nsaspy/zara

Fork 0

P0 — harden the bounded utterance rewriter on the current Prolog-RLM direct-mode path #122

New issue

Open

opened 2026-08-16 02:41:43 +00:00 by lost-rob0t · 0 comments

lost-rob0t commented

2026-08-16 02:41:43 +00:00

(Migrated from github.com)

Status reconciliation — 2026-09-07

The feature is still needed, but its original backend dependency is obsolete.

#124 was correctly closed not planned: merged PR #233 deliberately replaced the proposed Zara-owned full embedded Prolog LLM client with a pinned, optional Prolog-RLM direct-mode rewrite path over the current provider architecture.
Current master already contains modules/rlm_rewrite.pl and deterministic tests for the basic direct-mode adapter.
Do not close this issue as implemented. The current adapter does not yet prove the stronger bounded rewrite contract below. In particular, malformed/unparseable model output currently falls back to Intent = ask, Args = [Resp]; prompt-time enumeration of canonical intents is not the same as structural validation of returned intent/args; and literal-slot preservation/ambiguity behavior remains part of this issue's required contract.

This issue is therefore evolved in place rather than closed or duplicated.

Goal

Provide one narrow, typed utterance-rewrite stage that improves phrasing/misspelling tolerance without letting model/RLM output become command authority.

Target ordering for voice paths follows #219:

final STT transcript
  -> transcript normalization (#215/#219, when enabled)
  -> deterministic utterance normalization
  -> optional bounded Prolog-RLM direct-mode rewrite
  -> typed rewrite validation
  -> authoritative semantic/Prolog resolution (#150)
  -> capability/policy execution

CLI/text/plugin callers must converge on the same rewrite+resolve boundary rather than creating surface-specific rewriters.

Rewrite result contract

Freeze a typed result equivalent to:

rewrite(CanonicalIntentOrText, TypedSlots, Evidence);
no_rewrite;
ambiguous(Candidates);
rewrite_error(Error).

Exact term/schema is design-owned, but raw model text is never itself a successful command.

Required behavior

Deterministic first

Handle cheap safe cases without a model where possible:

casing/punctuation;
known aliases/paraphrases;
reviewed spelling variants such as forcast -> forecast;
exact existing commands remain semantically unchanged.

Bounded current model path

When configured, reuse the pinned Prolog-RLM direct-mode path introduced by #233 rather than resurrecting #124.

Maintain/strengthen its current bounds:

no arbitrary tools or context operations;
explicit model/provider configuration;
deadline/cancellation;
bounded model calls/tokens/output bytes/cost;
deterministic/low-temperature structured rewrite behavior;
typed failures for missing runtime/credential/contract/provider errors.

Structural validation after the model

Do not trust the prompt to enforce the contract. Validate returned data before semantic resolution:

returned intent/canonical command belongs to the closed registered vocabulary;
args/slots have the expected shape, count, type and bounds for that intent;
unknown fields/intents fail closed;
malformed JSON/schema mismatch returns typed rewrite failure, not generic ask success;
output growth is bounded;
no shell/Python/Prolog goal/executable/capability injection through generic string fields;
canonical result is still revalidated by #150/Prolog semantic resolution before execution.

Literal slot preservation

Rewriting may normalize the command frame but must not silently invent or mutate user-provided filenames, contacts, URLs, message bodies, durations, locations or other literal/typed arguments. Where preservation cannot be proven, return ambiguity/no-rewrite and let the dialogue layer clarify.

Honest ambiguity/failure

Model failure, timeout, invalid output or multiple materially different interpretations must remain explicit. Falling back to conversational ask is allowed only as the downstream product response to a failed/ambiguous rewrite, not as a fake successful rewrite result containing arbitrary model output.

Diagnostics

Emit bounded structured evidence/status suitable for runtime diagnostics and #93/#634-style reasoning inspection without exposing hidden model chain-of-thought or transcript bodies in ordinary metrics.

Required tests

Table-driven deterministic corpus must cover at least:

what is the weather?;
tell me the forecast;
tell me the forcast;
could you please tell me the weather;
existing exact deterministic commands unchanged;
ambiguous phrases not guessed;
quoted/literal slot values preserved byte/semantic-equivalently as appropriate;
model returns unknown intent;
model returns malformed JSON;
model returns wrong arg count/type;
model attempts shell/tool/Prolog/capability injection;
oversized output;
timeout/cancel/provider/RLM failure;
normalized voice text from #219 reaches this stage exactly once;
stale/cancelled rewrite cannot execute or append history;
same contract through text/voice shared runtime path.

