slash-tokens 1.5.0 → 1.5.2

package/dist/cli.js

@@ -41,16 +41,19 @@ function writeToMemory(content) {
 // src/slash.ts
 var WASM_INPUT_OFFSET2 = 4096;
 var CALIBRATION = {
-  "claude-opus": 1.85,
-  "claude-opus-4.7": 1.85,
-  "claude-sonnet": 1.85,
-  "claude-haiku": 1.4,
-  "gemini-3.1-pro": 1.4,
-  "gemini-2.5-flash": 1.4,
+  "claude-opus": 2.05,
+  "claude-opus-4.7": 2.05,
+  "claude-sonnet": 2.05,
+  "claude-haiku": 1.45,
+  "gemini-3.1-pro": 1.45,
+  "gemini-2.5-flash": 1.45,
   "grok-4.20": 1.15,
-  "grok-4-1-fast": 1.15
+  "grok-4-1-fast": 1.15,
+  "gpt-5.4": 1.15,
+  "gpt-5.4-mini": 1.15,
+  "gpt-5.4-nano": 1.15
 };
-var DEFAULT_UNKNOWN_MODEL_FACTOR = 1.85;
+var DEFAULT_UNKNOWN_MODEL_FACTOR = 2.05;
 function slash(content, model) {
   if (!content)
     return 0;

