps, one past the trained depth of 4).
+   The mechanism runs end-to-end; the decision it encodes does not yet depend
+   on your question. Use `/steps 4`, the default.
+7. **`<think>` tags occasionally leak** into output.
+8. **English only**, tested on short reasoning prompts.
+9. **No safety training, and refusal is inconsistent.** `SYSTEM_PROMPT` (top
    of `hyper_chat.py`) instructs it to refuse illegal requests. Observed
    behaviour across sessions:
    - In fresh sessions at depth 4 it *does* refuse — clearly and on-topic,