No default gate may require network/live provider credentials. Use scripted RLM/model handlers and local fixtures.

Acceptance

Natural variants of a registered capability can be canonicalized before authoritative intent resolution.
Exact existing commands are not semantically changed.
The rewriter cannot execute, grant authority, invent a capability, or bypass semantic/capability policy.
Model/RLM output is structurally validated rather than trusted because the prompt asked for JSON.
Literal user slot values are preserved or the rewrite fails/clarifies honestly.
Ambiguity and failures are typed and never masquerade as successful resolution.
One shared rewrite boundary serves voice/text/client surfaces.
Current pinned Prolog-RLM direct mode is reused when model-backed rewriting is enabled; #124 remains obsolete/closed.

Dependencies / integration

Current model implementation baseline: merged PR #233.
Voice ordering: #215/#219 run before this stage.
Semantic authority and typed slots/capabilities: #150.
This issue does not grant Prolog-RLM direct side-effect authority.

## Status reconciliation — 2026-09-07 The feature is still needed, but its original backend dependency is obsolete. - #124 was correctly closed `not planned`: merged PR #233 deliberately replaced the proposed Zara-owned full embedded Prolog LLM client with a pinned, optional Prolog-RLM direct-mode rewrite path over the current provider architecture. - Current `master` already contains `modules/rlm_rewrite.pl` and deterministic tests for the basic direct-mode adapter. - **Do not close this issue as implemented.** The current adapter does not yet prove the stronger bounded rewrite contract below. In particular, malformed/unparseable model output currently falls back to `Intent = ask, Args = [Resp]`; prompt-time enumeration of canonical intents is not the same as structural validation of returned intent/args; and literal-slot preservation/ambiguity behavior remains part of this issue's required contract. This issue is therefore evolved in place rather than closed or duplicated. ## Goal Provide one narrow, typed utterance-rewrite stage that improves phrasing/misspelling tolerance **without letting model/RLM output become command authority**. Target ordering for voice paths follows #219: ```text final STT transcript -> transcript normalization (#215/#219, when enabled) -> deterministic utterance normalization -> optional bounded Prolog-RLM direct-mode rewrite -> typed rewrite validation -> authoritative semantic/Prolog resolution (#150) -> capability/policy execution ``` CLI/text/plugin callers must converge on the same rewrite+resolve boundary rather than creating surface-specific rewriters. ## Rewrite result contract Freeze a typed result equivalent to: - `rewrite(CanonicalIntentOrText, TypedSlots, Evidence)`; - `no_rewrite`; - `ambiguous(Candidates)`; - `rewrite_error(Error)`. Exact term/schema is design-owned, but **raw model text is never itself a successful command**. ## Required behavior ### Deterministic first Handle cheap safe cases without a model where possible: - casing/punctuation; - known aliases/paraphrases; - reviewed spelling variants such as `forcast` -> `forecast`; - exact existing commands remain semantically unchanged. ### Bounded current model path When configured, reuse the pinned Prolog-RLM direct-mode path introduced by #233 rather than resurrecting #124. Maintain/strengthen its current bounds: - no arbitrary tools or context operations; - explicit model/provider configuration; - deadline/cancellation; - bounded model calls/tokens/output bytes/cost; - deterministic/low-temperature structured rewrite behavior; - typed failures for missing runtime/credential/contract/provider errors. ### Structural validation after the model Do not trust the prompt to enforce the contract. Validate returned data before semantic resolution: - returned intent/canonical command belongs to the closed registered vocabulary; - args/slots have the expected shape, count, type and bounds for that intent; - unknown fields/intents fail closed; - malformed JSON/schema mismatch returns typed rewrite failure, **not generic `ask` success**; - output growth is bounded; - no shell/Python/Prolog goal/executable/capability injection through generic string fields; - canonical result is still revalidated by #150/Prolog semantic resolution before execution. ### Literal slot preservation Rewriting may normalize the command frame but must not silently invent or mutate user-provided filenames, contacts, URLs, message bodies, durations, locations or other literal/typed arguments. Where preservation cannot be proven, return ambiguity/no-rewrite and let the dialogue layer clarify. ### Honest ambiguity/failure Model failure, timeout, invalid output or multiple materially different interpretations must remain explicit. Falling back to conversational `ask` is allowed only as the **downstream product response to a failed/ambiguous rewrite**, not as a fake successful rewrite result containing arbitrary model output. ### Diagnostics Emit bounded structured evidence/status suitable for runtime diagnostics and #93/#634-style reasoning inspection without exposing hidden model chain-of-thought or transcript bodies in ordinary metrics. ## Required tests Table-driven deterministic corpus must cover at least: - `what is the weather?`; - `tell me the forecast`; - `tell me the forcast`; - `could you please tell me the weather`; - existing exact deterministic commands unchanged; - ambiguous phrases not guessed; - quoted/literal slot values preserved byte/semantic-equivalently as appropriate; - model returns unknown intent; - model returns malformed JSON; - model returns wrong arg count/type; - model attempts shell/tool/Prolog/capability injection; - oversized output; - timeout/cancel/provider/RLM failure; - normalized voice text from #219 reaches this stage exactly once; - stale/cancelled rewrite cannot execute or append history; - same contract through text/voice shared runtime path. No default gate may require network/live provider credentials. Use scripted RLM/model handlers and local fixtures. ## Acceptance - Natural variants of a registered capability can be canonicalized before authoritative intent resolution. - Exact existing commands are not semantically changed. - The rewriter cannot execute, grant authority, invent a capability, or bypass semantic/capability policy. - Model/RLM output is structurally validated rather than trusted because the prompt asked for JSON. - Literal user slot values are preserved or the rewrite fails/clarifies honestly. - Ambiguity and failures are typed and never masquerade as successful resolution. - One shared rewrite boundary serves voice/text/client surfaces. - Current pinned Prolog-RLM direct mode is reused when model-backed rewriting is enabled; #124 remains obsolete/closed. ## Dependencies / integration - Current model implementation baseline: merged PR #233. - Voice ordering: #215/#219 run before this stage. - Semantic authority and typed slots/capabilities: #150. - This issue does not grant Prolog-RLM direct side-effect authority.