package/dist/slash.js

@@ -14,57 +14,94 @@ const WASM_INPUT_OFFSET = 4096;
  * 1 / min(observed_ratio) with a small safety margin, so it does not
  * under-report even against the worst sample in this benchmark:
  *
- * opus-4.7:   1.85 (was 1.50 — that value was insufficient: min observed
- *             ratio 0.571 needs >=1.75 just to avoid under-report, before
- *             any margin. Previously the factor also went UNUSED in
- *             production — see intercept.ts/preflight.ts history, fixed
- *             2026-08-23 same day)
- * claude-opus (generic, now routed to claude-opus-5): 1.85 — carried
+ * SUPERSEDED same day: the first pass (1.85 / 1.85 / 1.40 below) was
+ * measured against a 9-sample corpus with no non-English text and no
+ * SQL/bash. Expanding the corpus to 29 samples (still 2026-08-23) found
+ * a WORSE case for every Claude model — Spanish prose (ratio 0.517) and
+ * SQL/bash syntax (~0.53) drift far more than the original English/JS
+ * corpus ever showed. 1.85 was live in production (v1.5.0) as an
+ * insufficient factor for real Spanish-language and SQL/bash traffic
+ * until this second pass caught it — the exact failure mode this whole
+ * exercise exists to prevent. Re-verify again if the corpus grows further.
+ *
+ * opus-4.7:   2.05 (was 1.85, was 1.50 before that — min observed ratio
+ *             now 0.517 needs >=1.934 before margin, found on Spanish
+ *             prose, not on any sample in the original 9-item corpus)
+ * claude-opus (generic, now routed to claude-opus-5): 2.05 — carried
  *             over from opus-4.7's measurement as the closest tested
  *             sibling; opus-5 itself has not been directly benchmarked
- * claude-sonnet (routed to claude-sonnet-5): 1.85 — measured directly,
- *             nearly identical drift to opus-4.7 (min ratio 0.580)
- * claude-haiku (routed to claude-haiku-4.5): 1.40 — measured directly,
- *             drifts less than the two above (min ratio 0.762)
- * gemini-3.1-pro / gemini-2.5-flash: 1.40 — measured 2026-08-23 against
- *             Google's free countTokens endpoint via the gemini-pro-latest
- *             / gemini-flash-latest aliases (min ratio 0.768, identical
- *             for both — pro/flash share one tokenizer this generation).
+ * claude-sonnet (routed to claude-sonnet-5): 2.05 — measured directly,
+ *             nearly identical drift to opus-4.7 (min ratio 0.525, same
+ *             Spanish-prose sample)
+ * claude-haiku (routed to claude-haiku-4.5): 1.45 (was 1.40) — drifts
+ *             less than the two above but still worsened under the
+ *             bigger corpus (min ratio 0.744, was 0.762)
+ * gemini-3.1-pro / gemini-2.5-flash: 1.45 (was 1.40) — re-verified
+ *             2026-08-23 against the 29-sample corpus via Google's free
+ *             countTokens endpoint (gemini-pro-latest / gemini-flash-latest
+ *             aliases; identical ratios for both — pro/flash share one
+ *             tokenizer this generation). Worst case shifted from the
+ *             original 9-sample corpus: now graphql-response JSON (ratio
+ *             0.749, needs >=1.335), not the prose sample that mattered
+ *             before. 1.40 was technically still safe here (unlike
+ *             Claude's factor, which actually failed) but at the thinnest
+ *             margin of any provider (4.9%) — bumped to match the ~6-8%
+ *             band used everywhere else rather than leave the one factor
+ *             most likely to fail next time this corpus grows again.
  *             Note: 'gemini-3.1-pro' as a literal model ID does not exist
  *             on the live API (404) — same stale-hardcoded-ID class of bug
  *             as the retired Claude snapshots, caught the same day. See
  *             intercept.ts MODEL_API_NAMES: the wire name is now the
  *             '-latest' alias, which sidesteps this whole bug class going
  *             forward since Google repoints it, not us.
- * grok-4.20 / grok-4-1-fast: 1.15 — measured 2026-08-23 (min ratio 0.928,
- *             the least drift of any provider tested — Grok's tokenizer
- *             runs fewer tokens per content than the others, so raw WASM
- *             already over-reports on most samples). xAI has no free
- *             count-tokens endpoint; this used real chat completions
- *             (max_tokens: 1) with a per-model baseline subtracted to
- *             remove xAI's fixed system-preamble overhead (~185-193
- *             tokens, mostly cached) from the content-only count.
+ * gpt-5.4 / gpt-5.4-mini / gpt-5.4-nano: 1.15 — measured 2026-08-23
+ *             against js-tiktoken's o200k_base (free, local, exact — same
+ *             encoding for the whole GPT-5.4 family per OpenAI's own
+ *             tiktoken docs). Min ratio 0.923 against the expanded
+ *             29-sample corpus. This was a real gap, not just an accuracy
+ *             tweak: GPT had a calibration SCRIPT (bench/calibrate-openai.ts)
+ *             from the same day as the Claude fixes, but was never actually
+ *             added to this table — it was silently falling through to
+ *             DEFAULT_UNKNOWN_MODEL_FACTOR (1.85), a ~60% larger correction
+ *             than it needed. Still safe either way (1.85 > required
+ *             minimum), just needlessly inflated for real GPT users.
+ * grok-4.20 / grok-4-1-fast: 1.15 — re-verified 2026-08-23 against the
+ *             29-sample corpus (20 new samples: more languages, Spanish/
+ *             Japanese prose, more JSON shapes) and UNCHANGED — same
+ *             worst case (technical-docs prose, ratio 0.928) as the
+ *             original 9-sample measurement. The only one of the three
+ *             non-GPT providers whose factor held without a bump; Claude
+ *             and Gemini both needed one, on different content types each
+ *             (Spanish prose, then JSON). xAI has no free count-tokens
+ *             endpoint; this used real chat completions (max_tokens: 1)
+ *             with a per-model baseline subtracted to remove xAI's fixed
+ *             system-preamble overhead (~185-193 tokens, mostly cached)
+ *             from the content-only count.
  *             Both literal API IDs ('grok-4.20', 'grok-4-1-fast') are
  *             fully retired — not just old snapshots, gone entirely —
  *             remapped to grok-4.20-0309-non-reasoning / grok-4.3 in
  *             intercept.ts MODEL_API_NAMES the same day this was found.
  *
- * Known limitation: derived from a 9-sample benchmark corpus, several
- * of which are slash-tokens' own code/docs (not representative
- * third-party content). Re-run bench/calibrate.ts with a larger,
- * more diverse corpus before treating these as final.
+ * All four providers (Claude, Gemini, Grok, GPT) are now verified against
+ * the same 29-sample corpus (expanded 2026-08-23 from the original 9,
+ * which was mostly slash-tokens' own code/docs — not representative
+ * third-party content). Re-verify again if the corpus grows further —
+ * Claude and Gemini both needed real bumps the first time it grew.
  *
  * Slash must NEVER under-report. Over-reporting is safe (go/no-go only).
  */
 const CALIBRATION = {
-    'claude-opus': 1.85,
-    'claude-opus-4.7': 1.85,
-    'claude-sonnet': 1.85,
-    'claude-haiku': 1.40,
-    'gemini-3.1-pro': 1.40,
-    'gemini-2.5-flash': 1.40,
+    'claude-opus': 2.05,
+    'claude-opus-4.7': 2.05,
+    'claude-sonnet': 2.05,
+    'claude-haiku': 1.45,
+    'gemini-3.1-pro': 1.45,
+    'gemini-2.5-flash': 1.45,
     'grok-4.20': 1.15,
     'grok-4-1-fast': 1.15,
+    'gpt-5.4': 1.15,
+    'gpt-5.4-mini': 1.15,
+    'gpt-5.4-nano': 1.15,
 };
 /**
  * Fallback factor for any model not in CALIBRATION — new model IDs,
@@ -78,7 +115,7 @@ const CALIBRATION = {
  * haven't measured, the only safe assumption is the worst one we've
  * actually observed (see CALIBRATION comment above).
  */
-const DEFAULT_UNKNOWN_MODEL_FACTOR = 1.85;
+const DEFAULT_UNKNOWN_MODEL_FACTOR = 2.05;
 /**
  * Estimate token count for a string.
  * Sub-millisecond. Zero allocations in WASM.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "slash-tokens",
-  "version": "1.5.0",
+  "version": "1.5.2",
   "description": "Token Optimization for Context Engineers. 4.8 KB WASM. Sub-millisecond. Zero dependencies.",
   "main": "dist/index.js",
   "module": "dist/index.js",