nsaspy referenced this issue

2026-08-28 03:23:52 +00:00

P1 — wire transcript normalization into voice, dictation and semantic routing before #122 #216

nsaspy referenced this issue

2026-08-28 03:23:52 +00:00

P1 — add transcript-normalizer configuration, selection, status and diagnostics #217

nsaspy referenced this issue

2026-08-28 03:23:53 +00:00

P1 — define provider-neutral transcript normalization contract and raw/normalized runtime semantics #219

nsaspy referenced this issue

2026-08-28 03:23:53 +00:00

[EPIC] Selectable post-STT transcript normalization backends, including local S1-mini #220

nsaspy referenced this issue

2026-08-28 03:24:19 +00:00

[EPIC] Selectable post-STT transcript normalization backends, including local S1-mini #220

nsaspy referenced this issue

2026-08-29 10:40:44 +00:00

[ORDER] ZARA-000 — Repository repair and real-time voice roadmap #1

nsaspy referenced this issue

2026-09-04 14:39:03 +00:00

P0 — replace the minimal Prolog LLM wrapper with a full embedded async-capable client #124

nsaspy referenced this issue

2026-09-04 14:39:03 +00:00

P1 — RAGE research the post-client-split semantic intent architecture and freeze an IntentFrame contract #154

nsaspy referenced this issue

2026-09-04 14:39:03 +00:00

P1 — adapt Prolog resolution to IntentFrame and establish the canonical realistic command corpus #156

nsaspy referenced this issue

2026-09-04 22:44:08 +00:00

feat(prolog): adapt resolution to IntentFrame with canonical semantic corpus (#156) #280

nsaspy changed title from ~~P0 — add a bounded utterance rewriter before Prolog intent resolution~~ to P0 — harden the bounded utterance rewriter on the current Prolog-RLM direct-mode path

2026-09-08 02:03:28 +00:00

nsaspy referenced this issue

2026-09-08 02:03:29 +00:00

P1 — make RuntimeHost own semantic command routing and remove duplicated LLM prompt authority #160

nsaspy referenced this issue

2026-09-08 02:03:38 +00:00

P0 — bounded edit-distance normalization for registered command targets (#122 deterministic slice) #356

nsaspy referenced this issue

2026-09-08 02:03:44 +00:00

[ORDER] ZARA-000 — Repository repair and real-time voice roadmap #1

nsaspy referenced this issue

2026-09-08 02:07:54 +00:00